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