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