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 "classfile/javaClasses.hpp"
29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp"
30 #include "gc/g1/g1RegionToSpaceMapper.hpp"
31 #include "gc/g1/heapRegionSet.hpp"
32 #include "gc/shared/taskqueue.hpp"
33
34 class G1CollectedHeap;
35 class G1CMBitMap;
36 class G1CMTask;
37 class G1ConcurrentMark;
38 class ConcurrentGCTimer;
39 class G1OldTracer;
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 // Closure for iteration over bitmaps
114 class G1CMBitMapClosure VALUE_OBJ_CLASS_SPEC {
115 private:
116 G1ConcurrentMark* _cm;
117 G1CMTask* _task;
118 public:
119 G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm) : _task(task), _cm(cm) { }
120
121 bool do_addr(HeapWord* const addr);
122 };
123
124 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
125 private:
126 G1CMBitMap* _bm;
127 public:
128 G1CMBitMapMappingChangedListener() : _bm(NULL) {}
129
130 void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
131
132 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
133 };
134
135 // A generic mark bitmap for concurrent marking. This is essentially a wrapper
136 // around the BitMap class that is based on HeapWords, with one bit per (1<<_shifter) HeapWords.
137 class G1CMBitMap VALUE_OBJ_CLASS_SPEC {
138 private:
139 HeapWord* _start; // Base (heap) address of range covered by map.
140 size_t _word_size; // Covered range size (in #HeapWords covered)
141
142 const int _shifter; // map to char or bit
143
144 BitMapView _bm; // the bit map itself
145
146 G1CMBitMapMappingChangedListener _listener;
147
148 HeapWord* end_word() const { return _start + _word_size; }
149
150 // Convert from bit offset to address.
151 HeapWord* offset_to_addr(size_t offset) const {
152 return _start + (offset << _shifter);
153 }
154 // Convert from address to bit offset.
155 size_t addr_to_offset(const HeapWord* addr) const {
156 return pointer_delta(addr, _start) >> _shifter;
157 }
158
159 // debugging
160 NOT_PRODUCT(bool covers(MemRegion rs) const;)
161 public:
162 static size_t compute_size(size_t heap_size);
163 // Returns the amount of bytes on the heap between two marks in the bitmap.
164 static size_t mark_distance();
165 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
166 // mark bitmap corresponds to. This is the same as the mark distance above.
167 static size_t heap_map_factor() {
168 return mark_distance();
169 }
170
171 G1CMBitMap() : _start(NULL), _word_size(0), _bm(), _shifter(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
172
173 // Initializes the underlying BitMap to cover the given area.
174 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
175
176 // read marks
177 bool is_marked(HeapWord* addr) const {
178 assert(_start <= addr && addr < end_word(), "Address " PTR_FORMAT " is outside underlying space from " PTR_FORMAT " to " PTR_FORMAT, p2i(addr), p2i(_start), p2i(end_word()));
179 return _bm.at(addr_to_offset(addr));
180 }
181
182 // Apply the closure to the addresses that correspond to marked bits in the bitmap.
183 inline bool iterate(G1CMBitMapClosure* cl, MemRegion mr);
184
185 // Return the address corresponding to the next marked bit at or after
186 // "addr", and before "limit", if "limit" is non-NULL. If there is no
187 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
188 inline HeapWord* get_next_marked_addr(const HeapWord* addr,
189 const HeapWord* limit) const;
190
191 // The argument addr should be the start address of a valid object
192 inline HeapWord* addr_after_obj(HeapWord* addr);
193
194 void print_on_error(outputStream* st, const char* prefix) const;
195
196 // Write marks.
197 inline void mark(HeapWord* addr);
198 inline void clear(HeapWord* addr);
199 inline bool par_mark(HeapWord* addr);
200
201 void clear_range(MemRegion mr);
202 };
203
204 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
205 private:
206 G1CMBitMap* _bitmap;
207
208 public:
209 G1CMBitMapRO(G1CMBitMap* bitmap) : _bitmap(bitmap) { }
210
211 inline bool is_marked(HeapWord* addr) const;
212 inline HeapWord* get_next_marked_addr(const HeapWord* addr, const HeapWord* limit) const;
213 };
214
215 // Represents the overflow mark stack used by concurrent marking.
216 //
217 // Stores oops in a huge buffer in virtual memory that is always fully committed.
218 // Resizing may only happen during a STW pause when the stack is empty.
219 //
220 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
221 // stack memory is split into evenly sized chunks of oops. Users can only
222 // add or remove entries on that basis.
223 // Chunks are filled in increasing address order. Not completely filled chunks
224 // have a NULL element as a terminating element.
225 //
226 // Every chunk has a header containing a single pointer element used for memory
227 // management. This wastes some space, but is negligible (< .1% with current sizing).
228 //
229 // Memory management is done using a mix of tracking a high water-mark indicating
230 // that all chunks at a lower address are valid chunks, and a singly linked free
231 // list connecting all empty chunks.
232 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
233 public:
234 // Number of TaskQueueEntries that can fit in a single chunk.
235 static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
236 private:
237 struct TaskQueueEntryChunk {
238 TaskQueueEntryChunk* next;
239 G1TaskQueueEntry data[EntriesPerChunk];
240 };
241
242 size_t _max_chunk_capacity; // Maximum number of TaskQueueEntryChunk elements on the stack.
243
244 TaskQueueEntryChunk* _base; // Bottom address of allocated memory area.
245 size_t _chunk_capacity; // Current maximum number of TaskQueueEntryChunk elements.
246
247 char _pad0[DEFAULT_CACHE_LINE_SIZE];
248 TaskQueueEntryChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users.
249 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
250 TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
251 volatile size_t _chunks_in_chunk_list;
252 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
253
254 volatile size_t _hwm; // High water mark within the reserved space.
255 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
256
257 // Allocate a new chunk from the reserved memory, using the high water mark. Returns
258 // NULL if out of memory.
259 TaskQueueEntryChunk* allocate_new_chunk();
260
261 // Atomically add the given chunk to the list.
262 void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
263 // Atomically remove and return a chunk from the given list. Returns NULL if the
264 // list is empty.
265 TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
266
267 void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
268 void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
269
270 TaskQueueEntryChunk* remove_chunk_from_chunk_list();
271 TaskQueueEntryChunk* remove_chunk_from_free_list();
272
273 // Resizes the mark stack to the given new capacity. Releases any previous
274 // memory if successful.
275 bool resize(size_t new_capacity);
276
277 public:
278 G1CMMarkStack();
279 ~G1CMMarkStack();
280
281 // Alignment and minimum capacity of this mark stack in number of oops.
282 static size_t capacity_alignment();
283
284 // Allocate and initialize the mark stack with the given number of oops.
285 bool initialize(size_t initial_capacity, size_t max_capacity);
286
287 // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
288 // stack. If less than EntriesPerChunk elements are to be pushed, the array must
289 // be terminated with a NULL.
290 // Returns whether the buffer contents were successfully pushed to the global mark
291 // stack.
292 bool par_push_chunk(G1TaskQueueEntry* buffer);
293
294 // Pops a chunk from this mark stack, copying them into the given buffer. This
295 // chunk may contain up to EntriesPerChunk elements. If there are less, the last
296 // element in the array is a NULL pointer.
297 bool par_pop_chunk(G1TaskQueueEntry* buffer);
298
299 // Return whether the chunk list is empty. Racy due to unsynchronized access to
300 // _chunk_list.
301 bool is_empty() const { return _chunk_list == NULL; }
302
303 size_t capacity() const { return _chunk_capacity; }
304
305 // Expand the stack, typically in response to an overflow condition
306 void expand();
307
308 // Return the approximate number of oops on this mark stack. Racy due to
309 // unsynchronized access to _chunks_in_chunk_list.
310 size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
311
312 void set_empty();
313
314 // Apply Fn to every oop on the mark stack. The mark stack must not
315 // be modified while iterating.
316 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
317 };
318
319 // Root Regions are regions that are not empty at the beginning of a
320 // marking cycle and which we might collect during an evacuation pause
321 // while the cycle is active. Given that, during evacuation pauses, we
322 // do not copy objects that are explicitly marked, what we have to do
323 // for the root regions is to scan them and mark all objects reachable
324 // from them. According to the SATB assumptions, we only need to visit
325 // each object once during marking. So, as long as we finish this scan
326 // before the next evacuation pause, we can copy the objects from the
327 // root regions without having to mark them or do anything else to them.
328 //
329 // Currently, we only support root region scanning once (at the start
330 // of the marking cycle) and the root regions are all the survivor
331 // regions populated during the initial-mark pause.
332 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
333 private:
334 const G1SurvivorRegions* _survivors;
335 G1ConcurrentMark* _cm;
336
337 volatile bool _scan_in_progress;
338 volatile bool _should_abort;
339 volatile int _claimed_survivor_index;
340
341 void notify_scan_done();
342
343 public:
344 G1CMRootRegions();
345 // We actually do most of the initialization in this method.
346 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
347
348 // Reset the claiming / scanning of the root regions.
349 void prepare_for_scan();
350
351 // Forces get_next() to return NULL so that the iteration aborts early.
352 void abort() { _should_abort = true; }
353
354 // Return true if the CM thread are actively scanning root regions,
355 // false otherwise.
356 bool scan_in_progress() { return _scan_in_progress; }
357
358 // Claim the next root region to scan atomically, or return NULL if
359 // all have been claimed.
360 HeapRegion* claim_next();
361
362 // The number of root regions to scan.
363 uint num_root_regions() const;
364
365 void cancel_scan();
366
367 // Flag that we're done with root region scanning and notify anyone
368 // who's waiting on it. If aborted is false, assume that all regions
369 // have been claimed.
370 void scan_finished();
371
372 // If CM threads are still scanning root regions, wait until they
373 // are done. Return true if we had to wait, false otherwise.
374 bool wait_until_scan_finished();
375 };
376
377 class ConcurrentMarkThread;
378
379 class G1ConcurrentMark: public CHeapObj<mtGC> {
380 friend class ConcurrentMarkThread;
381 friend class G1ParNoteEndTask;
382 friend class G1VerifyLiveDataClosure;
383 friend class G1CMRefProcTaskProxy;
384 friend class G1CMRefProcTaskExecutor;
385 friend class G1CMKeepAliveAndDrainClosure;
386 friend class G1CMDrainMarkingStackClosure;
387 friend class G1CMBitMapClosure;
388 friend class G1CMConcurrentMarkingTask;
389 friend class G1CMRemarkTask;
390 friend class G1CMTask;
391
392 protected:
393 ConcurrentMarkThread* _cmThread; // The thread doing the work
394 G1CollectedHeap* _g1h; // The heap
395 uint _parallel_marking_threads; // The number of marking
396 // threads we're using
397 uint _max_parallel_marking_threads; // Max number of marking
398 // threads we'll ever use
399 double _sleep_factor; // How much we have to sleep, with
400 // respect to the work we just did, to
401 // meet the marking overhead goal
402 double _marking_task_overhead; // Marking target overhead for
403 // a single task
404
405 FreeRegionList _cleanup_list;
406
407 // Concurrent marking support structures
408 G1CMBitMap _markBitMap1;
409 G1CMBitMap _markBitMap2;
410 G1CMBitMap* _prevMarkBitMap; // Completed mark bitmap
411 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
412
413 // Heap bounds
414 HeapWord* _heap_start;
415 HeapWord* _heap_end;
416
417 // Root region tracking and claiming
418 G1CMRootRegions _root_regions;
419
420 // For gray objects
421 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger
422 HeapWord* volatile _finger; // The global finger, region aligned,
423 // always points to the end of the
424 // last claimed region
425
426 // Marking tasks
427 uint _max_worker_id;// Maximum worker id
428 uint _active_tasks; // Task num currently active
429 G1CMTask** _tasks; // Task queue array (max_worker_id len)
430 G1CMTaskQueueSet* _task_queues; // Task queue set
431 ParallelTaskTerminator _terminator; // For termination
432
433 // Two sync barriers that are used to synchronize tasks when an
434 // overflow occurs. The algorithm is the following. All tasks enter
435 // the first one to ensure that they have all stopped manipulating
436 // the global data structures. After they exit it, they re-initialize
437 // their data structures and task 0 re-initializes the global data
438 // structures. Then, they enter the second sync barrier. This
439 // ensure, that no task starts doing work before all data
440 // structures (local and global) have been re-initialized. When they
441 // exit it, they are free to start working again.
442 WorkGangBarrierSync _first_overflow_barrier_sync;
443 WorkGangBarrierSync _second_overflow_barrier_sync;
444
445 // This is set by any task, when an overflow on the global data
446 // structures is detected
447 volatile bool _has_overflown;
448 // True: marking is concurrent, false: we're in remark
449 volatile bool _concurrent;
450 // Set at the end of a Full GC so that marking aborts
451 volatile bool _has_aborted;
452
453 // Used when remark aborts due to an overflow to indicate that
454 // another concurrent marking phase should start
455 volatile bool _restart_for_overflow;
456
457 // This is true from the very start of concurrent marking until the
458 // point when all the tasks complete their work. It is really used
459 // to determine the points between the end of concurrent marking and
460 // time of remark.
461 volatile bool _concurrent_marking_in_progress;
462
463 ConcurrentGCTimer* _gc_timer_cm;
464
465 G1OldTracer* _gc_tracer_cm;
466
467 // All of these times are in ms
468 NumberSeq _init_times;
469 NumberSeq _remark_times;
470 NumberSeq _remark_mark_times;
471 NumberSeq _remark_weak_ref_times;
472 NumberSeq _cleanup_times;
473 double _total_counting_time;
474 double _total_rs_scrub_time;
475
476 double* _accum_task_vtime; // Accumulated task vtime
477
478 WorkGang* _parallel_workers;
479
480 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
481 void weakRefsWork(bool clear_all_soft_refs);
482
483 void swapMarkBitMaps();
484
485 // It resets the global marking data structures, as well as the
486 // task local ones; should be called during initial mark.
487 void reset();
488
489 // Resets all the marking data structures. Called when we have to restart
490 // marking or when marking completes (via set_non_marking_state below).
491 void reset_marking_state();
492
493 // We do this after we're done with marking so that the marking data
494 // structures are initialized to a sensible and predictable state.
495 void set_non_marking_state();
496
497 // Called to indicate how many threads are currently active.
498 void set_concurrency(uint active_tasks);
499
500 // It should be called to indicate which phase we're in (concurrent
501 // mark or remark) and how many threads are currently active.
502 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
503
504 // Prints all gathered CM-related statistics
505 void print_stats();
506
507 bool cleanup_list_is_empty() {
508 return _cleanup_list.is_empty();
509 }
510
511 // Accessor methods
512 uint parallel_marking_threads() const { return _parallel_marking_threads; }
513 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
514 double sleep_factor() { return _sleep_factor; }
515 double marking_task_overhead() { return _marking_task_overhead;}
516
517 HeapWord* finger() { return _finger; }
518 bool concurrent() { return _concurrent; }
519 uint active_tasks() { return _active_tasks; }
520 ParallelTaskTerminator* terminator() { return &_terminator; }
521
522 // It claims the next available region to be scanned by a marking
523 // task/thread. It might return NULL if the next region is empty or
524 // we have run out of regions. In the latter case, out_of_regions()
525 // determines whether we've really run out of regions or the task
526 // should call claim_region() again. This might seem a bit
527 // awkward. Originally, the code was written so that claim_region()
528 // either successfully returned with a non-empty region or there
529 // were no more regions to be claimed. The problem with this was
530 // that, in certain circumstances, it iterated over large chunks of
531 // the heap finding only empty regions and, while it was working, it
532 // was preventing the calling task to call its regular clock
533 // method. So, this way, each task will spend very little time in
534 // claim_region() and is allowed to call the regular clock method
535 // frequently.
536 HeapRegion* claim_region(uint worker_id);
537
538 // It determines whether we've run out of regions to scan. Note that
539 // the finger can point past the heap end in case the heap was expanded
540 // to satisfy an allocation without doing a GC. This is fine, because all
541 // objects in those regions will be considered live anyway because of
542 // SATB guarantees (i.e. their TAMS will be equal to bottom).
543 bool out_of_regions() { return _finger >= _heap_end; }
544
545 // Returns the task with the given id
546 G1CMTask* task(int id) {
547 assert(0 <= id && id < (int) _active_tasks,
548 "task id not within active bounds");
549 return _tasks[id];
550 }
551
552 // Returns the task queue with the given id
553 G1CMTaskQueue* task_queue(int id) {
554 assert(0 <= id && id < (int) _active_tasks,
555 "task queue id not within active bounds");
556 return (G1CMTaskQueue*) _task_queues->queue(id);
557 }
558
559 // Returns the task queue set
560 G1CMTaskQueueSet* task_queues() { return _task_queues; }
561
562 // Access / manipulation of the overflow flag which is set to
563 // indicate that the global stack has overflown
564 bool has_overflown() { return _has_overflown; }
565 void set_has_overflown() { _has_overflown = true; }
566 void clear_has_overflown() { _has_overflown = false; }
567 bool restart_for_overflow() { return _restart_for_overflow; }
568
569 // Methods to enter the two overflow sync barriers
570 void enter_first_sync_barrier(uint worker_id);
571 void enter_second_sync_barrier(uint worker_id);
572
573 // Card index of the bottom of the G1 heap. Used for biasing indices into
574 // the card bitmaps.
575 intptr_t _heap_bottom_card_num;
576
577 // Set to true when initialization is complete
578 bool _completed_initialization;
579
580 // end_timer, true to end gc timer after ending concurrent phase.
581 void register_concurrent_phase_end_common(bool end_timer);
582
583 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
584 // true, periodically insert checks to see if this method should exit prematurely.
585 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
586 public:
587 // Manipulation of the global mark stack.
588 // The push and pop operations are used by tasks for transfers
589 // between task-local queues and the global mark stack.
590 bool mark_stack_push(G1TaskQueueEntry* arr) {
591 if (!_global_mark_stack.par_push_chunk(arr)) {
592 set_has_overflown();
593 return false;
594 }
595 return true;
596 }
597 bool mark_stack_pop(G1TaskQueueEntry* arr) {
598 return _global_mark_stack.par_pop_chunk(arr);
599 }
600 size_t mark_stack_size() { return _global_mark_stack.size(); }
601 size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; }
602 bool mark_stack_empty() { return _global_mark_stack.is_empty(); }
603
604 G1CMRootRegions* root_regions() { return &_root_regions; }
605
606 bool concurrent_marking_in_progress() {
607 return _concurrent_marking_in_progress;
608 }
609 void set_concurrent_marking_in_progress() {
610 _concurrent_marking_in_progress = true;
611 }
612 void clear_concurrent_marking_in_progress() {
613 _concurrent_marking_in_progress = false;
614 }
615
616 void concurrent_cycle_start();
617 void concurrent_cycle_end();
618
619 void update_accum_task_vtime(int i, double vtime) {
620 _accum_task_vtime[i] += vtime;
621 }
622
623 double all_task_accum_vtime() {
624 double ret = 0.0;
625 for (uint i = 0; i < _max_worker_id; ++i)
626 ret += _accum_task_vtime[i];
627 return ret;
628 }
629
630 // Attempts to steal an object from the task queues of other tasks
631 bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry);
632
633 G1ConcurrentMark(G1CollectedHeap* g1h,
634 G1RegionToSpaceMapper* prev_bitmap_storage,
635 G1RegionToSpaceMapper* next_bitmap_storage);
636 ~G1ConcurrentMark();
637
638 ConcurrentMarkThread* cmThread() { return _cmThread; }
639
640 G1CMBitMapRO prevMarkBitMap() const { return G1CMBitMapRO(_prevMarkBitMap); }
641 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
642
643 // Returns the number of GC threads to be used in a concurrent
644 // phase based on the number of GC threads being used in a STW
645 // phase.
646 uint scale_parallel_threads(uint n_par_threads);
647
648 // Calculates the number of GC threads to be used in a concurrent phase.
649 uint calc_parallel_marking_threads();
650
651 // The following three are interaction between CM and
652 // G1CollectedHeap
653
654 // This notifies CM that a root during initial-mark needs to be
655 // grayed. It is MT-safe. hr is the region that
656 // contains the object and it's passed optionally from callers who
657 // might already have it (no point in recalculating it).
658 inline void grayRoot(oop obj,
659 HeapRegion* hr = NULL);
660
661 // Prepare internal data structures for the next mark cycle. This includes clearing
662 // the next mark bitmap and some internal data structures. This method is intended
663 // to be called concurrently to the mutator. It will yield to safepoint requests.
664 void cleanup_for_next_mark();
665
666 // Clear the previous marking bitmap during safepoint.
667 void clear_prev_bitmap(WorkGang* workers);
668
669 // Return whether the next mark bitmap has no marks set. To be used for assertions
670 // only. Will not yield to pause requests.
671 bool nextMarkBitmapIsClear();
672
673 // These two do the work that needs to be done before and after the
674 // initial root checkpoint. Since this checkpoint can be done at two
675 // different points (i.e. an explicit pause or piggy-backed on a
676 // young collection), then it's nice to be able to easily share the
677 // pre/post code. It might be the case that we can put everything in
678 // the post method. TP
679 void checkpointRootsInitialPre();
680 void checkpointRootsInitialPost();
681
682 // Scan all the root regions and mark everything reachable from
683 // them.
684 void scan_root_regions();
685
686 // Scan a single root region and mark everything reachable from it.
687 void scanRootRegion(HeapRegion* hr);
688
689 // Do concurrent phase of marking, to a tentative transitive closure.
690 void mark_from_roots();
691
692 void checkpointRootsFinal(bool clear_all_soft_refs);
693 void checkpointRootsFinalWork();
694 void cleanup();
695 void complete_cleanup();
696
697 // Mark in the previous bitmap. NB: this is usually read-only, so use
698 // this carefully!
699 inline void markPrev(oop p);
700
701 // Clears marks for all objects in the given range, for the prev or
702 // next bitmaps. NB: the previous bitmap is usually
703 // read-only, so use this carefully!
704 void clearRangePrevBitmap(MemRegion mr);
705
706 // Verify that there are no CSet oops on the stacks (taskqueues /
707 // global mark stack) and fingers (global / per-task).
708 // If marking is not in progress, it's a no-op.
709 void verify_no_cset_oops() PRODUCT_RETURN;
710
711 inline bool isPrevMarked(oop p) const;
712
713 inline bool do_yield_check();
714
715 // Abandon current marking iteration due to a Full GC.
716 void abort();
717
718 bool has_aborted() { return _has_aborted; }
719
720 void print_summary_info();
721
722 void print_worker_threads_on(outputStream* st) const;
723 void threads_do(ThreadClosure* tc) const;
724
725 void print_on_error(outputStream* st) const;
726
727 // Attempts to mark the given object on the next mark bitmap.
728 inline bool par_mark(oop obj);
729
730 // Returns true if initialization was successfully completed.
731 bool completed_initialization() const {
732 return _completed_initialization;
733 }
734
735 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
736 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
737
738 private:
739 // Clear (Reset) all liveness count data.
740 void clear_live_data(WorkGang* workers);
741
742 #ifdef ASSERT
743 // Verify all of the above data structures that they are in initial state.
744 void verify_live_data_clear();
745 #endif
746
747 // Aggregates the per-card liveness data based on the current marking. Also sets
748 // the amount of marked bytes for each region.
749 void create_live_data();
750
751 void finalize_live_data();
752
753 void verify_live_data();
754 };
755
756 // A class representing a marking task.
757 class G1CMTask : public TerminatorTerminator {
758 private:
759 enum PrivateConstants {
760 // The regular clock call is called once the scanned words reaches
761 // this limit
762 words_scanned_period = 12*1024,
763 // The regular clock call is called once the number of visited
764 // references reaches this limit
765 refs_reached_period = 1024,
766 // Initial value for the hash seed, used in the work stealing code
767 init_hash_seed = 17
768 };
769
770 G1CMObjArrayProcessor _objArray_processor;
771
772 uint _worker_id;
773 G1CollectedHeap* _g1h;
774 G1ConcurrentMark* _cm;
775 G1CMBitMap* _nextMarkBitMap;
776 // the task queue of this task
777 G1CMTaskQueue* _task_queue;
778 private:
779 // the task queue set---needed for stealing
780 G1CMTaskQueueSet* _task_queues;
781 // indicates whether the task has been claimed---this is only for
782 // debugging purposes
783 bool _claimed;
784
785 // number of calls to this task
786 int _calls;
787
788 // when the virtual timer reaches this time, the marking step should
789 // exit
790 double _time_target_ms;
791 // the start time of the current marking step
792 double _start_time_ms;
793
794 // the oop closure used for iterations over oops
795 G1CMOopClosure* _cm_oop_closure;
796
797 // the region this task is scanning, NULL if we're not scanning any
798 HeapRegion* _curr_region;
799 // the local finger of this task, NULL if we're not scanning a region
800 HeapWord* _finger;
801 // limit of the region this task is scanning, NULL if we're not scanning one
802 HeapWord* _region_limit;
803
804 // the number of words this task has scanned
805 size_t _words_scanned;
806 // When _words_scanned reaches this limit, the regular clock is
807 // called. Notice that this might be decreased under certain
808 // circumstances (i.e. when we believe that we did an expensive
809 // operation).
810 size_t _words_scanned_limit;
811 // the initial value of _words_scanned_limit (i.e. what it was
812 // before it was decreased).
813 size_t _real_words_scanned_limit;
814
815 // the number of references this task has visited
816 size_t _refs_reached;
817 // When _refs_reached reaches this limit, the regular clock is
818 // called. Notice this this might be decreased under certain
819 // circumstances (i.e. when we believe that we did an expensive
820 // operation).
821 size_t _refs_reached_limit;
822 // the initial value of _refs_reached_limit (i.e. what it was before
823 // it was decreased).
824 size_t _real_refs_reached_limit;
825
826 // used by the work stealing stuff
827 int _hash_seed;
828 // if this is true, then the task has aborted for some reason
829 bool _has_aborted;
830 // set when the task aborts because it has met its time quota
831 bool _has_timed_out;
832 // true when we're draining SATB buffers; this avoids the task
833 // aborting due to SATB buffers being available (as we're already
834 // dealing with them)
835 bool _draining_satb_buffers;
836
837 // number sequence of past step times
838 NumberSeq _step_times_ms;
839 // elapsed time of this task
840 double _elapsed_time_ms;
841 // termination time of this task
842 double _termination_time_ms;
843 // when this task got into the termination protocol
844 double _termination_start_time_ms;
845
846 // true when the task is during a concurrent phase, false when it is
847 // in the remark phase (so, in the latter case, we do not have to
848 // check all the things that we have to check during the concurrent
849 // phase, i.e. SATB buffer availability...)
850 bool _concurrent;
851
852 TruncatedSeq _marking_step_diffs_ms;
853
854 // it updates the local fields after this task has claimed
855 // a new region to scan
856 void setup_for_region(HeapRegion* hr);
857 // it brings up-to-date the limit of the region
858 void update_region_limit();
859
860 // called when either the words scanned or the refs visited limit
861 // has been reached
862 void reached_limit();
863 // recalculates the words scanned and refs visited limits
864 void recalculate_limits();
865 // decreases the words scanned and refs visited limits when we reach
866 // an expensive operation
867 void decrease_limits();
868 // it checks whether the words scanned or refs visited reached their
869 // respective limit and calls reached_limit() if they have
870 void check_limits() {
871 if (_words_scanned >= _words_scanned_limit ||
872 _refs_reached >= _refs_reached_limit) {
873 reached_limit();
874 }
875 }
876 // this is supposed to be called regularly during a marking step as
877 // it checks a bunch of conditions that might cause the marking step
878 // to abort
879 void regular_clock_call();
880 bool concurrent() { return _concurrent; }
881
882 // Test whether obj might have already been passed over by the
883 // mark bitmap scan, and so needs to be pushed onto the mark stack.
884 bool is_below_finger(oop obj, HeapWord* global_finger) const;
885
886 template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
887 public:
888 // Apply the closure on the given area of the objArray. Return the number of words
889 // scanned.
890 inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
891 // It resets the task; it should be called right at the beginning of
892 // a marking phase.
893 void reset(G1CMBitMap* _nextMarkBitMap);
894 // it clears all the fields that correspond to a claimed region.
895 void clear_region_fields();
896
897 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
898
899 // The main method of this class which performs a marking step
900 // trying not to exceed the given duration. However, it might exit
901 // prematurely, according to some conditions (i.e. SATB buffers are
902 // available for processing).
903 void do_marking_step(double target_ms,
904 bool do_termination,
905 bool is_serial);
906
907 // These two calls start and stop the timer
908 void record_start_time() {
909 _elapsed_time_ms = os::elapsedTime() * 1000.0;
910 }
911 void record_end_time() {
912 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
913 }
914
915 // returns the worker ID associated with this task.
916 uint worker_id() { return _worker_id; }
917
918 // From TerminatorTerminator. It determines whether this task should
919 // exit the termination protocol after it's entered it.
920 virtual bool should_exit_termination();
921
922 // Resets the local region fields after a task has finished scanning a
923 // region; or when they have become stale as a result of the region
924 // being evacuated.
925 void giveup_current_region();
926
927 HeapWord* finger() { return _finger; }
928
929 bool has_aborted() { return _has_aborted; }
930 void set_has_aborted() { _has_aborted = true; }
931 void clear_has_aborted() { _has_aborted = false; }
932 bool has_timed_out() { return _has_timed_out; }
933 bool claimed() { return _claimed; }
934
935 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
936
937 // Increment the number of references this task has visited.
938 void increment_refs_reached() { ++_refs_reached; }
939
940 // Grey the object by marking it. If not already marked, push it on
941 // the local queue if below the finger.
942 // obj is below its region's NTAMS.
943 inline void make_reference_grey(oop obj);
944
945 // Grey the object (by calling make_grey_reference) if required,
946 // e.g. obj is below its containing region's NTAMS.
947 // Precondition: obj is a valid heap object.
948 inline void deal_with_reference(oop obj);
949
950 // It scans an object and visits its children.
951 inline void scan_task_entry(G1TaskQueueEntry task_entry);
952
953 // It pushes an object on the local queue.
954 inline void push(G1TaskQueueEntry task_entry);
955
956 // Move entries to the global stack.
957 void move_entries_to_global_stack();
958 // Move entries from the global stack, return true if we were successful to do so.
959 bool get_entries_from_global_stack();
960
961 // It pops and scans objects from the local queue. If partially is
962 // true, then it stops when the queue size is of a given limit. If
963 // partially is false, then it stops when the queue is empty.
964 void drain_local_queue(bool partially);
965 // It moves entries from the global stack to the local queue and
966 // drains the local queue. If partially is true, then it stops when
967 // both the global stack and the local queue reach a given size. If
968 // partially if false, it tries to empty them totally.
969 void drain_global_stack(bool partially);
970 // It keeps picking SATB buffers and processing them until no SATB
971 // buffers are available.
972 void drain_satb_buffers();
973
974 // moves the local finger to a new location
975 inline void move_finger_to(HeapWord* new_finger) {
976 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
977 _finger = new_finger;
978 }
979
980 G1CMTask(uint worker_id,
981 G1ConcurrentMark *cm,
982 G1CMTaskQueue* task_queue,
983 G1CMTaskQueueSet* task_queues);
984
985 // it prints statistics associated with this task
986 void print_stats();
987 };
988
989 // Class that's used to to print out per-region liveness
990 // information. It's currently used at the end of marking and also
991 // after we sort the old regions at the end of the cleanup operation.
992 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
993 private:
994 // Accumulators for these values.
995 size_t _total_used_bytes;
996 size_t _total_capacity_bytes;
997 size_t _total_prev_live_bytes;
998 size_t _total_next_live_bytes;
999
1000 // Accumulator for the remembered set size
1001 size_t _total_remset_bytes;
1002
1003 // Accumulator for strong code roots memory size
1004 size_t _total_strong_code_roots_bytes;
1005
1006 static double perc(size_t val, size_t total) {
1007 if (total == 0) {
1008 return 0.0;
1009 } else {
1010 return 100.0 * ((double) val / (double) total);
1011 }
1012 }
1013
1014 static double bytes_to_mb(size_t val) {
1015 return (double) val / (double) M;
1016 }
1017
1018 public:
1019 // The header and footer are printed in the constructor and
1020 // destructor respectively.
1021 G1PrintRegionLivenessInfoClosure(const char* phase_name);
1022 virtual bool doHeapRegion(HeapRegion* r);
1023 ~G1PrintRegionLivenessInfoClosure();
1024 };
1025
1026 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP