1 /* 2 * Copyright (c) 2001, 2018, 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 "gc/g1/g1ConcurrentMarkBitMap.hpp" 29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp" 30 #include "gc/g1/g1RegionMarkStatsCache.hpp" 31 #include "gc/g1/heapRegionSet.hpp" 32 #include "gc/shared/taskqueue.hpp" 33 #include "memory/allocation.hpp" 34 35 class ConcurrentGCTimer; 36 class ConcurrentMarkThread; 37 class G1CollectedHeap; 38 class G1CMTask; 39 class G1ConcurrentMark; 40 class G1OldTracer; 41 class G1RegionToSpaceMapper; 42 class G1SurvivorRegions; 43 44 #ifdef _MSC_VER 45 #pragma warning(push) 46 // warning C4522: multiple assignment operators specified 47 #pragma warning(disable:4522) 48 #endif 49 50 // This is a container class for either an oop or a continuation address for 51 // mark stack entries. Both are pushed onto the mark stack. 52 class G1TaskQueueEntry { 53 private: 54 void* _holder; 55 56 static const uintptr_t ArraySliceBit = 1; 57 58 G1TaskQueueEntry(oop obj) : _holder(obj) { 59 assert(_holder != NULL, "Not allowed to set NULL task queue element"); 60 } 61 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { } 62 public: 63 G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; } 64 G1TaskQueueEntry() : _holder(NULL) { } 65 66 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); } 67 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); } 68 69 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) { 70 _holder = t._holder; 71 return *this; 72 } 73 74 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile { 75 _holder = t._holder; 76 return *this; 77 } 78 79 oop obj() const { 80 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder)); 81 return (oop)_holder; 82 } 83 84 HeapWord* slice() const { 85 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder)); 86 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit); 87 } 88 89 bool is_oop() const { return !is_array_slice(); } 90 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; } 91 bool is_null() const { return _holder == NULL; } 92 }; 93 94 #ifdef _MSC_VER 95 #pragma warning(pop) 96 #endif 97 98 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue; 99 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet; 100 101 // Closure used by CM during concurrent reference discovery 102 // and reference processing (during remarking) to determine 103 // if a particular object is alive. It is primarily used 104 // to determine if referents of discovered reference objects 105 // are alive. An instance is also embedded into the 106 // reference processor as the _is_alive_non_header field 107 class G1CMIsAliveClosure : public BoolObjectClosure { 108 G1CollectedHeap* _g1h; 109 public: 110 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1h(g1) { } 111 112 bool do_object_b(oop obj); 113 }; 114 115 // Represents the overflow mark stack used by concurrent marking. 116 // 117 // Stores oops in a huge buffer in virtual memory that is always fully committed. 118 // Resizing may only happen during a STW pause when the stack is empty. 119 // 120 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark 121 // stack memory is split into evenly sized chunks of oops. Users can only 122 // add or remove entries on that basis. 123 // Chunks are filled in increasing address order. Not completely filled chunks 124 // have a NULL element as a terminating element. 125 // 126 // Every chunk has a header containing a single pointer element used for memory 127 // management. This wastes some space, but is negligible (< .1% with current sizing). 128 // 129 // Memory management is done using a mix of tracking a high water-mark indicating 130 // that all chunks at a lower address are valid chunks, and a singly linked free 131 // list connecting all empty chunks. 132 class G1CMMarkStack { 133 public: 134 // Number of TaskQueueEntries that can fit in a single chunk. 135 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */; 136 private: 137 struct TaskQueueEntryChunk { 138 TaskQueueEntryChunk* next; 139 G1TaskQueueEntry data[EntriesPerChunk]; 140 }; 141 142 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack. 143 144 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area. 145 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements. 146 147 char _pad0[DEFAULT_CACHE_LINE_SIZE]; 148 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users. 149 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)]; 150 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data. 151 volatile size_t _chunks_in_chunk_list; 152 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)]; 153 154 volatile size_t _hwm; // High water mark within the reserved space. 155 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)]; 156 157 // Allocate a new chunk from the reserved memory, using the high water mark. Returns 158 // NULL if out of memory. 159 TaskQueueEntryChunk* allocate_new_chunk(); 160 161 // Atomically add the given chunk to the list. 162 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem); 163 // Atomically remove and return a chunk from the given list. Returns NULL if the 164 // list is empty. 165 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list); 166 167 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem); 168 void add_chunk_to_free_list(TaskQueueEntryChunk* elem); 169 170 TaskQueueEntryChunk* remove_chunk_from_chunk_list(); 171 TaskQueueEntryChunk* remove_chunk_from_free_list(); 172 173 // Resizes the mark stack to the given new capacity. Releases any previous 174 // memory if successful. 175 bool resize(size_t new_capacity); 176 177 public: 178 G1CMMarkStack(); 179 ~G1CMMarkStack(); 180 181 // Alignment and minimum capacity of this mark stack in number of oops. 182 static size_t capacity_alignment(); 183 184 // Allocate and initialize the mark stack with the given number of oops. 185 bool initialize(size_t initial_capacity, size_t max_capacity); 186 187 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark 188 // stack. If less than EntriesPerChunk elements are to be pushed, the array must 189 // be terminated with a NULL. 190 // Returns whether the buffer contents were successfully pushed to the global mark 191 // stack. 192 bool par_push_chunk(G1TaskQueueEntry* buffer); 193 194 // Pops a chunk from this mark stack, copying them into the given buffer. This 195 // chunk may contain up to EntriesPerChunk elements. If there are less, the last 196 // element in the array is a NULL pointer. 197 bool par_pop_chunk(G1TaskQueueEntry* buffer); 198 199 // Return whether the chunk list is empty. Racy due to unsynchronized access to 200 // _chunk_list. 201 bool is_empty() const { return _chunk_list == NULL; } 202 203 size_t capacity() const { return _chunk_capacity; } 204 205 // Expand the stack, typically in response to an overflow condition 206 void expand(); 207 208 // Return the approximate number of oops on this mark stack. Racy due to 209 // unsynchronized access to _chunks_in_chunk_list. 210 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; } 211 212 void set_empty(); 213 214 // Apply Fn to every oop on the mark stack. The mark stack must not 215 // be modified while iterating. 216 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN; 217 }; 218 219 // Root Regions are regions that are not empty at the beginning of a 220 // marking cycle and which we might collect during an evacuation pause 221 // while the cycle is active. Given that, during evacuation pauses, we 222 // do not copy objects that are explicitly marked, what we have to do 223 // for the root regions is to scan them and mark all objects reachable 224 // from them. According to the SATB assumptions, we only need to visit 225 // each object once during marking. So, as long as we finish this scan 226 // before the next evacuation pause, we can copy the objects from the 227 // root regions without having to mark them or do anything else to them. 228 // 229 // Currently, we only support root region scanning once (at the start 230 // of the marking cycle) and the root regions are all the survivor 231 // regions populated during the initial-mark pause. 232 class G1CMRootRegions { 233 private: 234 const G1SurvivorRegions* _survivors; 235 G1ConcurrentMark* _cm; 236 237 volatile bool _scan_in_progress; 238 volatile bool _should_abort; 239 volatile int _claimed_survivor_index; 240 241 void notify_scan_done(); 242 243 public: 244 G1CMRootRegions(); 245 // We actually do most of the initialization in this method. 246 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm); 247 248 // Reset the claiming / scanning of the root regions. 249 void prepare_for_scan(); 250 251 // Forces get_next() to return NULL so that the iteration aborts early. 252 void abort() { _should_abort = true; } 253 254 // Return true if the CM thread are actively scanning root regions, 255 // false otherwise. 256 bool scan_in_progress() { return _scan_in_progress; } 257 258 // Claim the next root region to scan atomically, or return NULL if 259 // all have been claimed. 260 HeapRegion* claim_next(); 261 262 // The number of root regions to scan. 263 uint num_root_regions() const; 264 265 void cancel_scan(); 266 267 // Flag that we're done with root region scanning and notify anyone 268 // who's waiting on it. If aborted is false, assume that all regions 269 // have been claimed. 270 void scan_finished(); 271 272 // If CM threads are still scanning root regions, wait until they 273 // are done. Return true if we had to wait, false otherwise. 274 bool wait_until_scan_finished(); 275 }; 276 277 // This class manages data structures and methods for doing liveness analysis in 278 // G1's concurrent cycle. 279 class G1ConcurrentMark : public CHeapObj<mtGC> { 280 friend class ConcurrentMarkThread; 281 friend class G1CMRefProcTaskProxy; 282 friend class G1CMRefProcTaskExecutor; 283 friend class G1CMKeepAliveAndDrainClosure; 284 friend class G1CMDrainMarkingStackClosure; 285 friend class G1CMBitMapClosure; 286 friend class G1CMConcurrentMarkingTask; 287 friend class G1CMRemarkTask; 288 friend class G1CMTask; 289 290 ConcurrentMarkThread* _cm_thread; // The thread doing the work 291 G1CollectedHeap* _g1h; // The heap 292 bool _completed_initialization; // Set to true when initialization is complete 293 294 // Concurrent marking support structures 295 G1CMBitMap _mark_bitmap_1; 296 G1CMBitMap _mark_bitmap_2; 297 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap 298 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap 299 300 // Heap bounds 301 MemRegion const _heap; 302 303 // Root region tracking and claiming 304 G1CMRootRegions _root_regions; 305 306 // For grey objects 307 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger 308 HeapWord* volatile _finger; // The global finger, region aligned, 309 // always pointing to the end of the 310 // last claimed region 311 312 uint _worker_id_offset; 313 uint _max_num_tasks; // Maximum number of marking tasks 314 uint _num_active_tasks; // Number of tasks currently active 315 G1CMTask** _tasks; // Task queue array (max_worker_id length) 316 317 G1CMTaskQueueSet* _task_queues; // Task queue set 318 ParallelTaskTerminator _terminator; // For termination 319 320 // Two sync barriers that are used to synchronize tasks when an 321 // overflow occurs. The algorithm is the following. All tasks enter 322 // the first one to ensure that they have all stopped manipulating 323 // the global data structures. After they exit it, they re-initialize 324 // their data structures and task 0 re-initializes the global data 325 // structures. Then, they enter the second sync barrier. This 326 // ensure, that no task starts doing work before all data 327 // structures (local and global) have been re-initialized. When they 328 // exit it, they are free to start working again. 329 WorkGangBarrierSync _first_overflow_barrier_sync; 330 WorkGangBarrierSync _second_overflow_barrier_sync; 331 332 // This is set by any task, when an overflow on the global data 333 // structures is detected 334 volatile bool _has_overflown; 335 // True: marking is concurrent, false: we're in remark 336 volatile bool _concurrent; 337 // Set at the end of a Full GC so that marking aborts 338 volatile bool _has_aborted; 339 340 // Used when remark aborts due to an overflow to indicate that 341 // another concurrent marking phase should start 342 volatile bool _restart_for_overflow; 343 344 // This is true from the very start of concurrent marking until the 345 // point when all the tasks complete their work. It is really used 346 // to determine the points between the end of concurrent marking and 347 // time of remark. 348 volatile bool _concurrent_marking_in_progress; 349 350 ConcurrentGCTimer* _gc_timer_cm; 351 352 G1OldTracer* _gc_tracer_cm; 353 354 // Timing statistics. All of them are in ms 355 NumberSeq _init_times; 356 NumberSeq _remark_times; 357 NumberSeq _remark_mark_times; 358 NumberSeq _remark_weak_ref_times; 359 NumberSeq _cleanup_times; 360 double _total_cleanup_time; 361 362 double* _accum_task_vtime; // Accumulated task vtime 363 364 WorkGang* _concurrent_workers; 365 uint _num_concurrent_workers; // The number of marking worker threads we're using 366 uint _max_concurrent_workers; // Maximum number of marking worker threads 367 368 void finalize_marking(); 369 370 void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes); 371 void weak_refs_work(bool clear_all_soft_refs); 372 373 void swap_mark_bitmaps(); 374 375 void reclaim_empty_regions(); 376 377 // Resets the global marking data structures, as well as the 378 // task local ones; should be called during initial mark. 379 void reset(); 380 381 // Resets all the marking data structures. Called when we have to restart 382 // marking or when marking completes (via set_non_marking_state below). 383 void reset_marking_for_restart(); 384 385 // We do this after we're done with marking so that the marking data 386 // structures are initialized to a sensible and predictable state. 387 void reset_at_marking_complete(); 388 389 // Called to indicate how many threads are currently active. 390 void set_concurrency(uint active_tasks); 391 392 // Should be called to indicate which phase we're in (concurrent 393 // mark or remark) and how many threads are currently active. 394 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 395 396 // Prints all gathered CM-related statistics 397 void print_stats(); 398 399 HeapWord* finger() { return _finger; } 400 bool concurrent() { return _concurrent; } 401 uint active_tasks() { return _num_active_tasks; } 402 ParallelTaskTerminator* terminator() { return &_terminator; } 403 404 // Claims the next available region to be scanned by a marking 405 // task/thread. It might return NULL if the next region is empty or 406 // we have run out of regions. In the latter case, out_of_regions() 407 // determines whether we've really run out of regions or the task 408 // should call claim_region() again. This might seem a bit 409 // awkward. Originally, the code was written so that claim_region() 410 // either successfully returned with a non-empty region or there 411 // were no more regions to be claimed. The problem with this was 412 // that, in certain circumstances, it iterated over large chunks of 413 // the heap finding only empty regions and, while it was working, it 414 // was preventing the calling task to call its regular clock 415 // method. So, this way, each task will spend very little time in 416 // claim_region() and is allowed to call the regular clock method 417 // frequently. 418 HeapRegion* claim_region(uint worker_id); 419 420 // Determines whether we've run out of regions to scan. Note that 421 // the finger can point past the heap end in case the heap was expanded 422 // to satisfy an allocation without doing a GC. This is fine, because all 423 // objects in those regions will be considered live anyway because of 424 // SATB guarantees (i.e. their TAMS will be equal to bottom). 425 bool out_of_regions() { return _finger >= _heap.end(); } 426 427 // Returns the task with the given id 428 G1CMTask* task(uint id) { 429 // During initial mark we use the parallel gc threads to do some work, so 430 // we can only compare against _max_num_tasks. 431 assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks); 432 return _tasks[id]; 433 } 434 435 // Access / manipulation of the overflow flag which is set to 436 // indicate that the global stack has overflown 437 bool has_overflown() { return _has_overflown; } 438 void set_has_overflown() { _has_overflown = true; } 439 void clear_has_overflown() { _has_overflown = false; } 440 bool restart_for_overflow() { return _restart_for_overflow; } 441 442 // Methods to enter the two overflow sync barriers 443 void enter_first_sync_barrier(uint worker_id); 444 void enter_second_sync_barrier(uint worker_id); 445 446 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is 447 // true, periodically insert checks to see if this method should exit prematurely. 448 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield); 449 450 // Clear statistics gathered during the concurrent cycle for the given region after 451 // it has been reclaimed. 452 void clear_statistics_in_region(uint region_idx); 453 // Region statistics gathered during marking. 454 G1RegionMarkStats* _region_mark_stats; 455 // Top pointer for each region at the start of the rebuild remembered set process 456 // for regions which remembered sets need to be rebuilt. A NULL for a given region 457 // means that this region does not be scanned during the rebuilding remembered 458 // set phase at all. 459 HeapWord* volatile* _top_at_rebuild_starts; 460 public: 461 void add_to_liveness(uint worker_id, oop const obj, size_t size); 462 // Liveness of the given region as determined by concurrent marking, i.e. the amount of 463 // live words between bottom and nTAMS. 464 size_t liveness(uint region) { return _region_mark_stats[region]._live_words; } 465 466 // Sets the internal top_at_region_start for the given region to current top of the region. 467 inline void update_top_at_rebuild_start(HeapRegion* r); 468 // TARS for the given region during remembered set rebuilding. 469 inline HeapWord* top_at_rebuild_start(uint region) const; 470 471 // Notification for eagerly reclaimed regions to clean up. 472 void humongous_object_eagerly_reclaimed(HeapRegion* r); 473 // Manipulation of the global mark stack. 474 // The push and pop operations are used by tasks for transfers 475 // between task-local queues and the global mark stack. 476 bool mark_stack_push(G1TaskQueueEntry* arr) { 477 if (!_global_mark_stack.par_push_chunk(arr)) { 478 set_has_overflown(); 479 return false; 480 } 481 return true; 482 } 483 bool mark_stack_pop(G1TaskQueueEntry* arr) { 484 return _global_mark_stack.par_pop_chunk(arr); 485 } 486 size_t mark_stack_size() const { return _global_mark_stack.size(); } 487 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; } 488 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); } 489 490 G1CMRootRegions* root_regions() { return &_root_regions; } 491 492 bool concurrent_marking_in_progress() const { 493 return _concurrent_marking_in_progress; 494 } 495 void set_concurrent_marking_in_progress() { 496 _concurrent_marking_in_progress = true; 497 } 498 void clear_concurrent_marking_in_progress() { 499 _concurrent_marking_in_progress = false; 500 } 501 502 void concurrent_cycle_start(); 503 // Abandon current marking iteration due to a Full GC. 504 void concurrent_cycle_abort(); 505 void concurrent_cycle_end(); 506 507 void update_accum_task_vtime(int i, double vtime) { 508 _accum_task_vtime[i] += vtime; 509 } 510 511 double all_task_accum_vtime() { 512 double ret = 0.0; 513 for (uint i = 0; i < _max_num_tasks; ++i) 514 ret += _accum_task_vtime[i]; 515 return ret; 516 } 517 518 // Attempts to steal an object from the task queues of other tasks 519 bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry); 520 521 G1ConcurrentMark(G1CollectedHeap* g1h, 522 G1RegionToSpaceMapper* prev_bitmap_storage, 523 G1RegionToSpaceMapper* next_bitmap_storage); 524 ~G1ConcurrentMark(); 525 526 ConcurrentMarkThread* cm_thread() { return _cm_thread; } 527 528 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; } 529 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; } 530 531 // Calculates the number of concurrent GC threads to be used in the marking phase. 532 uint calc_active_marking_workers(); 533 534 // Moves all per-task cached data into global state. 535 void flush_all_task_caches(); 536 // Prepare internal data structures for the next mark cycle. This includes clearing 537 // the next mark bitmap and some internal data structures. This method is intended 538 // to be called concurrently to the mutator. It will yield to safepoint requests. 539 void cleanup_for_next_mark(); 540 541 // Clear the previous marking bitmap during safepoint. 542 void clear_prev_bitmap(WorkGang* workers); 543 544 // Return whether the next mark bitmap has no marks set. To be used for assertions 545 // only. Will not yield to pause requests. 546 bool next_mark_bitmap_is_clear(); 547 548 // These two methods do the work that needs to be done at the start and end of the 549 // initial mark pause. 550 void pre_initial_mark(); 551 void post_initial_mark(); 552 553 // Scan all the root regions and mark everything reachable from 554 // them. 555 void scan_root_regions(); 556 557 // Scan a single root region and mark everything reachable from it. 558 void scan_root_region(HeapRegion* hr, uint worker_id); 559 560 // Do concurrent phase of marking, to a tentative transitive closure. 561 void mark_from_roots(); 562 563 void remark(); 564 565 void cleanup(); 566 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use 567 // this carefully. 568 inline void mark_in_prev_bitmap(oop p); 569 570 // Clears marks for all objects in the given range, for the prev or 571 // next bitmaps. Caution: the previous bitmap is usually 572 // read-only, so use this carefully! 573 void clear_range_in_prev_bitmap(MemRegion mr); 574 575 inline bool is_marked_in_prev_bitmap(oop p) const; 576 577 // Verify that there are no collection set oops on the stacks (taskqueues / 578 // global mark stack) and fingers (global / per-task). 579 // If marking is not in progress, it's a no-op. 580 void verify_no_cset_oops() PRODUCT_RETURN; 581 582 inline bool do_yield_check(); 583 584 bool has_aborted() { return _has_aborted; } 585 586 void print_summary_info(); 587 588 void print_worker_threads_on(outputStream* st) const; 589 void threads_do(ThreadClosure* tc) const; 590 591 void print_on_error(outputStream* st) const; 592 593 // Mark the given object on the next bitmap if it is below nTAMS. 594 // If the passed obj_size is zero, it is recalculated from the given object if 595 // needed. This is to be as lazy as possible with accessing the object's size. 596 inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj, size_t const obj_size = 0); 597 inline bool mark_in_next_bitmap(uint worker_id, oop const obj, size_t const obj_size = 0); 598 599 // Returns true if initialization was successfully completed. 600 bool completed_initialization() const { 601 return _completed_initialization; 602 } 603 604 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; } 605 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; } 606 607 private: 608 // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application. 609 void rebuild_rem_set_concurrently(); 610 }; 611 612 // A class representing a marking task. 613 class G1CMTask : public TerminatorTerminator { 614 private: 615 enum PrivateConstants { 616 // The regular clock call is called once the scanned words reaches 617 // this limit 618 words_scanned_period = 12*1024, 619 // The regular clock call is called once the number of visited 620 // references reaches this limit 621 refs_reached_period = 1024, 622 // Initial value for the hash seed, used in the work stealing code 623 init_hash_seed = 17 624 }; 625 626 // Number of entries in the per-task stats entry. This seems enough to have a very 627 // low cache miss rate. 628 static const uint RegionMarkStatsCacheSize = 1024; 629 630 G1CMObjArrayProcessor _objArray_processor; 631 632 uint _worker_id; 633 G1CollectedHeap* _g1h; 634 G1ConcurrentMark* _cm; 635 G1CMBitMap* _next_mark_bitmap; 636 // the task queue of this task 637 G1CMTaskQueue* _task_queue; 638 639 G1RegionMarkStatsCache _mark_stats_cache; 640 // Number of calls to this task 641 uint _calls; 642 643 // When the virtual timer reaches this time, the marking step should exit 644 double _time_target_ms; 645 // Start time of the current marking step 646 double _start_time_ms; 647 648 // Oop closure used for iterations over oops 649 G1CMOopClosure* _cm_oop_closure; 650 651 // Region this task is scanning, NULL if we're not scanning any 652 HeapRegion* _curr_region; 653 // Local finger of this task, NULL if we're not scanning a region 654 HeapWord* _finger; 655 // Limit of the region this task is scanning, NULL if we're not scanning one 656 HeapWord* _region_limit; 657 658 // Number of words this task has scanned 659 size_t _words_scanned; 660 // When _words_scanned reaches this limit, the regular clock is 661 // called. Notice that this might be decreased under certain 662 // circumstances (i.e. when we believe that we did an expensive 663 // operation). 664 size_t _words_scanned_limit; 665 // Initial value of _words_scanned_limit (i.e. what it was 666 // before it was decreased). 667 size_t _real_words_scanned_limit; 668 669 // Number of references this task has visited 670 size_t _refs_reached; 671 // When _refs_reached reaches this limit, the regular clock is 672 // called. Notice this this might be decreased under certain 673 // circumstances (i.e. when we believe that we did an expensive 674 // operation). 675 size_t _refs_reached_limit; 676 // Initial value of _refs_reached_limit (i.e. what it was before 677 // it was decreased). 678 size_t _real_refs_reached_limit; 679 680 // Used by the work stealing 681 int _hash_seed; 682 // If true, then the task has aborted for some reason 683 bool _has_aborted; 684 // Set when the task aborts because it has met its time quota 685 bool _has_timed_out; 686 // True when we're draining SATB buffers; this avoids the task 687 // aborting due to SATB buffers being available (as we're already 688 // dealing with them) 689 bool _draining_satb_buffers; 690 691 // Number sequence of past step times 692 NumberSeq _step_times_ms; 693 // Elapsed time of this task 694 double _elapsed_time_ms; 695 // Termination time of this task 696 double _termination_time_ms; 697 // When this task got into the termination protocol 698 double _termination_start_time_ms; 699 700 // True when the task is during a concurrent phase, false when it is 701 // in the remark phase (so, in the latter case, we do not have to 702 // check all the things that we have to check during the concurrent 703 // phase, i.e. SATB buffer availability...) 704 bool _concurrent; 705 706 TruncatedSeq _marking_step_diffs_ms; 707 708 // Updates the local fields after this task has claimed 709 // a new region to scan 710 void setup_for_region(HeapRegion* hr); 711 // Makes the limit of the region up-to-date 712 void update_region_limit(); 713 714 // Called when either the words scanned or the refs visited limit 715 // has been reached 716 void reached_limit(); 717 // Recalculates the words scanned and refs visited limits 718 void recalculate_limits(); 719 // Decreases the words scanned and refs visited limits when we reach 720 // an expensive operation 721 void decrease_limits(); 722 // Checks whether the words scanned or refs visited reached their 723 // respective limit and calls reached_limit() if they have 724 void check_limits() { 725 if (_words_scanned >= _words_scanned_limit || 726 _refs_reached >= _refs_reached_limit) { 727 reached_limit(); 728 } 729 } 730 // Supposed to be called regularly during a marking step as 731 // it checks a bunch of conditions that might cause the marking step 732 // to abort 733 void regular_clock_call(); 734 735 // Test whether obj might have already been passed over by the 736 // mark bitmap scan, and so needs to be pushed onto the mark stack. 737 bool is_below_finger(oop obj, HeapWord* global_finger) const; 738 739 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry); 740 public: 741 // Apply the closure on the given area of the objArray. Return the number of words 742 // scanned. 743 inline size_t scan_objArray(objArrayOop obj, MemRegion mr); 744 // Resets the task; should be called right at the beginning of a marking phase. 745 void reset(G1CMBitMap* next_mark_bitmap); 746 // Clears all the fields that correspond to a claimed region. 747 void clear_region_fields(); 748 749 void set_concurrent(bool concurrent) { _concurrent = concurrent; } 750 751 // The main method of this class which performs a marking step 752 // trying not to exceed the given duration. However, it might exit 753 // prematurely, according to some conditions (i.e. SATB buffers are 754 // available for processing). 755 void do_marking_step(double target_ms, 756 bool do_termination, 757 bool is_serial); 758 759 // These two calls start and stop the timer 760 void record_start_time() { 761 _elapsed_time_ms = os::elapsedTime() * 1000.0; 762 } 763 void record_end_time() { 764 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 765 } 766 767 // Returns the worker ID associated with this task. 768 uint worker_id() { return _worker_id; } 769 770 // From TerminatorTerminator. It determines whether this task should 771 // exit the termination protocol after it's entered it. 772 virtual bool should_exit_termination(); 773 774 // Resets the local region fields after a task has finished scanning a 775 // region; or when they have become stale as a result of the region 776 // being evacuated. 777 void giveup_current_region(); 778 779 HeapWord* finger() { return _finger; } 780 781 bool has_aborted() { return _has_aborted; } 782 void set_has_aborted() { _has_aborted = true; } 783 void clear_has_aborted() { _has_aborted = false; } 784 785 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 786 787 // Increment the number of references this task has visited. 788 void increment_refs_reached() { ++_refs_reached; } 789 790 // Grey the object by marking it. If not already marked, push it on 791 // the local queue if below the finger. 792 // obj is below its region's NTAMS. 793 inline void make_reference_grey(oop obj); 794 795 // Grey the object (by calling make_grey_reference) if required, 796 // e.g. obj is below its containing region's NTAMS. 797 // Precondition: obj is a valid heap object. 798 template <class T> 799 inline void deal_with_reference(T* p); 800 801 // Scans an object and visits its children. 802 inline void scan_task_entry(G1TaskQueueEntry task_entry); 803 804 // Pushes an object on the local queue. 805 inline void push(G1TaskQueueEntry task_entry); 806 807 // Move entries to the global stack. 808 void move_entries_to_global_stack(); 809 // Move entries from the global stack, return true if we were successful to do so. 810 bool get_entries_from_global_stack(); 811 812 // Pops and scans objects from the local queue. If partially is 813 // true, then it stops when the queue size is of a given limit. If 814 // partially is false, then it stops when the queue is empty. 815 void drain_local_queue(bool partially); 816 // Moves entries from the global stack to the local queue and 817 // drains the local queue. If partially is true, then it stops when 818 // both the global stack and the local queue reach a given size. If 819 // partially if false, it tries to empty them totally. 820 void drain_global_stack(bool partially); 821 // Keeps picking SATB buffers and processing them until no SATB 822 // buffers are available. 823 void drain_satb_buffers(); 824 825 // Moves the local finger to a new location 826 inline void move_finger_to(HeapWord* new_finger) { 827 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 828 _finger = new_finger; 829 } 830 831 G1CMTask(uint worker_id, 832 G1ConcurrentMark *cm, 833 G1CMTaskQueue* task_queue, 834 G1RegionMarkStats* mark_stats, 835 uint max_regions); 836 837 inline void update_liveness(oop const obj, size_t const obj_size); 838 839 // Clear (without flushing) the mark cache entry for the given region. 840 void clear_mark_stats_cache(uint region_idx); 841 // Evict the whole statistics cache into the global statistics. Returns the 842 // number of cache hits and misses so far. 843 Pair<size_t, size_t> flush_mark_stats_cache(); 844 // Prints statistics associated with this task 845 void print_stats(); 846 }; 847 848 // Class that's used to to print out per-region liveness 849 // information. It's currently used at the end of marking and also 850 // after we sort the old regions at the end of the cleanup operation. 851 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure { 852 private: 853 // Accumulators for these values. 854 size_t _total_used_bytes; 855 size_t _total_capacity_bytes; 856 size_t _total_prev_live_bytes; 857 size_t _total_next_live_bytes; 858 859 // Accumulator for the remembered set size 860 size_t _total_remset_bytes; 861 862 // Accumulator for strong code roots memory size 863 size_t _total_strong_code_roots_bytes; 864 865 static double bytes_to_mb(size_t val) { 866 return (double) val / (double) M; 867 } 868 869 public: 870 // The header and footer are printed in the constructor and 871 // destructor respectively. 872 G1PrintRegionLivenessInfoClosure(const char* phase_name); 873 virtual bool do_heap_region(HeapRegion* r); 874 ~G1PrintRegionLivenessInfoClosure(); 875 }; 876 877 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP