1 /* 2 * Copyright (c) 2001, 2019, 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_GC_G1_G1CONCURRENTMARK_HPP 26 #define SHARE_GC_G1_G1CONCURRENTMARK_HPP 27 28 #include "gc/g1/g1ConcurrentMarkBitMap.hpp" 29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp" 30 #include "gc/g1/g1HeapVerifier.hpp" 31 #include "gc/g1/g1RegionMarkStatsCache.hpp" 32 #include "gc/g1/heapRegionSet.hpp" 33 #include "gc/shared/taskqueue.hpp" 34 #include "gc/shared/verifyOption.hpp" 35 #include "gc/shared/workgroup.hpp" 36 #include "memory/allocation.hpp" 37 #include "utilities/compilerWarnings.hpp" 38 39 class ConcurrentGCTimer; 40 class G1ConcurrentMarkThread; 41 class G1CollectedHeap; 42 class G1CMOopClosure; 43 class G1CMTask; 44 class G1ConcurrentMark; 45 class G1OldTracer; 46 class G1RegionToSpaceMapper; 47 class G1SurvivorRegions; 48 class ThreadClosure; 49 50 PRAGMA_DIAG_PUSH 51 // warning C4522: multiple assignment operators specified 52 PRAGMA_DISABLE_MSVC_WARNING(4522) 53 54 // This is a container class for either an oop or a continuation address for 55 // mark stack entries. Both are pushed onto the mark stack. 56 class G1TaskQueueEntry { 57 private: 58 void* _holder; 59 60 static const uintptr_t ArraySliceBit = 1; 61 62 G1TaskQueueEntry(oop obj) : _holder(obj) { 63 assert(_holder != NULL, "Not allowed to set NULL task queue element"); 64 } 65 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { } 66 public: 67 G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; } 68 G1TaskQueueEntry() : _holder(NULL) { } 69 70 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); } 71 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); } 72 73 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) { 74 _holder = t._holder; 75 return *this; 76 } 77 78 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile { 79 _holder = t._holder; 80 return *this; 81 } 82 83 oop obj() const { 84 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder)); 85 return (oop)_holder; 86 } 87 88 HeapWord* slice() const { 89 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder)); 90 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit); 91 } 92 93 bool is_oop() const { return !is_array_slice(); } 94 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; } 95 bool is_null() const { return _holder == NULL; } 96 }; 97 98 PRAGMA_DIAG_POP 99 100 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue; 101 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet; 102 103 // Closure used by CM during concurrent reference discovery 104 // and reference processing (during remarking) to determine 105 // if a particular object is alive. It is primarily used 106 // to determine if referents of discovered reference objects 107 // are alive. An instance is also embedded into the 108 // reference processor as the _is_alive_non_header field 109 class G1CMIsAliveClosure : public BoolObjectClosure { 110 G1CollectedHeap* _g1h; 111 public: 112 G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { } 113 bool do_object_b(oop obj); 114 }; 115 116 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure { 117 G1CollectedHeap* _g1h; 118 public: 119 G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { } 120 bool do_object_b(oop obj); 121 }; 122 123 // Represents the overflow mark stack used by concurrent marking. 124 // 125 // Stores oops in a huge buffer in virtual memory that is always fully committed. 126 // Resizing may only happen during a STW pause when the stack is empty. 127 // 128 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark 129 // stack memory is split into evenly sized chunks of oops. Users can only 130 // add or remove entries on that basis. 131 // Chunks are filled in increasing address order. Not completely filled chunks 132 // have a NULL element as a terminating element. 133 // 134 // Every chunk has a header containing a single pointer element used for memory 135 // management. This wastes some space, but is negligible (< .1% with current sizing). 136 // 137 // Memory management is done using a mix of tracking a high water-mark indicating 138 // that all chunks at a lower address are valid chunks, and a singly linked free 139 // list connecting all empty chunks. 140 class G1CMMarkStack { 141 public: 142 // Number of TaskQueueEntries that can fit in a single chunk. 143 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */; 144 private: 145 struct TaskQueueEntryChunk { 146 TaskQueueEntryChunk* next; 147 G1TaskQueueEntry data[EntriesPerChunk]; 148 }; 149 150 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack. 151 152 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area. 153 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements. 154 155 char _pad0[DEFAULT_CACHE_LINE_SIZE]; 156 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users. 157 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)]; 158 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data. 159 volatile size_t _chunks_in_chunk_list; 160 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)]; 161 162 volatile size_t _hwm; // High water mark within the reserved space. 163 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)]; 164 165 // Allocate a new chunk from the reserved memory, using the high water mark. Returns 166 // NULL if out of memory. 167 TaskQueueEntryChunk* allocate_new_chunk(); 168 169 // Atomically add the given chunk to the list. 170 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem); 171 // Atomically remove and return a chunk from the given list. Returns NULL if the 172 // list is empty. 173 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list); 174 175 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem); 176 void add_chunk_to_free_list(TaskQueueEntryChunk* elem); 177 178 TaskQueueEntryChunk* remove_chunk_from_chunk_list(); 179 TaskQueueEntryChunk* remove_chunk_from_free_list(); 180 181 // Resizes the mark stack to the given new capacity. Releases any previous 182 // memory if successful. 183 bool resize(size_t new_capacity); 184 185 public: 186 G1CMMarkStack(); 187 ~G1CMMarkStack(); 188 189 // Alignment and minimum capacity of this mark stack in number of oops. 190 static size_t capacity_alignment(); 191 192 // Allocate and initialize the mark stack with the given number of oops. 193 bool initialize(size_t initial_capacity, size_t max_capacity); 194 195 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark 196 // stack. If less than EntriesPerChunk elements are to be pushed, the array must 197 // be terminated with a NULL. 198 // Returns whether the buffer contents were successfully pushed to the global mark 199 // stack. 200 bool par_push_chunk(G1TaskQueueEntry* buffer); 201 202 // Pops a chunk from this mark stack, copying them into the given buffer. This 203 // chunk may contain up to EntriesPerChunk elements. If there are less, the last 204 // element in the array is a NULL pointer. 205 bool par_pop_chunk(G1TaskQueueEntry* buffer); 206 207 // Return whether the chunk list is empty. Racy due to unsynchronized access to 208 // _chunk_list. 209 bool is_empty() const { return _chunk_list == NULL; } 210 211 size_t capacity() const { return _chunk_capacity; } 212 213 // Expand the stack, typically in response to an overflow condition 214 void expand(); 215 216 // Return the approximate number of oops on this mark stack. Racy due to 217 // unsynchronized access to _chunks_in_chunk_list. 218 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; } 219 220 void set_empty(); 221 222 // Apply Fn to every oop on the mark stack. The mark stack must not 223 // be modified while iterating. 224 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN; 225 }; 226 227 // Root MemRegions are memory areas that contain objects which references are 228 // roots wrt to the marking. They must be scanned before marking to maintain the 229 // SATB invariant. 230 // Typically they contain the areas from nTAMS to top of the regions. 231 // We could scan and mark through these objects during the initial-mark pause, but for 232 // pause time reasons we move this work to the concurrent phase. 233 // We need to complete this procedure before the next GC because it might determine 234 // that some of these "root objects" are dead, potentially dropping some required 235 // references. 236 // Root MemRegions comprise of the contents of survivor regions at the end 237 // of the GC, and any objects copied into the old gen during GC. 238 class G1CMRootMemRegions { 239 // The set of root MemRegions. 240 MemRegion* _root_regions; 241 size_t const _max_regions; 242 243 volatile size_t _num_root_regions; // Actual number of root regions. 244 245 volatile size_t _claimed_root_regions; // Number of root regions currently claimed. 246 247 volatile bool _scan_in_progress; 248 volatile bool _should_abort; 249 250 void notify_scan_done(); 251 252 public: 253 G1CMRootMemRegions(uint const max_regions); 254 ~G1CMRootMemRegions(); 255 256 // Reset the data structure to allow addition of new root regions. 257 void reset(); 258 259 void add(HeapWord* start, HeapWord* end); 260 261 // Reset the claiming / scanning of the root regions. 262 void prepare_for_scan(); 263 264 // Forces get_next() to return NULL so that the iteration aborts early. 265 void abort() { _should_abort = true; } 266 267 // Return true if the CM thread are actively scanning root regions, 268 // false otherwise. 269 bool scan_in_progress() { return _scan_in_progress; } 270 271 // Claim the next root MemRegion to scan atomically, or return NULL if 272 // all have been claimed. 273 const MemRegion* claim_next(); 274 275 // The number of root regions to scan. 276 uint num_root_regions() const; 277 278 void cancel_scan(); 279 280 // Flag that we're done with root region scanning and notify anyone 281 // who's waiting on it. If aborted is false, assume that all regions 282 // have been claimed. 283 void scan_finished(); 284 285 // If CM threads are still scanning root regions, wait until they 286 // are done. Return true if we had to wait, false otherwise. 287 bool wait_until_scan_finished(); 288 }; 289 290 // This class manages data structures and methods for doing liveness analysis in 291 // G1's concurrent cycle. 292 class G1ConcurrentMark : public CHeapObj<mtGC> { 293 friend class G1ConcurrentMarkThread; 294 friend class G1CMRefProcTaskProxy; 295 friend class G1CMRefProcTaskExecutor; 296 friend class G1CMKeepAliveAndDrainClosure; 297 friend class G1CMDrainMarkingStackClosure; 298 friend class G1CMBitMapClosure; 299 friend class G1CMConcurrentMarkingTask; 300 friend class G1CMRemarkTask; 301 friend class G1CMTask; 302 303 G1ConcurrentMarkThread* _cm_thread; // The thread doing the work 304 G1CollectedHeap* _g1h; // The heap 305 bool _completed_initialization; // Set to true when initialization is complete 306 307 // Concurrent marking support structures 308 G1CMBitMap _mark_bitmap_1; 309 G1CMBitMap _mark_bitmap_2; 310 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap 311 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap 312 313 // Heap bounds 314 MemRegion const _heap; 315 316 // Root region tracking and claiming 317 G1CMRootMemRegions _root_regions; 318 319 // For grey objects 320 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger 321 HeapWord* volatile _finger; // The global finger, region aligned, 322 // always pointing to the end of the 323 // last claimed region 324 325 uint _worker_id_offset; 326 uint _max_num_tasks; // Maximum number of marking tasks 327 uint _num_active_tasks; // Number of tasks currently active 328 G1CMTask** _tasks; // Task queue array (max_worker_id length) 329 330 G1CMTaskQueueSet* _task_queues; // Task queue set 331 TaskTerminator _terminator; // For termination 332 333 // Two sync barriers that are used to synchronize tasks when an 334 // overflow occurs. The algorithm is the following. All tasks enter 335 // the first one to ensure that they have all stopped manipulating 336 // the global data structures. After they exit it, they re-initialize 337 // their data structures and task 0 re-initializes the global data 338 // structures. Then, they enter the second sync barrier. This 339 // ensure, that no task starts doing work before all data 340 // structures (local and global) have been re-initialized. When they 341 // exit it, they are free to start working again. 342 WorkGangBarrierSync _first_overflow_barrier_sync; 343 WorkGangBarrierSync _second_overflow_barrier_sync; 344 345 // This is set by any task, when an overflow on the global data 346 // structures is detected 347 volatile bool _has_overflown; 348 // True: marking is concurrent, false: we're in remark 349 volatile bool _concurrent; 350 // Set at the end of a Full GC so that marking aborts 351 volatile bool _has_aborted; 352 353 // Used when remark aborts due to an overflow to indicate that 354 // another concurrent marking phase should start 355 volatile bool _restart_for_overflow; 356 357 ConcurrentGCTimer* _gc_timer_cm; 358 359 G1OldTracer* _gc_tracer_cm; 360 361 // Timing statistics. All of them are in ms 362 NumberSeq _init_times; 363 NumberSeq _remark_times; 364 NumberSeq _remark_mark_times; 365 NumberSeq _remark_weak_ref_times; 366 NumberSeq _cleanup_times; 367 double _total_cleanup_time; 368 369 double* _accum_task_vtime; // Accumulated task vtime 370 371 WorkGang* _concurrent_workers; 372 uint _num_concurrent_workers; // The number of marking worker threads we're using 373 uint _max_concurrent_workers; // Maximum number of marking worker threads 374 375 void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller); 376 377 void finalize_marking(); 378 379 void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes); 380 void weak_refs_work(bool clear_all_soft_refs); 381 382 void report_object_count(bool mark_completed); 383 384 void swap_mark_bitmaps(); 385 386 void reclaim_empty_regions(); 387 388 // After reclaiming empty regions, update heap sizes. 389 void compute_new_sizes(); 390 391 // Clear statistics gathered during the concurrent cycle for the given region after 392 // it has been reclaimed. 393 void clear_statistics(HeapRegion* r); 394 395 // Resets the global marking data structures, as well as the 396 // task local ones; should be called during initial mark. 397 void reset(); 398 399 // Resets all the marking data structures. Called when we have to restart 400 // marking or when marking completes (via set_non_marking_state below). 401 void reset_marking_for_restart(); 402 403 // We do this after we're done with marking so that the marking data 404 // structures are initialized to a sensible and predictable state. 405 void reset_at_marking_complete(); 406 407 // Called to indicate how many threads are currently active. 408 void set_concurrency(uint active_tasks); 409 410 // Should be called to indicate which phase we're in (concurrent 411 // mark or remark) and how many threads are currently active. 412 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 413 414 // Prints all gathered CM-related statistics 415 void print_stats(); 416 417 HeapWord* finger() { return _finger; } 418 bool concurrent() { return _concurrent; } 419 uint active_tasks() { return _num_active_tasks; } 420 ParallelTaskTerminator* terminator() const { return _terminator.terminator(); } 421 422 // Claims the next available region to be scanned by a marking 423 // task/thread. It might return NULL if the next region is empty or 424 // we have run out of regions. In the latter case, out_of_regions() 425 // determines whether we've really run out of regions or the task 426 // should call claim_region() again. This might seem a bit 427 // awkward. Originally, the code was written so that claim_region() 428 // either successfully returned with a non-empty region or there 429 // were no more regions to be claimed. The problem with this was 430 // that, in certain circumstances, it iterated over large chunks of 431 // the heap finding only empty regions and, while it was working, it 432 // was preventing the calling task to call its regular clock 433 // method. So, this way, each task will spend very little time in 434 // claim_region() and is allowed to call the regular clock method 435 // frequently. 436 HeapRegion* claim_region(uint worker_id); 437 438 // Determines whether we've run out of regions to scan. Note that 439 // the finger can point past the heap end in case the heap was expanded 440 // to satisfy an allocation without doing a GC. This is fine, because all 441 // objects in those regions will be considered live anyway because of 442 // SATB guarantees (i.e. their TAMS will be equal to bottom). 443 bool out_of_regions() { return _finger >= _heap.end(); } 444 445 // Returns the task with the given id 446 G1CMTask* task(uint id) { 447 // During initial mark we use the parallel gc threads to do some work, so 448 // we can only compare against _max_num_tasks. 449 assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks); 450 return _tasks[id]; 451 } 452 453 // Access / manipulation of the overflow flag which is set to 454 // indicate that the global stack has overflown 455 bool has_overflown() { return _has_overflown; } 456 void set_has_overflown() { _has_overflown = true; } 457 void clear_has_overflown() { _has_overflown = false; } 458 bool restart_for_overflow() { return _restart_for_overflow; } 459 460 // Methods to enter the two overflow sync barriers 461 void enter_first_sync_barrier(uint worker_id); 462 void enter_second_sync_barrier(uint worker_id); 463 464 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is 465 // true, periodically insert checks to see if this method should exit prematurely. 466 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield); 467 468 // Region statistics gathered during marking. 469 G1RegionMarkStats* _region_mark_stats; 470 // Top pointer for each region at the start of the rebuild remembered set process 471 // for regions which remembered sets need to be rebuilt. A NULL for a given region 472 // means that this region does not be scanned during the rebuilding remembered 473 // set phase at all. 474 HeapWord* volatile* _top_at_rebuild_starts; 475 public: 476 void add_to_liveness(uint worker_id, oop const obj, size_t size); 477 // Liveness of the given region as determined by concurrent marking, i.e. the amount of 478 // live words between bottom and nTAMS. 479 size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; } 480 481 // Sets the internal top_at_region_start for the given region to current top of the region. 482 inline void update_top_at_rebuild_start(HeapRegion* r); 483 // TARS for the given region during remembered set rebuilding. 484 inline HeapWord* top_at_rebuild_start(uint region) const; 485 486 // Clear statistics gathered during the concurrent cycle for the given region after 487 // it has been reclaimed. 488 void clear_statistics_in_region(uint region_idx); 489 // Notification for eagerly reclaimed regions to clean up. 490 void humongous_object_eagerly_reclaimed(HeapRegion* r); 491 // Manipulation of the global mark stack. 492 // The push and pop operations are used by tasks for transfers 493 // between task-local queues and the global mark stack. 494 bool mark_stack_push(G1TaskQueueEntry* arr) { 495 if (!_global_mark_stack.par_push_chunk(arr)) { 496 set_has_overflown(); 497 return false; 498 } 499 return true; 500 } 501 bool mark_stack_pop(G1TaskQueueEntry* arr) { 502 return _global_mark_stack.par_pop_chunk(arr); 503 } 504 size_t mark_stack_size() const { return _global_mark_stack.size(); } 505 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; } 506 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); } 507 508 G1CMRootMemRegions* root_regions() { return &_root_regions; } 509 510 void concurrent_cycle_start(); 511 // Abandon current marking iteration due to a Full GC. 512 void concurrent_cycle_abort(); 513 void concurrent_cycle_end(); 514 515 void update_accum_task_vtime(int i, double vtime) { 516 _accum_task_vtime[i] += vtime; 517 } 518 519 double all_task_accum_vtime() { 520 double ret = 0.0; 521 for (uint i = 0; i < _max_num_tasks; ++i) 522 ret += _accum_task_vtime[i]; 523 return ret; 524 } 525 526 // Attempts to steal an object from the task queues of other tasks 527 bool try_stealing(uint worker_id, G1TaskQueueEntry& task_entry); 528 529 G1ConcurrentMark(G1CollectedHeap* g1h, 530 G1RegionToSpaceMapper* prev_bitmap_storage, 531 G1RegionToSpaceMapper* next_bitmap_storage); 532 ~G1ConcurrentMark(); 533 534 G1ConcurrentMarkThread* cm_thread() { return _cm_thread; } 535 536 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; } 537 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; } 538 539 // Calculates the number of concurrent GC threads to be used in the marking phase. 540 uint calc_active_marking_workers(); 541 542 // Moves all per-task cached data into global state. 543 void flush_all_task_caches(); 544 // Prepare internal data structures for the next mark cycle. This includes clearing 545 // the next mark bitmap and some internal data structures. This method is intended 546 // to be called concurrently to the mutator. It will yield to safepoint requests. 547 void cleanup_for_next_mark(); 548 549 // Clear the previous marking bitmap during safepoint. 550 void clear_prev_bitmap(WorkGang* workers); 551 552 // These two methods do the work that needs to be done at the start and end of the 553 // initial mark pause. 554 void pre_initial_mark(); 555 void post_initial_mark(); 556 557 // Scan all the root regions and mark everything reachable from 558 // them. 559 void scan_root_regions(); 560 561 // Scan a single root MemRegion to mark everything reachable from it. 562 void scan_root_region(const MemRegion* region, uint worker_id); 563 564 // Do concurrent phase of marking, to a tentative transitive closure. 565 void mark_from_roots(); 566 567 // Do concurrent preclean work. 568 void preclean(); 569 570 void remark(); 571 572 void cleanup(); 573 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use 574 // this carefully. 575 inline void mark_in_prev_bitmap(oop p); 576 577 // Clears marks for all objects in the given range, for the prev or 578 // next bitmaps. Caution: the previous bitmap is usually 579 // read-only, so use this carefully! 580 void clear_range_in_prev_bitmap(MemRegion mr); 581 582 inline bool is_marked_in_prev_bitmap(oop p) const; 583 584 // Verify that there are no collection set oops on the stacks (taskqueues / 585 // global mark stack) and fingers (global / per-task). 586 // If marking is not in progress, it's a no-op. 587 void verify_no_collection_set_oops() PRODUCT_RETURN; 588 589 inline bool do_yield_check(); 590 591 bool has_aborted() { return _has_aborted; } 592 593 void print_summary_info(); 594 595 void print_worker_threads_on(outputStream* st) const; 596 void threads_do(ThreadClosure* tc) const; 597 598 void print_on_error(outputStream* st) const; 599 600 // Mark the given object on the next bitmap if it is below nTAMS. 601 inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj); 602 inline bool mark_in_next_bitmap(uint worker_id, oop const obj); 603 604 inline bool is_marked_in_next_bitmap(oop p) const; 605 606 // Returns true if initialization was successfully completed. 607 bool completed_initialization() const { 608 return _completed_initialization; 609 } 610 611 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; } 612 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; } 613 614 private: 615 // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application. 616 void rebuild_rem_set_concurrently(); 617 }; 618 619 // A class representing a marking task. 620 class G1CMTask : public TerminatorTerminator { 621 private: 622 enum PrivateConstants { 623 // The regular clock call is called once the scanned words reaches 624 // this limit 625 words_scanned_period = 12*1024, 626 // The regular clock call is called once the number of visited 627 // references reaches this limit 628 refs_reached_period = 1024, 629 // Initial value for the hash seed, used in the work stealing code 630 init_hash_seed = 17 631 }; 632 633 // Number of entries in the per-task stats entry. This seems enough to have a very 634 // low cache miss rate. 635 static const uint RegionMarkStatsCacheSize = 1024; 636 637 G1CMObjArrayProcessor _objArray_processor; 638 639 uint _worker_id; 640 G1CollectedHeap* _g1h; 641 G1ConcurrentMark* _cm; 642 G1CMBitMap* _next_mark_bitmap; 643 // the task queue of this task 644 G1CMTaskQueue* _task_queue; 645 646 G1RegionMarkStatsCache _mark_stats_cache; 647 // Number of calls to this task 648 uint _calls; 649 650 // When the virtual timer reaches this time, the marking step should exit 651 double _time_target_ms; 652 // Start time of the current marking step 653 double _start_time_ms; 654 655 // Oop closure used for iterations over oops 656 G1CMOopClosure* _cm_oop_closure; 657 658 // Region this task is scanning, NULL if we're not scanning any 659 HeapRegion* _curr_region; 660 // Local finger of this task, NULL if we're not scanning a region 661 HeapWord* _finger; 662 // Limit of the region this task is scanning, NULL if we're not scanning one 663 HeapWord* _region_limit; 664 665 // Number of words this task has scanned 666 size_t _words_scanned; 667 // When _words_scanned reaches this limit, the regular clock is 668 // called. Notice that this might be decreased under certain 669 // circumstances (i.e. when we believe that we did an expensive 670 // operation). 671 size_t _words_scanned_limit; 672 // Initial value of _words_scanned_limit (i.e. what it was 673 // before it was decreased). 674 size_t _real_words_scanned_limit; 675 676 // Number of references this task has visited 677 size_t _refs_reached; 678 // When _refs_reached reaches this limit, the regular clock is 679 // called. Notice this this might be decreased under certain 680 // circumstances (i.e. when we believe that we did an expensive 681 // operation). 682 size_t _refs_reached_limit; 683 // Initial value of _refs_reached_limit (i.e. what it was before 684 // it was decreased). 685 size_t _real_refs_reached_limit; 686 687 // If true, then the task has aborted for some reason 688 bool _has_aborted; 689 // Set when the task aborts because it has met its time quota 690 bool _has_timed_out; 691 // True when we're draining SATB buffers; this avoids the task 692 // aborting due to SATB buffers being available (as we're already 693 // dealing with them) 694 bool _draining_satb_buffers; 695 696 // Number sequence of past step times 697 NumberSeq _step_times_ms; 698 // Elapsed time of this task 699 double _elapsed_time_ms; 700 // Termination time of this task 701 double _termination_time_ms; 702 // When this task got into the termination protocol 703 double _termination_start_time_ms; 704 705 TruncatedSeq _marking_step_diff_ms; 706 707 // Updates the local fields after this task has claimed 708 // a new region to scan 709 void setup_for_region(HeapRegion* hr); 710 // Makes the limit of the region up-to-date 711 void update_region_limit(); 712 713 // Called when either the words scanned or the refs visited limit 714 // has been reached 715 void reached_limit(); 716 // Recalculates the words scanned and refs visited limits 717 void recalculate_limits(); 718 // Decreases the words scanned and refs visited limits when we reach 719 // an expensive operation 720 void decrease_limits(); 721 // Checks whether the words scanned or refs visited reached their 722 // respective limit and calls reached_limit() if they have 723 void check_limits() { 724 if (_words_scanned >= _words_scanned_limit || 725 _refs_reached >= _refs_reached_limit) { 726 reached_limit(); 727 } 728 } 729 // Supposed to be called regularly during a marking step as 730 // it checks a bunch of conditions that might cause the marking step 731 // to abort 732 // Return true if the marking step should continue. Otherwise, return false to abort 733 bool regular_clock_call(); 734 735 // Set abort flag if regular_clock_call() check fails 736 inline void abort_marking_if_regular_check_fail(); 737 738 // Test whether obj might have already been passed over by the 739 // mark bitmap scan, and so needs to be pushed onto the mark stack. 740 bool is_below_finger(oop obj, HeapWord* global_finger) const; 741 742 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry); 743 public: 744 // Apply the closure on the given area of the objArray. Return the number of words 745 // scanned. 746 inline size_t scan_objArray(objArrayOop obj, MemRegion mr); 747 // Resets the task; should be called right at the beginning of a marking phase. 748 void reset(G1CMBitMap* next_mark_bitmap); 749 // Clears all the fields that correspond to a claimed region. 750 void clear_region_fields(); 751 752 // The main method of this class which performs a marking step 753 // trying not to exceed the given duration. However, it might exit 754 // prematurely, according to some conditions (i.e. SATB buffers are 755 // available for processing). 756 void do_marking_step(double target_ms, 757 bool do_termination, 758 bool is_serial); 759 760 // These two calls start and stop the timer 761 void record_start_time() { 762 _elapsed_time_ms = os::elapsedTime() * 1000.0; 763 } 764 void record_end_time() { 765 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 766 } 767 768 // Returns the worker ID associated with this task. 769 uint worker_id() { return _worker_id; } 770 771 // From TerminatorTerminator. It determines whether this task should 772 // exit the termination protocol after it's entered it. 773 virtual bool should_exit_termination(); 774 775 // Resets the local region fields after a task has finished scanning a 776 // region; or when they have become stale as a result of the region 777 // being evacuated. 778 void giveup_current_region(); 779 780 HeapWord* finger() { return _finger; } 781 782 bool has_aborted() { return _has_aborted; } 783 void set_has_aborted() { _has_aborted = true; } 784 void clear_has_aborted() { _has_aborted = false; } 785 786 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 787 788 // Increment the number of references this task has visited. 789 void increment_refs_reached() { ++_refs_reached; } 790 791 // Grey the object by marking it. If not already marked, push it on 792 // the local queue if below the finger. obj is required to be below its region's NTAMS. 793 // Returns whether there has been a mark to the bitmap. 794 inline bool make_reference_grey(oop obj); 795 796 // Grey the object (by calling make_grey_reference) if required, 797 // e.g. obj is below its containing region's NTAMS. 798 // Precondition: obj is a valid heap object. 799 // Returns true if the reference caused a mark to be set in the next bitmap. 800 template <class T> 801 inline bool deal_with_reference(T* p); 802 803 // Scans an object and visits its children. 804 inline void scan_task_entry(G1TaskQueueEntry task_entry); 805 806 // Pushes an object on the local queue. 807 inline void push(G1TaskQueueEntry task_entry); 808 809 // Move entries to the global stack. 810 void move_entries_to_global_stack(); 811 // Move entries from the global stack, return true if we were successful to do so. 812 bool get_entries_from_global_stack(); 813 814 // Pops and scans objects from the local queue. If partially is 815 // true, then it stops when the queue size is of a given limit. If 816 // partially is false, then it stops when the queue is empty. 817 void drain_local_queue(bool partially); 818 // Moves entries from the global stack to the local queue and 819 // drains the local queue. If partially is true, then it stops when 820 // both the global stack and the local queue reach a given size. If 821 // partially if false, it tries to empty them totally. 822 void drain_global_stack(bool partially); 823 // Keeps picking SATB buffers and processing them until no SATB 824 // buffers are available. 825 void drain_satb_buffers(); 826 827 // Moves the local finger to a new location 828 inline void move_finger_to(HeapWord* new_finger) { 829 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 830 _finger = new_finger; 831 } 832 833 G1CMTask(uint worker_id, 834 G1ConcurrentMark *cm, 835 G1CMTaskQueue* task_queue, 836 G1RegionMarkStats* mark_stats, 837 uint max_regions); 838 839 inline void update_liveness(oop const obj, size_t const obj_size); 840 841 // Clear (without flushing) the mark cache entry for the given region. 842 void clear_mark_stats_cache(uint region_idx); 843 // Evict the whole statistics cache into the global statistics. Returns the 844 // number of cache hits and misses so far. 845 Pair<size_t, size_t> flush_mark_stats_cache(); 846 // Prints statistics associated with this task 847 void print_stats(); 848 }; 849 850 // Class that's used to to print out per-region liveness 851 // information. It's currently used at the end of marking and also 852 // after we sort the old regions at the end of the cleanup operation. 853 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure { 854 // Accumulators for these values. 855 size_t _total_used_bytes; 856 size_t _total_capacity_bytes; 857 size_t _total_prev_live_bytes; 858 size_t _total_next_live_bytes; 859 860 // Accumulator for the remembered set size 861 size_t _total_remset_bytes; 862 863 // Accumulator for strong code roots memory size 864 size_t _total_strong_code_roots_bytes; 865 866 static double bytes_to_mb(size_t val) { 867 return (double) val / (double) M; 868 } 869 870 public: 871 // The header and footer are printed in the constructor and 872 // destructor respectively. 873 G1PrintRegionLivenessInfoClosure(const char* phase_name); 874 virtual bool do_heap_region(HeapRegion* r); 875 ~G1PrintRegionLivenessInfoClosure(); 876 }; 877 878 #endif // SHARE_GC_G1_G1CONCURRENTMARK_HPP