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