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/g1HeapVerifier.hpp" 31 #include "gc/g1/g1RegionMarkStatsCache.hpp" 32 #include "gc/g1/heapRegionSet.hpp" 33 #include "gc/shared/taskqueue.hpp" 34 #include "memory/allocation.hpp" 35 36 class ConcurrentGCTimer; 37 class G1ConcurrentMarkThread; 38 class G1CollectedHeap; 39 class G1CMTask; 40 class G1ConcurrentMark; 41 class G1OldTracer; 42 class G1RegionToSpaceMapper; 43 class G1SurvivorRegions; 44 45 #ifdef _MSC_VER 46 #pragma warning(push) 47 // warning C4522: multiple assignment operators specified 48 #pragma warning(disable:4522) 49 #endif 50 51 // This is a container class for either an oop or a continuation address for 52 // mark stack entries. Both are pushed onto the mark stack. 53 class G1TaskQueueEntry { 54 private: 55 void* _holder; 56 57 static const uintptr_t ArraySliceBit = 1; 58 59 G1TaskQueueEntry(oop obj) : _holder(obj) { 60 assert(_holder != NULL, "Not allowed to set NULL task queue element"); 61 } 62 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { } 63 public: 64 G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; } 65 G1TaskQueueEntry() : _holder(NULL) { } 66 67 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); } 68 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); } 69 70 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) { 71 _holder = t._holder; 72 return *this; 73 } 74 75 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile { 76 _holder = t._holder; 77 return *this; 78 } 79 80 oop obj() const { 81 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder)); 82 return (oop)_holder; 83 } 84 85 HeapWord* slice() const { 86 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder)); 87 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit); 88 } 89 90 bool is_oop() const { return !is_array_slice(); } 91 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; } 92 bool is_null() const { return _holder == NULL; } 93 }; 94 95 #ifdef _MSC_VER 96 #pragma warning(pop) 97 #endif 98 99 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue; 100 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet; 101 102 // Closure used by CM during concurrent reference discovery 103 // and reference processing (during remarking) to determine 104 // if a particular object is alive. It is primarily used 105 // to determine if referents of discovered reference objects 106 // are alive. An instance is also embedded into the 107 // reference processor as the _is_alive_non_header field 108 class G1CMIsAliveClosure : public BoolObjectClosure { 109 G1CollectedHeap* _g1h; 110 public: 111 G1CMIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { } 112 bool do_object_b(oop obj); 113 }; 114 115 class G1CMSubjectToDiscoveryClosure : public BoolObjectClosure { 116 G1CollectedHeap* _g1h; 117 public: 118 G1CMSubjectToDiscoveryClosure(G1CollectedHeap* g1h) : _g1h(g1h) { } 119 bool do_object_b(oop obj); 120 }; 121 122 // Represents the overflow mark stack used by concurrent marking. 123 // 124 // Stores oops in a huge buffer in virtual memory that is always fully committed. 125 // Resizing may only happen during a STW pause when the stack is empty. 126 // 127 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark 128 // stack memory is split into evenly sized chunks of oops. Users can only 129 // add or remove entries on that basis. 130 // Chunks are filled in increasing address order. Not completely filled chunks 131 // have a NULL element as a terminating element. 132 // 133 // Every chunk has a header containing a single pointer element used for memory 134 // management. This wastes some space, but is negligible (< .1% with current sizing). 135 // 136 // Memory management is done using a mix of tracking a high water-mark indicating 137 // that all chunks at a lower address are valid chunks, and a singly linked free 138 // list connecting all empty chunks. 139 class G1CMMarkStack { 140 public: 141 // Number of TaskQueueEntries that can fit in a single chunk. 142 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */; 143 private: 144 struct TaskQueueEntryChunk { 145 TaskQueueEntryChunk* next; 146 G1TaskQueueEntry data[EntriesPerChunk]; 147 }; 148 149 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack. 150 151 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area. 152 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements. 153 154 char _pad0[DEFAULT_CACHE_LINE_SIZE]; 155 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users. 156 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)]; 157 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data. 158 volatile size_t _chunks_in_chunk_list; 159 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)]; 160 161 volatile size_t _hwm; // High water mark within the reserved space. 162 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)]; 163 164 // Allocate a new chunk from the reserved memory, using the high water mark. Returns 165 // NULL if out of memory. 166 TaskQueueEntryChunk* allocate_new_chunk(); 167 168 // Atomically add the given chunk to the list. 169 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem); 170 // Atomically remove and return a chunk from the given list. Returns NULL if the 171 // list is empty. 172 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list); 173 174 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem); 175 void add_chunk_to_free_list(TaskQueueEntryChunk* elem); 176 177 TaskQueueEntryChunk* remove_chunk_from_chunk_list(); 178 TaskQueueEntryChunk* remove_chunk_from_free_list(); 179 180 // Resizes the mark stack to the given new capacity. Releases any previous 181 // memory if successful. 182 bool resize(size_t new_capacity); 183 184 public: 185 G1CMMarkStack(); 186 ~G1CMMarkStack(); 187 188 // Alignment and minimum capacity of this mark stack in number of oops. 189 static size_t capacity_alignment(); 190 191 // Allocate and initialize the mark stack with the given number of oops. 192 bool initialize(size_t initial_capacity, size_t max_capacity); 193 194 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark 195 // stack. If less than EntriesPerChunk elements are to be pushed, the array must 196 // be terminated with a NULL. 197 // Returns whether the buffer contents were successfully pushed to the global mark 198 // stack. 199 bool par_push_chunk(G1TaskQueueEntry* buffer); 200 201 // Pops a chunk from this mark stack, copying them into the given buffer. This 202 // chunk may contain up to EntriesPerChunk elements. If there are less, the last 203 // element in the array is a NULL pointer. 204 bool par_pop_chunk(G1TaskQueueEntry* buffer); 205 206 // Return whether the chunk list is empty. Racy due to unsynchronized access to 207 // _chunk_list. 208 bool is_empty() const { return _chunk_list == NULL; } 209 210 size_t capacity() const { return _chunk_capacity; } 211 212 // Expand the stack, typically in response to an overflow condition 213 void expand(); 214 215 // Return the approximate number of oops on this mark stack. Racy due to 216 // unsynchronized access to _chunks_in_chunk_list. 217 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; } 218 219 void set_empty(); 220 221 // Apply Fn to every oop on the mark stack. The mark stack must not 222 // be modified while iterating. 223 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN; 224 }; 225 226 // Root Regions are regions that are not empty at the beginning of a 227 // marking cycle and which we might collect during an evacuation pause 228 // while the cycle is active. Given that, during evacuation pauses, we 229 // do not copy objects that are explicitly marked, what we have to do 230 // for the root regions is to scan them and mark all objects reachable 231 // from them. According to the SATB assumptions, we only need to visit 232 // each object once during marking. So, as long as we finish this scan 233 // before the next evacuation pause, we can copy the objects from the 234 // root regions without having to mark them or do anything else to them. 235 // 236 // Currently, we only support root region scanning once (at the start 237 // of the marking cycle) and the root regions are all the survivor 238 // regions populated during the initial-mark pause. 239 class G1CMRootRegions { 240 private: 241 const G1SurvivorRegions* _survivors; 242 G1ConcurrentMark* _cm; 243 244 volatile bool _scan_in_progress; 245 volatile bool _should_abort; 246 volatile int _claimed_survivor_index; 247 248 void notify_scan_done(); 249 250 public: 251 G1CMRootRegions(); 252 // We actually do most of the initialization in this method. 253 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm); 254 255 // Reset the claiming / scanning of the root regions. 256 void prepare_for_scan(); 257 258 // Forces get_next() to return NULL so that the iteration aborts early. 259 void abort() { _should_abort = true; } 260 261 // Return true if the CM thread are actively scanning root regions, 262 // false otherwise. 263 bool scan_in_progress() { return _scan_in_progress; } 264 265 // Claim the next root region to scan atomically, or return NULL if 266 // all have been claimed. 267 HeapRegion* claim_next(); 268 269 // The number of root regions to scan. 270 uint num_root_regions() const; 271 272 void cancel_scan(); 273 274 // Flag that we're done with root region scanning and notify anyone 275 // who's waiting on it. If aborted is false, assume that all regions 276 // have been claimed. 277 void scan_finished(); 278 279 // If CM threads are still scanning root regions, wait until they 280 // are done. Return true if we had to wait, false otherwise. 281 bool wait_until_scan_finished(); 282 }; 283 284 // This class manages data structures and methods for doing liveness analysis in 285 // G1's concurrent cycle. 286 class G1ConcurrentMark : public CHeapObj<mtGC> { 287 friend class G1ConcurrentMarkThread; 288 friend class G1CMRefProcTaskProxy; 289 friend class G1CMRefProcTaskExecutor; 290 friend class G1CMKeepAliveAndDrainClosure; 291 friend class G1CMDrainMarkingStackClosure; 292 friend class G1CMBitMapClosure; 293 friend class G1CMConcurrentMarkingTask; 294 friend class G1CMRemarkTask; 295 friend class G1CMTask; 296 297 G1ConcurrentMarkThread* _cm_thread; // The thread doing the work 298 G1CollectedHeap* _g1h; // The heap 299 bool _completed_initialization; // Set to true when initialization is complete 300 301 // Concurrent marking support structures 302 G1CMBitMap _mark_bitmap_1; 303 G1CMBitMap _mark_bitmap_2; 304 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap 305 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap 306 307 // Heap bounds 308 MemRegion const _heap; 309 310 // Root region tracking and claiming 311 G1CMRootRegions _root_regions; 312 313 // For grey objects 314 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger 315 HeapWord* volatile _finger; // The global finger, region aligned, 316 // always pointing to the end of the 317 // last claimed region 318 319 uint _worker_id_offset; 320 uint _max_num_tasks; // Maximum number of marking tasks 321 uint _num_active_tasks; // Number of tasks currently active 322 G1CMTask** _tasks; // Task queue array (max_worker_id length) 323 324 G1CMTaskQueueSet* _task_queues; // Task queue set 325 ParallelTaskTerminator _terminator; // For termination 326 327 // Two sync barriers that are used to synchronize tasks when an 328 // overflow occurs. The algorithm is the following. All tasks enter 329 // the first one to ensure that they have all stopped manipulating 330 // the global data structures. After they exit it, they re-initialize 331 // their data structures and task 0 re-initializes the global data 332 // structures. Then, they enter the second sync barrier. This 333 // ensure, that no task starts doing work before all data 334 // structures (local and global) have been re-initialized. When they 335 // exit it, they are free to start working again. 336 WorkGangBarrierSync _first_overflow_barrier_sync; 337 WorkGangBarrierSync _second_overflow_barrier_sync; 338 339 // This is set by any task, when an overflow on the global data 340 // structures is detected 341 volatile bool _has_overflown; 342 // True: marking is concurrent, false: we're in remark 343 volatile bool _concurrent; 344 // Set at the end of a Full GC so that marking aborts 345 volatile bool _has_aborted; 346 347 // Used when remark aborts due to an overflow to indicate that 348 // another concurrent marking phase should start 349 volatile bool _restart_for_overflow; 350 351 ConcurrentGCTimer* _gc_timer_cm; 352 353 G1OldTracer* _gc_tracer_cm; 354 355 // Timing statistics. All of them are in ms 356 NumberSeq _init_times; 357 NumberSeq _remark_times; 358 NumberSeq _remark_mark_times; 359 NumberSeq _remark_weak_ref_times; 360 NumberSeq _cleanup_times; 361 double _total_cleanup_time; 362 363 double* _accum_task_vtime; // Accumulated task vtime 364 365 WorkGang* _concurrent_workers; 366 uint _num_concurrent_workers; // The number of marking worker threads we're using 367 uint _max_concurrent_workers; // Maximum number of marking worker threads 368 369 void verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller); 370 371 void finalize_marking(); 372 373 void weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes); 374 void weak_refs_work(bool clear_all_soft_refs); 375 376 void report_object_count(bool mark_completed); 377 378 void swap_mark_bitmaps(); 379 380 void reclaim_empty_regions(); 381 382 // After reclaiming empty regions, update heap sizes. 383 void compute_new_sizes(); 384 385 // Clear statistics gathered during the concurrent cycle for the given region after 386 // it has been reclaimed. 387 void clear_statistics(HeapRegion* r); 388 389 // Resets the global marking data structures, as well as the 390 // task local ones; should be called during initial mark. 391 void reset(); 392 393 // Resets all the marking data structures. Called when we have to restart 394 // marking or when marking completes (via set_non_marking_state below). 395 void reset_marking_for_restart(); 396 397 // We do this after we're done with marking so that the marking data 398 // structures are initialized to a sensible and predictable state. 399 void reset_at_marking_complete(); 400 401 // Called to indicate how many threads are currently active. 402 void set_concurrency(uint active_tasks); 403 404 // Should be called to indicate which phase we're in (concurrent 405 // mark or remark) and how many threads are currently active. 406 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 407 408 // Prints all gathered CM-related statistics 409 void print_stats(); 410 411 HeapWord* finger() { return _finger; } 412 bool concurrent() { return _concurrent; } 413 uint active_tasks() { return _num_active_tasks; } 414 ParallelTaskTerminator* terminator() { return &_terminator; } 415 416 // Claims the next available region to be scanned by a marking 417 // task/thread. It might return NULL if the next region is empty or 418 // we have run out of regions. In the latter case, out_of_regions() 419 // determines whether we've really run out of regions or the task 420 // should call claim_region() again. This might seem a bit 421 // awkward. Originally, the code was written so that claim_region() 422 // either successfully returned with a non-empty region or there 423 // were no more regions to be claimed. The problem with this was 424 // that, in certain circumstances, it iterated over large chunks of 425 // the heap finding only empty regions and, while it was working, it 426 // was preventing the calling task to call its regular clock 427 // method. So, this way, each task will spend very little time in 428 // claim_region() and is allowed to call the regular clock method 429 // frequently. 430 HeapRegion* claim_region(uint worker_id); 431 432 // Determines whether we've run out of regions to scan. Note that 433 // the finger can point past the heap end in case the heap was expanded 434 // to satisfy an allocation without doing a GC. This is fine, because all 435 // objects in those regions will be considered live anyway because of 436 // SATB guarantees (i.e. their TAMS will be equal to bottom). 437 bool out_of_regions() { return _finger >= _heap.end(); } 438 439 // Returns the task with the given id 440 G1CMTask* task(uint id) { 441 // During initial mark we use the parallel gc threads to do some work, so 442 // we can only compare against _max_num_tasks. 443 assert(id < _max_num_tasks, "Task id %u not within bounds up to %u", id, _max_num_tasks); 444 return _tasks[id]; 445 } 446 447 // Access / manipulation of the overflow flag which is set to 448 // indicate that the global stack has overflown 449 bool has_overflown() { return _has_overflown; } 450 void set_has_overflown() { _has_overflown = true; } 451 void clear_has_overflown() { _has_overflown = false; } 452 bool restart_for_overflow() { return _restart_for_overflow; } 453 454 // Methods to enter the two overflow sync barriers 455 void enter_first_sync_barrier(uint worker_id); 456 void enter_second_sync_barrier(uint worker_id); 457 458 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is 459 // true, periodically insert checks to see if this method should exit prematurely. 460 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield); 461 462 // Region statistics gathered during marking. 463 G1RegionMarkStats* _region_mark_stats; 464 // Top pointer for each region at the start of the rebuild remembered set process 465 // for regions which remembered sets need to be rebuilt. A NULL for a given region 466 // means that this region does not be scanned during the rebuilding remembered 467 // set phase at all. 468 HeapWord* volatile* _top_at_rebuild_starts; 469 public: 470 void add_to_liveness(uint worker_id, oop const obj, size_t size); 471 // Liveness of the given region as determined by concurrent marking, i.e. the amount of 472 // live words between bottom and nTAMS. 473 size_t liveness(uint region) const { return _region_mark_stats[region]._live_words; } 474 475 // Sets the internal top_at_region_start for the given region to current top of the region. 476 inline void update_top_at_rebuild_start(HeapRegion* r); 477 // TARS for the given region during remembered set rebuilding. 478 inline HeapWord* top_at_rebuild_start(uint region) const; 479 480 // Clear statistics gathered during the concurrent cycle for the given region after 481 // it has been reclaimed. 482 void clear_statistics_in_region(uint region_idx); 483 // Notification for eagerly reclaimed regions to clean up. 484 void humongous_object_eagerly_reclaimed(HeapRegion* r); 485 // Manipulation of the global mark stack. 486 // The push and pop operations are used by tasks for transfers 487 // between task-local queues and the global mark stack. 488 bool mark_stack_push(G1TaskQueueEntry* arr) { 489 if (!_global_mark_stack.par_push_chunk(arr)) { 490 set_has_overflown(); 491 return false; 492 } 493 return true; 494 } 495 bool mark_stack_pop(G1TaskQueueEntry* arr) { 496 return _global_mark_stack.par_pop_chunk(arr); 497 } 498 size_t mark_stack_size() const { return _global_mark_stack.size(); } 499 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; } 500 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); } 501 502 G1CMRootRegions* root_regions() { return &_root_regions; } 503 504 void concurrent_cycle_start(); 505 // Abandon current marking iteration due to a Full GC. 506 void concurrent_cycle_abort(); 507 void concurrent_cycle_end(); 508 509 void update_accum_task_vtime(int i, double vtime) { 510 _accum_task_vtime[i] += vtime; 511 } 512 513 double all_task_accum_vtime() { 514 double ret = 0.0; 515 for (uint i = 0; i < _max_num_tasks; ++i) 516 ret += _accum_task_vtime[i]; 517 return ret; 518 } 519 520 // Attempts to steal an object from the task queues of other tasks 521 bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry); 522 523 G1ConcurrentMark(G1CollectedHeap* g1h, 524 G1RegionToSpaceMapper* prev_bitmap_storage, 525 G1RegionToSpaceMapper* next_bitmap_storage); 526 ~G1ConcurrentMark(); 527 528 G1ConcurrentMarkThread* cm_thread() { return _cm_thread; } 529 530 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; } 531 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; } 532 533 // Calculates the number of concurrent GC threads to be used in the marking phase. 534 uint calc_active_marking_workers(); 535 536 // Moves all per-task cached data into global state. 537 void flush_all_task_caches(); 538 // Prepare internal data structures for the next mark cycle. This includes clearing 539 // the next mark bitmap and some internal data structures. This method is intended 540 // to be called concurrently to the mutator. It will yield to safepoint requests. 541 void cleanup_for_next_mark(); 542 543 // Clear the previous marking bitmap during safepoint. 544 void clear_prev_bitmap(WorkGang* workers); 545 546 // Return whether the next mark bitmap has no marks set. To be used for assertions 547 // only. Will not yield to pause requests. 548 bool next_mark_bitmap_is_clear(); 549 550 // These two methods do the work that needs to be done at the start and end of the 551 // initial mark pause. 552 void pre_initial_mark(); 553 void post_initial_mark(); 554 555 // Scan all the root regions and mark everything reachable from 556 // them. 557 void scan_root_regions(); 558 559 // Scan a single root region and mark everything reachable from it. 560 void scan_root_region(HeapRegion* hr, uint worker_id); 561 562 // Do concurrent phase of marking, to a tentative transitive closure. 563 void mark_from_roots(); 564 565 void remark(); 566 567 void cleanup(); 568 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use 569 // this carefully. 570 inline void mark_in_prev_bitmap(oop p); 571 572 // Clears marks for all objects in the given range, for the prev or 573 // next bitmaps. Caution: the previous bitmap is usually 574 // read-only, so use this carefully! 575 void clear_range_in_prev_bitmap(MemRegion mr); 576 577 inline bool is_marked_in_prev_bitmap(oop p) const; 578 579 // Verify that there are no collection set oops on the stacks (taskqueues / 580 // global mark stack) and fingers (global / per-task). 581 // If marking is not in progress, it's a no-op. 582 void verify_no_cset_oops() PRODUCT_RETURN; 583 584 inline bool do_yield_check(); 585 586 bool has_aborted() { return _has_aborted; } 587 588 void print_summary_info(); 589 590 void print_worker_threads_on(outputStream* st) const; 591 void threads_do(ThreadClosure* tc) const; 592 593 void print_on_error(outputStream* st) const; 594 595 // Mark the given object on the next bitmap if it is below nTAMS. 596 // If the passed obj_size is zero, it is recalculated from the given object if 597 // needed. This is to be as lazy as possible with accessing the object's size. 598 inline bool mark_in_next_bitmap(uint worker_id, HeapRegion* const hr, oop const obj, size_t const obj_size = 0); 599 inline bool mark_in_next_bitmap(uint worker_id, oop const obj, size_t const obj_size = 0); 600 601 inline bool is_marked_in_next_bitmap(oop p) const; 602 603 // Returns true if initialization was successfully completed. 604 bool completed_initialization() const { 605 return _completed_initialization; 606 } 607 608 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; } 609 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; } 610 611 private: 612 // Rebuilds the remembered sets for chosen regions in parallel and concurrently to the application. 613 void rebuild_rem_set_concurrently(); 614 }; 615 616 // A class representing a marking task. 617 class G1CMTask : public TerminatorTerminator { 618 private: 619 enum PrivateConstants { 620 // The regular clock call is called once the scanned words reaches 621 // this limit 622 words_scanned_period = 12*1024, 623 // The regular clock call is called once the number of visited 624 // references reaches this limit 625 refs_reached_period = 1024, 626 // Initial value for the hash seed, used in the work stealing code 627 init_hash_seed = 17 628 }; 629 630 // Number of entries in the per-task stats entry. This seems enough to have a very 631 // low cache miss rate. 632 static const uint RegionMarkStatsCacheSize = 1024; 633 634 G1CMObjArrayProcessor _objArray_processor; 635 636 uint _worker_id; 637 G1CollectedHeap* _g1h; 638 G1ConcurrentMark* _cm; 639 G1CMBitMap* _next_mark_bitmap; 640 // the task queue of this task 641 G1CMTaskQueue* _task_queue; 642 643 G1RegionMarkStatsCache _mark_stats_cache; 644 // Number of calls to this task 645 uint _calls; 646 647 // When the virtual timer reaches this time, the marking step should exit 648 double _time_target_ms; 649 // Start time of the current marking step 650 double _start_time_ms; 651 652 // Oop closure used for iterations over oops 653 G1CMOopClosure* _cm_oop_closure; 654 655 // Region this task is scanning, NULL if we're not scanning any 656 HeapRegion* _curr_region; 657 // Local finger of this task, NULL if we're not scanning a region 658 HeapWord* _finger; 659 // Limit of the region this task is scanning, NULL if we're not scanning one 660 HeapWord* _region_limit; 661 662 // Number of words this task has scanned 663 size_t _words_scanned; 664 // When _words_scanned reaches this limit, the regular clock is 665 // called. Notice that this might be decreased under certain 666 // circumstances (i.e. when we believe that we did an expensive 667 // operation). 668 size_t _words_scanned_limit; 669 // Initial value of _words_scanned_limit (i.e. what it was 670 // before it was decreased). 671 size_t _real_words_scanned_limit; 672 673 // Number of references this task has visited 674 size_t _refs_reached; 675 // When _refs_reached reaches this limit, the regular clock is 676 // called. Notice this this might be decreased under certain 677 // circumstances (i.e. when we believe that we did an expensive 678 // operation). 679 size_t _refs_reached_limit; 680 // Initial value of _refs_reached_limit (i.e. what it was before 681 // it was decreased). 682 size_t _real_refs_reached_limit; 683 684 // Used by the work stealing 685 int _hash_seed; 686 // If true, then the task has aborted for some reason 687 bool _has_aborted; 688 // Set when the task aborts because it has met its time quota 689 bool _has_timed_out; 690 // True when we're draining SATB buffers; this avoids the task 691 // aborting due to SATB buffers being available (as we're already 692 // dealing with them) 693 bool _draining_satb_buffers; 694 695 // Number sequence of past step times 696 NumberSeq _step_times_ms; 697 // Elapsed time of this task 698 double _elapsed_time_ms; 699 // Termination time of this task 700 double _termination_time_ms; 701 // When this task got into the termination protocol 702 double _termination_start_time_ms; 703 704 TruncatedSeq _marking_step_diffs_ms; 705 706 // Updates the local fields after this task has claimed 707 // a new region to scan 708 void setup_for_region(HeapRegion* hr); 709 // Makes the limit of the region up-to-date 710 void update_region_limit(); 711 712 // Called when either the words scanned or the refs visited limit 713 // has been reached 714 void reached_limit(); 715 // Recalculates the words scanned and refs visited limits 716 void recalculate_limits(); 717 // Decreases the words scanned and refs visited limits when we reach 718 // an expensive operation 719 void decrease_limits(); 720 // Checks whether the words scanned or refs visited reached their 721 // respective limit and calls reached_limit() if they have 722 void check_limits() { 723 if (_words_scanned >= _words_scanned_limit || 724 _refs_reached >= _refs_reached_limit) { 725 reached_limit(); 726 } 727 } 728 // Supposed to be called regularly during a marking step as 729 // it checks a bunch of conditions that might cause the marking step 730 // to abort 731 void regular_clock_call(); 732 733 // Test whether obj might have already been passed over by the 734 // mark bitmap scan, and so needs to be pushed onto the mark stack. 735 bool is_below_finger(oop obj, HeapWord* global_finger) const; 736 737 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry); 738 public: 739 // Apply the closure on the given area of the objArray. Return the number of words 740 // scanned. 741 inline size_t scan_objArray(objArrayOop obj, MemRegion mr); 742 // Resets the task; should be called right at the beginning of a marking phase. 743 void reset(G1CMBitMap* next_mark_bitmap); 744 // Clears all the fields that correspond to a claimed region. 745 void clear_region_fields(); 746 747 // The main method of this class which performs a marking step 748 // trying not to exceed the given duration. However, it might exit 749 // prematurely, according to some conditions (i.e. SATB buffers are 750 // available for processing). 751 void do_marking_step(double target_ms, 752 bool do_termination, 753 bool is_serial); 754 755 // These two calls start and stop the timer 756 void record_start_time() { 757 _elapsed_time_ms = os::elapsedTime() * 1000.0; 758 } 759 void record_end_time() { 760 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 761 } 762 763 // Returns the worker ID associated with this task. 764 uint worker_id() { return _worker_id; } 765 766 // From TerminatorTerminator. It determines whether this task should 767 // exit the termination protocol after it's entered it. 768 virtual bool should_exit_termination(); 769 770 // Resets the local region fields after a task has finished scanning a 771 // region; or when they have become stale as a result of the region 772 // being evacuated. 773 void giveup_current_region(); 774 775 HeapWord* finger() { return _finger; } 776 777 bool has_aborted() { return _has_aborted; } 778 void set_has_aborted() { _has_aborted = true; } 779 void clear_has_aborted() { _has_aborted = false; } 780 781 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 782 783 // Increment the number of references this task has visited. 784 void increment_refs_reached() { ++_refs_reached; } 785 786 // Grey the object by marking it. If not already marked, push it on 787 // the local queue if below the finger. obj is required to be below its region's NTAMS. 788 // Returns whether there has been a mark to the bitmap. 789 inline bool make_reference_grey(oop obj); 790 791 // Grey the object (by calling make_grey_reference) if required, 792 // e.g. obj is below its containing region's NTAMS. 793 // Precondition: obj is a valid heap object. 794 // Returns true if the reference caused a mark to be set in the next bitmap. 795 template <class T> 796 inline bool deal_with_reference(T* p); 797 798 // Scans an object and visits its children. 799 inline void scan_task_entry(G1TaskQueueEntry task_entry); 800 801 // Pushes an object on the local queue. 802 inline void push(G1TaskQueueEntry task_entry); 803 804 // Move entries to the global stack. 805 void move_entries_to_global_stack(); 806 // Move entries from the global stack, return true if we were successful to do so. 807 bool get_entries_from_global_stack(); 808 809 // Pops and scans objects from the local queue. If partially is 810 // true, then it stops when the queue size is of a given limit. If 811 // partially is false, then it stops when the queue is empty. 812 void drain_local_queue(bool partially); 813 // Moves entries from the global stack to the local queue and 814 // drains the local queue. If partially is true, then it stops when 815 // both the global stack and the local queue reach a given size. If 816 // partially if false, it tries to empty them totally. 817 void drain_global_stack(bool partially); 818 // Keeps picking SATB buffers and processing them until no SATB 819 // buffers are available. 820 void drain_satb_buffers(); 821 822 // Moves the local finger to a new location 823 inline void move_finger_to(HeapWord* new_finger) { 824 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 825 _finger = new_finger; 826 } 827 828 G1CMTask(uint worker_id, 829 G1ConcurrentMark *cm, 830 G1CMTaskQueue* task_queue, 831 G1RegionMarkStats* mark_stats, 832 uint max_regions); 833 834 inline void update_liveness(oop const obj, size_t const obj_size); 835 836 // Clear (without flushing) the mark cache entry for the given region. 837 void clear_mark_stats_cache(uint region_idx); 838 // Evict the whole statistics cache into the global statistics. Returns the 839 // number of cache hits and misses so far. 840 Pair<size_t, size_t> flush_mark_stats_cache(); 841 // Prints statistics associated with this task 842 void print_stats(); 843 }; 844 845 // Class that's used to to print out per-region liveness 846 // information. It's currently used at the end of marking and also 847 // after we sort the old regions at the end of the cleanup operation. 848 class G1PrintRegionLivenessInfoClosure : public HeapRegionClosure { 849 // Accumulators for these values. 850 size_t _total_used_bytes; 851 size_t _total_capacity_bytes; 852 size_t _total_prev_live_bytes; 853 size_t _total_next_live_bytes; 854 855 // Accumulator for the remembered set size 856 size_t _total_remset_bytes; 857 858 // Accumulator for strong code roots memory size 859 size_t _total_strong_code_roots_bytes; 860 861 static double bytes_to_mb(size_t val) { 862 return (double) val / (double) M; 863 } 864 865 public: 866 // The header and footer are printed in the constructor and 867 // destructor respectively. 868 G1PrintRegionLivenessInfoClosure(const char* phase_name); 869 virtual bool do_heap_region(HeapRegion* r); 870 ~G1PrintRegionLivenessInfoClosure(); 871 }; 872 873 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP