1 /*
2 * Copyright (c) 2001, 2017, 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/heapRegionSet.hpp"
31 #include "gc/shared/taskqueue.hpp"
32
33 class ConcurrentGCTimer;
34 class ConcurrentMarkThread;
35 class G1CollectedHeap;
36 class G1CMTask;
37 class G1ConcurrentMark;
38 class G1OldTracer;
39 class G1RegionToSpaceMapper;
40 class G1SurvivorRegions;
41
42 #ifdef _MSC_VER
43 #pragma warning(push)
44 // warning C4522: multiple assignment operators specified
45 #pragma warning(disable:4522)
46 #endif
47
48 // This is a container class for either an oop or a continuation address for
49 // mark stack entries. Both are pushed onto the mark stack.
50 class G1TaskQueueEntry VALUE_OBJ_CLASS_SPEC {
51 private:
52 void* _holder;
53
54 static const uintptr_t ArraySliceBit = 1;
55
56 G1TaskQueueEntry(oop obj) : _holder(obj) {
57 assert(_holder != NULL, "Not allowed to set NULL task queue element");
58 }
59 G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
60 public:
61 G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
62 G1TaskQueueEntry() : _holder(NULL) { }
63
64 static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
65 static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
66
67 G1TaskQueueEntry& operator=(const G1TaskQueueEntry& t) {
68 _holder = t._holder;
69 return *this;
70 }
71
72 volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
73 _holder = t._holder;
74 return *this;
75 }
76
77 oop obj() const {
78 assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
79 return (oop)_holder;
80 }
81
82 HeapWord* slice() const {
83 assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
84 return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
85 }
86
87 bool is_oop() const { return !is_array_slice(); }
88 bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
89 bool is_null() const { return _holder == NULL; }
90 };
91
92 #ifdef _MSC_VER
93 #pragma warning(pop)
94 #endif
95
96 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
97 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
98
99 // Closure used by CM during concurrent reference discovery
100 // and reference processing (during remarking) to determine
101 // if a particular object is alive. It is primarily used
102 // to determine if referents of discovered reference objects
103 // are alive. An instance is also embedded into the
104 // reference processor as the _is_alive_non_header field
105 class G1CMIsAliveClosure: public BoolObjectClosure {
106 G1CollectedHeap* _g1;
107 public:
108 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
109
110 bool do_object_b(oop obj);
111 };
112
113 // Represents the overflow mark stack used by concurrent marking.
114 //
115 // Stores oops in a huge buffer in virtual memory that is always fully committed.
116 // Resizing may only happen during a STW pause when the stack is empty.
117 //
118 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
119 // stack memory is split into evenly sized chunks of oops. Users can only
120 // add or remove entries on that basis.
121 // Chunks are filled in increasing address order. Not completely filled chunks
122 // have a NULL element as a terminating element.
123 //
124 // Every chunk has a header containing a single pointer element used for memory
125 // management. This wastes some space, but is negligible (< .1% with current sizing).
126 //
127 // Memory management is done using a mix of tracking a high water-mark indicating
128 // that all chunks at a lower address are valid chunks, and a singly linked free
129 // list connecting all empty chunks.
130 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
131 public:
132 // Number of TaskQueueEntries that can fit in a single chunk.
133 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
134 private:
135 struct TaskQueueEntryChunk {
136 TaskQueueEntryChunk* next;
137 G1TaskQueueEntry data[EntriesPerChunk];
138 };
139
140 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack.
141
142 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area.
143 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements.
144
145 char _pad0[DEFAULT_CACHE_LINE_SIZE];
146 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
147 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
148 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
149 volatile size_t _chunks_in_chunk_list;
150 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
151
152 volatile size_t _hwm; // High water mark within the reserved space.
153 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
154
155 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
156 // NULL if out of memory.
157 TaskQueueEntryChunk* allocate_new_chunk();
158
159 // Atomically add the given chunk to the list.
160 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
161 // Atomically remove and return a chunk from the given list. Returns NULL if the
162 // list is empty.
163 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
164
165 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
166 void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
167
168 TaskQueueEntryChunk* remove_chunk_from_chunk_list();
169 TaskQueueEntryChunk* remove_chunk_from_free_list();
170
171 // Resizes the mark stack to the given new capacity. Releases any previous
172 // memory if successful.
173 bool resize(size_t new_capacity);
174
175 public:
176 G1CMMarkStack();
177 ~G1CMMarkStack();
178
179 // Alignment and minimum capacity of this mark stack in number of oops.
180 static size_t capacity_alignment();
181
182 // Allocate and initialize the mark stack with the given number of oops.
183 bool initialize(size_t initial_capacity, size_t max_capacity);
184
185 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
186 // stack. If less than EntriesPerChunk elements are to be pushed, the array must
187 // be terminated with a NULL.
188 // Returns whether the buffer contents were successfully pushed to the global mark
189 // stack.
190 bool par_push_chunk(G1TaskQueueEntry* buffer);
191
192 // Pops a chunk from this mark stack, copying them into the given buffer. This
193 // chunk may contain up to EntriesPerChunk elements. If there are less, the last
194 // element in the array is a NULL pointer.
195 bool par_pop_chunk(G1TaskQueueEntry* buffer);
196
197 // Return whether the chunk list is empty. Racy due to unsynchronized access to
198 // _chunk_list.
199 bool is_empty() const { return _chunk_list == NULL; }
200
201 size_t capacity() const { return _chunk_capacity; }
202
203 // Expand the stack, typically in response to an overflow condition
204 void expand();
205
206 // Return the approximate number of oops on this mark stack. Racy due to
207 // unsynchronized access to _chunks_in_chunk_list.
208 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
209
210 void set_empty();
211
212 // Apply Fn to every oop on the mark stack. The mark stack must not
213 // be modified while iterating.
214 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
215 };
216
217 // Root Regions are regions that are not empty at the beginning of a
218 // marking cycle and which we might collect during an evacuation pause
219 // while the cycle is active. Given that, during evacuation pauses, we
220 // do not copy objects that are explicitly marked, what we have to do
221 // for the root regions is to scan them and mark all objects reachable
222 // from them. According to the SATB assumptions, we only need to visit
223 // each object once during marking. So, as long as we finish this scan
224 // before the next evacuation pause, we can copy the objects from the
225 // root regions without having to mark them or do anything else to them.
226 //
227 // Currently, we only support root region scanning once (at the start
228 // of the marking cycle) and the root regions are all the survivor
229 // regions populated during the initial-mark pause.
230 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
231 private:
232 const G1SurvivorRegions* _survivors;
233 G1ConcurrentMark* _cm;
234
235 volatile bool _scan_in_progress;
236 volatile bool _should_abort;
237 volatile int _claimed_survivor_index;
238
239 void notify_scan_done();
240
241 public:
242 G1CMRootRegions();
243 // We actually do most of the initialization in this method.
244 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
245
246 // Reset the claiming / scanning of the root regions.
247 void prepare_for_scan();
248
249 // Forces get_next() to return NULL so that the iteration aborts early.
250 void abort() { _should_abort = true; }
251
252 // Return true if the CM thread are actively scanning root regions,
253 // false otherwise.
254 bool scan_in_progress() { return _scan_in_progress; }
255
256 // Claim the next root region to scan atomically, or return NULL if
257 // all have been claimed.
258 HeapRegion* claim_next();
259
260 // The number of root regions to scan.
261 uint num_root_regions() const;
262
263 void cancel_scan();
264
265 // Flag that we're done with root region scanning and notify anyone
266 // who's waiting on it. If aborted is false, assume that all regions
267 // have been claimed.
268 void scan_finished();
269
270 // If CM threads are still scanning root regions, wait until they
271 // are done. Return true if we had to wait, false otherwise.
272 bool wait_until_scan_finished();
273 };
274
275 // This class manages data structures and methods for doing liveness analysis in
276 // G1's concurrent cycle.
277 class G1ConcurrentMark: public CHeapObj<mtGC> {
278 friend class ConcurrentMarkThread;
279 friend class G1CMRefProcTaskProxy;
280 friend class G1CMRefProcTaskExecutor;
281 friend class G1CMKeepAliveAndDrainClosure;
282 friend class G1CMDrainMarkingStackClosure;
283 friend class G1CMBitMapClosure;
284 friend class G1CMConcurrentMarkingTask;
285 friend class G1CMRemarkTask;
286 friend class G1CMTask;
287
288 ConcurrentMarkThread* _cm_thread; // The thread doing the work
289 G1CollectedHeap* _g1h; // The heap
290 bool _completed_initialization; // Set to true when initialization is complete
291
292 FreeRegionList _cleanup_list;
293
294 // Concurrent marking support structures
295 G1CMBitMap _mark_bitmap_1;
296 G1CMBitMap _mark_bitmap_2;
297 G1CMBitMap* _prev_mark_bitmap; // Completed mark bitmap
298 G1CMBitMap* _next_mark_bitmap; // Under-construction mark bitmap
299
300 // Heap bounds
301 HeapWord* _heap_start;
302 HeapWord* _heap_end;
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 _max_num_tasks; // Maximum number of marking tasks
314 uint _num_active_tasks; // Number of tasks currently active
315 G1CMTask** _tasks; // Task queue array (max_worker_id length)
316
317 G1CMTaskQueueSet* _task_queues; // Task queue set
318 ParallelTaskTerminator _terminator; // For termination
319
320 // Two sync barriers that are used to synchronize tasks when an
321 // overflow occurs. The algorithm is the following. All tasks enter
322 // the first one to ensure that they have all stopped manipulating
323 // the global data structures. After they exit it, they re-initialize
324 // their data structures and task 0 re-initializes the global data
325 // structures. Then, they enter the second sync barrier. This
326 // ensure, that no task starts doing work before all data
327 // structures (local and global) have been re-initialized. When they
328 // exit it, they are free to start working again.
329 WorkGangBarrierSync _first_overflow_barrier_sync;
330 WorkGangBarrierSync _second_overflow_barrier_sync;
331
332 // This is set by any task, when an overflow on the global data
333 // structures is detected
334 volatile bool _has_overflown;
335 // True: marking is concurrent, false: we're in remark
336 volatile bool _concurrent;
337 // Set at the end of a Full GC so that marking aborts
338 volatile bool _has_aborted;
339
340 // Used when remark aborts due to an overflow to indicate that
341 // another concurrent marking phase should start
342 volatile bool _restart_for_overflow;
343
344 // This is true from the very start of concurrent marking until the
345 // point when all the tasks complete their work. It is really used
346 // to determine the points between the end of concurrent marking and
347 // time of remark.
348 volatile bool _concurrent_marking_in_progress;
349
350 ConcurrentGCTimer* _gc_timer_cm;
351
352 G1OldTracer* _gc_tracer_cm;
353
354 // Timing statistics. All of them are in ms
355 NumberSeq _init_times;
356 NumberSeq _remark_times;
357 NumberSeq _remark_mark_times;
358 NumberSeq _remark_weak_ref_times;
359 NumberSeq _cleanup_times;
360 double _total_counting_time;
361 double _total_rs_scrub_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 weak_refs_work_parallel_part(BoolObjectClosure* is_alive, bool purged_classes);
370 void weak_refs_work(bool clear_all_soft_refs);
371
372 void swap_mark_bitmaps();
373
374 // Resets the global marking data structures, as well as the
375 // task local ones; should be called during initial mark.
376 void reset();
377
378 // Resets all the marking data structures. Called when we have to restart
379 // marking or when marking completes (via set_non_marking_state below).
380 void reset_marking_state();
381
382 // We do this after we're done with marking so that the marking data
383 // structures are initialized to a sensible and predictable state.
384 void set_non_marking_state();
385
386 // Called to indicate how many threads are currently active.
387 void set_concurrency(uint active_tasks);
388
389 // Should be called to indicate which phase we're in (concurrent
390 // mark or remark) and how many threads are currently active.
391 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
392
393 // Prints all gathered CM-related statistics
394 void print_stats();
395
396 bool cleanup_list_is_empty() {
397 return _cleanup_list.is_empty();
398 }
399
400 HeapWord* finger() { return _finger; }
401 bool concurrent() { return _concurrent; }
402 uint active_tasks() { return _num_active_tasks; }
403 ParallelTaskTerminator* terminator() { return &_terminator; }
404
405 // Claims the next available region to be scanned by a marking
406 // task/thread. It might return NULL if the next region is empty or
407 // we have run out of regions. In the latter case, out_of_regions()
408 // determines whether we've really run out of regions or the task
409 // should call claim_region() again. This might seem a bit
410 // awkward. Originally, the code was written so that claim_region()
411 // either successfully returned with a non-empty region or there
412 // were no more regions to be claimed. The problem with this was
413 // that, in certain circumstances, it iterated over large chunks of
414 // the heap finding only empty regions and, while it was working, it
415 // was preventing the calling task to call its regular clock
416 // method. So, this way, each task will spend very little time in
417 // claim_region() and is allowed to call the regular clock method
418 // frequently.
419 HeapRegion* claim_region(uint worker_id);
420
421 // Determines whether we've run out of regions to scan. Note that
422 // the finger can point past the heap end in case the heap was expanded
423 // to satisfy an allocation without doing a GC. This is fine, because all
424 // objects in those regions will be considered live anyway because of
425 // SATB guarantees (i.e. their TAMS will be equal to bottom).
426 bool out_of_regions() { return _finger >= _heap_end; }
427
428 // Returns the task with the given id
429 G1CMTask* task(uint id) {
430 assert(id < _num_active_tasks, "Task id %u not within active bounds up to %u", id, _num_active_tasks);
431 return _tasks[id];
432 }
433
434 // Access / manipulation of the overflow flag which is set to
435 // indicate that the global stack has overflown
436 bool has_overflown() { return _has_overflown; }
437 void set_has_overflown() { _has_overflown = true; }
438 void clear_has_overflown() { _has_overflown = false; }
439 bool restart_for_overflow() { return _restart_for_overflow; }
440
441 // Methods to enter the two overflow sync barriers
442 void enter_first_sync_barrier(uint worker_id);
443 void enter_second_sync_barrier(uint worker_id);
444
445 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
446 // true, periodically insert checks to see if this method should exit prematurely.
447 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
448 public:
449 // Manipulation of the global mark stack.
450 // The push and pop operations are used by tasks for transfers
451 // between task-local queues and the global mark stack.
452 bool mark_stack_push(G1TaskQueueEntry* arr) {
453 if (!_global_mark_stack.par_push_chunk(arr)) {
454 set_has_overflown();
455 return false;
456 }
457 return true;
458 }
459 bool mark_stack_pop(G1TaskQueueEntry* arr) {
460 return _global_mark_stack.par_pop_chunk(arr);
461 }
462 size_t mark_stack_size() const { return _global_mark_stack.size(); }
463 size_t partial_mark_stack_size_target() const { return _global_mark_stack.capacity() / 3; }
464 bool mark_stack_empty() const { return _global_mark_stack.is_empty(); }
465
466 G1CMRootRegions* root_regions() { return &_root_regions; }
467
468 bool concurrent_marking_in_progress() const {
469 return _concurrent_marking_in_progress;
470 }
471 void set_concurrent_marking_in_progress() {
472 _concurrent_marking_in_progress = true;
473 }
474 void clear_concurrent_marking_in_progress() {
475 _concurrent_marking_in_progress = false;
476 }
477
478 void concurrent_cycle_start();
479 void concurrent_cycle_end();
480
481 void update_accum_task_vtime(int i, double vtime) {
482 _accum_task_vtime[i] += vtime;
483 }
484
485 double all_task_accum_vtime() {
486 double ret = 0.0;
487 for (uint i = 0; i < _max_num_tasks; ++i)
488 ret += _accum_task_vtime[i];
489 return ret;
490 }
491
492 // Attempts to steal an object from the task queues of other tasks
493 bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry);
494
495 G1ConcurrentMark(G1CollectedHeap* g1h,
496 G1RegionToSpaceMapper* prev_bitmap_storage,
497 G1RegionToSpaceMapper* next_bitmap_storage);
498 ~G1ConcurrentMark();
499
500 ConcurrentMarkThread* cm_thread() { return _cm_thread; }
501
502 const G1CMBitMap* const prev_mark_bitmap() const { return _prev_mark_bitmap; }
503 G1CMBitMap* next_mark_bitmap() const { return _next_mark_bitmap; }
504
505 // Calculates the number of concurrent GC threads to be used in the marking phase.
506 uint calc_active_marking_workers();
507
508 // Prepare internal data structures for the next mark cycle. This includes clearing
509 // the next mark bitmap and some internal data structures. This method is intended
510 // to be called concurrently to the mutator. It will yield to safepoint requests.
511 void cleanup_for_next_mark();
512
513 // Clear the previous marking bitmap during safepoint.
514 void clear_prev_bitmap(WorkGang* workers);
515
516 // Return whether the next mark bitmap has no marks set. To be used for assertions
517 // only. Will not yield to pause requests.
518 bool next_mark_bitmap_is_clear();
519
520 // These two do the work that needs to be done before and after the
521 // initial root checkpoint. Since this checkpoint can be done at two
522 // different points (i.e. an explicit pause or piggy-backed on a
523 // young collection), then it's nice to be able to easily share the
524 // pre/post code. It might be the case that we can put everything in
525 // the post method.
526 void checkpoint_roots_initial_pre();
527 void checkpoint_roots_initial_post();
528
529 // Scan all the root regions and mark everything reachable from
530 // them.
531 void scan_root_regions();
532
533 // Scan a single root region and mark everything reachable from it.
534 void scan_root_region(HeapRegion* hr);
535
536 // Do concurrent phase of marking, to a tentative transitive closure.
537 void mark_from_roots();
538
539 void checkpoint_roots_final(bool clear_all_soft_refs);
540 void checkpoint_roots_final_work();
541
542 void cleanup();
543 void complete_cleanup();
544
545 // Mark in the previous bitmap. Caution: the prev bitmap is usually read-only, so use
546 // this carefully.
547 inline void mark_in_prev_bitmap(oop p);
548
549 // Clears marks for all objects in the given range, for the prev or
550 // next bitmaps. Caution: the previous bitmap is usually
551 // read-only, so use this carefully!
552 void clear_range_in_prev_bitmap(MemRegion mr);
553
554 inline bool is_marked_in_prev_bitmap(oop p) const;
555
556 // Verify that there are no CSet oops on the stacks (taskqueues /
557 // global mark stack) and fingers (global / per-task).
558 // If marking is not in progress, it's a no-op.
559 void verify_no_cset_oops() PRODUCT_RETURN;
560
561 inline bool do_yield_check();
562
563 // Abandon current marking iteration due to a Full GC.
564 void abort();
565
566 bool has_aborted() { return _has_aborted; }
567
568 void print_summary_info();
569
570 void print_worker_threads_on(outputStream* st) const;
571 void threads_do(ThreadClosure* tc) const;
572
573 void print_on_error(outputStream* st) const;
574
575 // Mark the given object on the next bitmap if it is below nTAMS.
576 inline bool mark_in_next_bitmap(HeapRegion* const hr, oop const obj);
577 inline bool mark_in_next_bitmap(oop const obj);
578
579 // Returns true if initialization was successfully completed.
580 bool completed_initialization() const {
581 return _completed_initialization;
582 }
583
584 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
585 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
586
587 private:
588 // Clear (Reset) all liveness count data.
589 void clear_live_data(WorkGang* workers);
590
591 #ifdef ASSERT
592 // Verify all of the above data structures that they are in initial state.
593 void verify_live_data_clear();
594 #endif
595
596 // Aggregates the per-card liveness data based on the current marking. Also sets
597 // the amount of marked bytes for each region.
598 void create_live_data();
599
600 void finalize_live_data();
601
602 void verify_live_data();
603 };
604
605 // A class representing a marking task.
606 class G1CMTask : public TerminatorTerminator {
607 private:
608 enum PrivateConstants {
609 // The regular clock call is called once the scanned words reaches
610 // this limit
611 words_scanned_period = 12*1024,
612 // The regular clock call is called once the number of visited
613 // references reaches this limit
614 refs_reached_period = 1024,
615 // Initial value for the hash seed, used in the work stealing code
616 init_hash_seed = 17
617 };
618
619 G1CMObjArrayProcessor _objArray_processor;
620
621 uint _worker_id;
622 G1CollectedHeap* _g1h;
623 G1ConcurrentMark* _cm;
624 G1CMBitMap* _next_mark_bitmap;
625 // the task queue of this task
626 G1CMTaskQueue* _task_queue;
627
628 // Number of calls to this task
629 uint _calls;
630
631 // When the virtual timer reaches this time, the marking step should exit
632 double _time_target_ms;
633 // Start time of the current marking step
634 double _start_time_ms;
635
636 // Oop closure used for iterations over oops
637 G1CMOopClosure* _cm_oop_closure;
638
639 // Region this task is scanning, NULL if we're not scanning any
640 HeapRegion* _curr_region;
641 // Local finger of this task, NULL if we're not scanning a region
642 HeapWord* _finger;
643 // Limit of the region this task is scanning, NULL if we're not scanning one
644 HeapWord* _region_limit;
645
646 // Number of words this task has scanned
647 size_t _words_scanned;
648 // When _words_scanned reaches this limit, the regular clock is
649 // called. Notice that this might be decreased under certain
650 // circumstances (i.e. when we believe that we did an expensive
651 // operation).
652 size_t _words_scanned_limit;
653 // Initial value of _words_scanned_limit (i.e. what it was
654 // before it was decreased).
655 size_t _real_words_scanned_limit;
656
657 // Number of references this task has visited
658 size_t _refs_reached;
659 // When _refs_reached reaches this limit, the regular clock is
660 // called. Notice this this might be decreased under certain
661 // circumstances (i.e. when we believe that we did an expensive
662 // operation).
663 size_t _refs_reached_limit;
664 // Initial value of _refs_reached_limit (i.e. what it was before
665 // it was decreased).
666 size_t _real_refs_reached_limit;
667
668 // Used by the work stealing
669 int _hash_seed;
670 // If true, then the task has aborted for some reason
671 bool _has_aborted;
672 // Set when the task aborts because it has met its time quota
673 bool _has_timed_out;
674 // True when we're draining SATB buffers; this avoids the task
675 // aborting due to SATB buffers being available (as we're already
676 // dealing with them)
677 bool _draining_satb_buffers;
678
679 // Number sequence of past step times
680 NumberSeq _step_times_ms;
681 // Elapsed time of this task
682 double _elapsed_time_ms;
683 // Termination time of this task
684 double _termination_time_ms;
685 // When this task got into the termination protocol
686 double _termination_start_time_ms;
687
688 // True when the task is during a concurrent phase, false when it is
689 // in the remark phase (so, in the latter case, we do not have to
690 // check all the things that we have to check during the concurrent
691 // phase, i.e. SATB buffer availability...)
692 bool _concurrent;
693
694 TruncatedSeq _marking_step_diffs_ms;
695
696 // Updates the local fields after this task has claimed
697 // a new region to scan
698 void setup_for_region(HeapRegion* hr);
699 // Makes the limit of the region up-to-date
700 void update_region_limit();
701
702 // Called when either the words scanned or the refs visited limit
703 // has been reached
704 void reached_limit();
705 // Recalculates the words scanned and refs visited limits
706 void recalculate_limits();
707 // Decreases the words scanned and refs visited limits when we reach
708 // an expensive operation
709 void decrease_limits();
710 // Checks whether the words scanned or refs visited reached their
711 // respective limit and calls reached_limit() if they have
712 void check_limits() {
713 if (_words_scanned >= _words_scanned_limit ||
714 _refs_reached >= _refs_reached_limit) {
715 reached_limit();
716 }
717 }
718 // Supposed to be called regularly during a marking step as
719 // it checks a bunch of conditions that might cause the marking step
720 // to abort
721 void regular_clock_call();
722
723 // Test whether obj might have already been passed over by the
724 // mark bitmap scan, and so needs to be pushed onto the mark stack.
725 bool is_below_finger(oop obj, HeapWord* global_finger) const;
726
727 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
728 public:
729 // Apply the closure on the given area of the objArray. Return the number of words
730 // scanned.
731 inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
732 // Resets the task; should be called right at the beginning of a marking phase.
733 void reset(G1CMBitMap* next_mark_bitmap);
734 // Clears all the fields that correspond to a claimed region.
735 void clear_region_fields();
736
737 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
738
739 // The main method of this class which performs a marking step
740 // trying not to exceed the given duration. However, it might exit
741 // prematurely, according to some conditions (i.e. SATB buffers are
742 // available for processing).
743 void do_marking_step(double target_ms,
744 bool do_termination,
745 bool is_serial);
746
747 // These two calls start and stop the timer
748 void record_start_time() {
749 _elapsed_time_ms = os::elapsedTime() * 1000.0;
750 }
751 void record_end_time() {
752 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
753 }
754
755 // Returns the worker ID associated with this task.
756 uint worker_id() { return _worker_id; }
757
758 // From TerminatorTerminator. It determines whether this task should
759 // exit the termination protocol after it's entered it.
760 virtual bool should_exit_termination();
761
762 // Resets the local region fields after a task has finished scanning a
763 // region; or when they have become stale as a result of the region
764 // being evacuated.
765 void giveup_current_region();
766
767 HeapWord* finger() { return _finger; }
768
769 bool has_aborted() { return _has_aborted; }
770 void set_has_aborted() { _has_aborted = true; }
771 void clear_has_aborted() { _has_aborted = false; }
772
773 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
774
775 // Increment the number of references this task has visited.
776 void increment_refs_reached() { ++_refs_reached; }
777
778 // Grey the object by marking it. If not already marked, push it on
779 // the local queue if below the finger.
780 // obj is below its region's NTAMS.
781 inline void make_reference_grey(oop obj);
782
783 // Grey the object (by calling make_grey_reference) if required,
784 // e.g. obj is below its containing region's NTAMS.
785 // Precondition: obj is a valid heap object.
786 inline void deal_with_reference(oop obj);
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
822 // Prints statistics associated with this task
823 void print_stats();
824 };
825
826 // Class that's used to to print out per-region liveness
827 // information. It's currently used at the end of marking and also
828 // after we sort the old regions at the end of the cleanup operation.
829 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
830 private:
831 // Accumulators for these values.
832 size_t _total_used_bytes;
833 size_t _total_capacity_bytes;
834 size_t _total_prev_live_bytes;
835 size_t _total_next_live_bytes;
836
837 // Accumulator for the remembered set size
838 size_t _total_remset_bytes;
839
840 // Accumulator for strong code roots memory size
841 size_t _total_strong_code_roots_bytes;
842
843 static double bytes_to_mb(size_t val) {
844 return (double) val / (double) M;
845 }
846
847 public:
848 // The header and footer are printed in the constructor and
849 // destructor respectively.
850 G1PrintRegionLivenessInfoClosure(const char* phase_name);
851 virtual bool doHeapRegion(HeapRegion* r);
852 ~G1PrintRegionLivenessInfoClosure();
853 };
854
855 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP