1 /*
2 * Copyright (c) 2001, 2013, 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_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
27
28 #include "gc_implementation/g1/heapRegionSet.hpp"
29 #include "gc_implementation/shared/gcId.hpp"
30 #include "utilities/taskqueue.hpp"
31
32 class G1CollectedHeap;
33 class CMTask;
34 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
35 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
36
37 // Closure used by CM during concurrent reference discovery
38 // and reference processing (during remarking) to determine
39 // if a particular object is alive. It is primarily used
40 // to determine if referents of discovered reference objects
41 // are alive. An instance is also embedded into the
42 // reference processor as the _is_alive_non_header field
43 class G1CMIsAliveClosure: public BoolObjectClosure {
44 G1CollectedHeap* _g1;
45 public:
46 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
47
48 bool do_object_b(oop obj);
49 };
50
51 // A generic CM bit map. This is essentially a wrapper around the BitMap
52 // class, with one bit per (1<<_shifter) HeapWords.
53
54 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
55 protected:
56 HeapWord* _bmStartWord; // base address of range covered by map
57 size_t _bmWordSize; // map size (in #HeapWords covered)
58 const int _shifter; // map to char or bit
59 VirtualSpace _virtual_space; // underlying the bit map
60 BitMap _bm; // the bit map itself
61
62 public:
63 // constructor
64 CMBitMapRO(int shifter);
65
66 enum { do_yield = true };
67
68 // inquiries
69 HeapWord* startWord() const { return _bmStartWord; }
70 size_t sizeInWords() const { return _bmWordSize; }
71 // the following is one past the last word in space
72 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
73
74 // read marks
75
76 bool isMarked(HeapWord* addr) const {
77 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
78 "outside underlying space?");
79 return _bm.at(heapWordToOffset(addr));
80 }
81
82 // iteration
83 inline bool iterate(BitMapClosure* cl, MemRegion mr);
84 inline bool iterate(BitMapClosure* cl);
85
86 // Return the address corresponding to the next marked bit at or after
87 // "addr", and before "limit", if "limit" is non-NULL. If there is no
88 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
89 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
90 HeapWord* limit = NULL) const;
91 // Return the address corresponding to the next unmarked bit at or after
92 // "addr", and before "limit", if "limit" is non-NULL. If there is no
93 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
94 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
95 HeapWord* limit = NULL) const;
96
97 // conversion utilities
98 HeapWord* offsetToHeapWord(size_t offset) const {
99 return _bmStartWord + (offset << _shifter);
100 }
101 size_t heapWordToOffset(HeapWord* addr) const {
102 return pointer_delta(addr, _bmStartWord) >> _shifter;
103 }
104 int heapWordDiffToOffsetDiff(size_t diff) const;
105
106 // The argument addr should be the start address of a valid object
107 HeapWord* nextObject(HeapWord* addr) {
108 oop obj = (oop) addr;
109 HeapWord* res = addr + obj->size();
110 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
111 return res;
112 }
113
114 void print_on_error(outputStream* st, const char* prefix) const;
115
116 // debugging
117 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
118 };
119
120 class CMBitMap : public CMBitMapRO {
121
122 public:
123 // constructor
124 CMBitMap(int shifter) :
125 CMBitMapRO(shifter) {}
126
127 // Allocates the back store for the marking bitmap
128 bool allocate(ReservedSpace heap_rs);
129
130 // write marks
131 void mark(HeapWord* addr) {
132 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
133 "outside underlying space?");
134 _bm.set_bit(heapWordToOffset(addr));
135 }
136 void clear(HeapWord* addr) {
137 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
138 "outside underlying space?");
139 _bm.clear_bit(heapWordToOffset(addr));
140 }
141 bool parMark(HeapWord* addr) {
142 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
143 "outside underlying space?");
144 return _bm.par_set_bit(heapWordToOffset(addr));
145 }
146 bool parClear(HeapWord* addr) {
147 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
148 "outside underlying space?");
149 return _bm.par_clear_bit(heapWordToOffset(addr));
150 }
151 void markRange(MemRegion mr);
152 void clearAll();
153 void clearRange(MemRegion mr);
154
155 // Starting at the bit corresponding to "addr" (inclusive), find the next
156 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
157 // the end of this run (stopping at "end_addr"). Return the MemRegion
158 // covering from the start of the region corresponding to the first bit
159 // of the run to the end of the region corresponding to the last bit of
160 // the run. If there is no "1" bit at or after "addr", return an empty
161 // MemRegion.
162 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
163 };
164
165 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
166 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
167 VirtualSpace _virtual_space; // Underlying backing store for actual stack
168 ConcurrentMark* _cm;
169 oop* _base; // bottom of stack
170 jint _index; // one more than last occupied index
171 jint _capacity; // max #elements
172 jint _saved_index; // value of _index saved at start of GC
173 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
174
175 bool _overflow;
176 bool _should_expand;
177 DEBUG_ONLY(bool _drain_in_progress;)
178 DEBUG_ONLY(bool _drain_in_progress_yields;)
179
180 public:
181 CMMarkStack(ConcurrentMark* cm);
182 ~CMMarkStack();
183
184 #ifndef PRODUCT
185 jint max_depth() const {
186 return _max_depth;
187 }
188 #endif
189
190 bool allocate(size_t capacity);
191
192 oop pop() {
193 if (!isEmpty()) {
194 return _base[--_index] ;
195 }
196 return NULL;
197 }
198
199 // If overflow happens, don't do the push, and record the overflow.
200 // *Requires* that "ptr" is already marked.
201 void push(oop ptr) {
202 if (isFull()) {
203 // Record overflow.
204 _overflow = true;
205 return;
206 } else {
207 _base[_index++] = ptr;
208 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
209 }
210 }
211 // Non-block impl. Note: concurrency is allowed only with other
212 // "par_push" operations, not with "pop" or "drain". We would need
213 // parallel versions of them if such concurrency was desired.
214 void par_push(oop ptr);
215
216 // Pushes the first "n" elements of "ptr_arr" on the stack.
217 // Non-block impl. Note: concurrency is allowed only with other
218 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
219 void par_adjoin_arr(oop* ptr_arr, int n);
220
221 // Pushes the first "n" elements of "ptr_arr" on the stack.
222 // Locking impl: concurrency is allowed only with
223 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
224 // locking strategy.
225 void par_push_arr(oop* ptr_arr, int n);
226
227 // If returns false, the array was empty. Otherwise, removes up to "max"
228 // elements from the stack, and transfers them to "ptr_arr" in an
229 // unspecified order. The actual number transferred is given in "n" ("n
230 // == 0" is deliberately redundant with the return value.) Locking impl:
231 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
232 // operations, which use the same locking strategy.
233 bool par_pop_arr(oop* ptr_arr, int max, int* n);
234
235 // Drain the mark stack, applying the given closure to all fields of
236 // objects on the stack. (That is, continue until the stack is empty,
237 // even if closure applications add entries to the stack.) The "bm"
238 // argument, if non-null, may be used to verify that only marked objects
239 // are on the mark stack. If "yield_after" is "true", then the
240 // concurrent marker performing the drain offers to yield after
241 // processing each object. If a yield occurs, stops the drain operation
242 // and returns false. Otherwise, returns true.
243 template<class OopClosureClass>
244 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
245
246 bool isEmpty() { return _index == 0; }
247 bool isFull() { return _index == _capacity; }
248 int maxElems() { return _capacity; }
249
250 bool overflow() { return _overflow; }
251 void clear_overflow() { _overflow = false; }
252
253 bool should_expand() const { return _should_expand; }
254 void set_should_expand();
255
256 // Expand the stack, typically in response to an overflow condition
257 void expand();
258
259 int size() { return _index; }
260
261 void setEmpty() { _index = 0; clear_overflow(); }
262
263 // Record the current index.
264 void note_start_of_gc();
265
266 // Make sure that we have not added any entries to the stack during GC.
267 void note_end_of_gc();
268
269 // iterate over the oops in the mark stack, up to the bound recorded via
270 // the call above.
271 void oops_do(OopClosure* f);
272 };
273
274 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
275 private:
276 #ifndef PRODUCT
277 uintx _num_remaining;
278 bool _force;
279 #endif // !defined(PRODUCT)
280
281 public:
282 void init() PRODUCT_RETURN;
283 void update() PRODUCT_RETURN;
284 bool should_force() PRODUCT_RETURN_( return false; );
285 };
286
287 // this will enable a variety of different statistics per GC task
288 #define _MARKING_STATS_ 0
289 // this will enable the higher verbose levels
290 #define _MARKING_VERBOSE_ 0
291
292 #if _MARKING_STATS_
293 #define statsOnly(statement) \
294 do { \
295 statement ; \
296 } while (0)
297 #else // _MARKING_STATS_
298 #define statsOnly(statement) \
299 do { \
300 } while (0)
301 #endif // _MARKING_STATS_
302
303 typedef enum {
304 no_verbose = 0, // verbose turned off
305 stats_verbose, // only prints stats at the end of marking
306 low_verbose, // low verbose, mostly per region and per major event
307 medium_verbose, // a bit more detailed than low
308 high_verbose // per object verbose
309 } CMVerboseLevel;
310
311 class YoungList;
312
313 // Root Regions are regions that are not empty at the beginning of a
314 // marking cycle and which we might collect during an evacuation pause
315 // while the cycle is active. Given that, during evacuation pauses, we
316 // do not copy objects that are explicitly marked, what we have to do
317 // for the root regions is to scan them and mark all objects reachable
318 // from them. According to the SATB assumptions, we only need to visit
319 // each object once during marking. So, as long as we finish this scan
320 // before the next evacuation pause, we can copy the objects from the
321 // root regions without having to mark them or do anything else to them.
322 //
323 // Currently, we only support root region scanning once (at the start
324 // of the marking cycle) and the root regions are all the survivor
325 // regions populated during the initial-mark pause.
326 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
327 private:
328 YoungList* _young_list;
329 ConcurrentMark* _cm;
330
331 volatile bool _scan_in_progress;
332 volatile bool _should_abort;
333 HeapRegion* volatile _next_survivor;
334
335 public:
336 CMRootRegions();
337 // We actually do most of the initialization in this method.
338 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
339
340 // Reset the claiming / scanning of the root regions.
341 void prepare_for_scan();
342
343 // Forces get_next() to return NULL so that the iteration aborts early.
344 void abort() { _should_abort = true; }
345
346 // Return true if the CM thread are actively scanning root regions,
347 // false otherwise.
348 bool scan_in_progress() { return _scan_in_progress; }
349
350 // Claim the next root region to scan atomically, or return NULL if
351 // all have been claimed.
352 HeapRegion* claim_next();
353
354 // Flag that we're done with root region scanning and notify anyone
355 // who's waiting on it. If aborted is false, assume that all regions
356 // have been claimed.
357 void scan_finished();
358
359 // If CM threads are still scanning root regions, wait until they
360 // are done. Return true if we had to wait, false otherwise.
361 bool wait_until_scan_finished();
362 };
363
364 class ConcurrentMarkThread;
365
366 class ConcurrentMark: public CHeapObj<mtGC> {
367 friend class CMMarkStack;
368 friend class ConcurrentMarkThread;
369 friend class CMTask;
370 friend class CMBitMapClosure;
371 friend class CMGlobalObjectClosure;
372 friend class CMRemarkTask;
373 friend class CMConcurrentMarkingTask;
374 friend class G1ParNoteEndTask;
375 friend class CalcLiveObjectsClosure;
376 friend class G1CMRefProcTaskProxy;
377 friend class G1CMRefProcTaskExecutor;
378 friend class G1CMKeepAliveAndDrainClosure;
379 friend class G1CMDrainMarkingStackClosure;
380
381 protected:
382 ConcurrentMarkThread* _cmThread; // The thread doing the work
383 G1CollectedHeap* _g1h; // The heap
384 uint _parallel_marking_threads; // The number of marking
385 // threads we're using
386 uint _max_parallel_marking_threads; // Max number of marking
387 // threads we'll ever use
388 double _sleep_factor; // How much we have to sleep, with
389 // respect to the work we just did, to
390 // meet the marking overhead goal
391 double _marking_task_overhead; // Marking target overhead for
392 // a single task
393
394 // Same as the two above, but for the cleanup task
395 double _cleanup_sleep_factor;
396 double _cleanup_task_overhead;
397
398 FreeRegionList _cleanup_list;
399
400 // Concurrent marking support structures
401 CMBitMap _markBitMap1;
402 CMBitMap _markBitMap2;
403 CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
404 CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
405
406 BitMap _region_bm;
407 BitMap _card_bm;
408
409 // Heap bounds
410 HeapWord* _heap_start;
411 HeapWord* _heap_end;
412
413 // Root region tracking and claiming
414 CMRootRegions _root_regions;
415
416 // For gray objects
417 CMMarkStack _markStack; // Grey objects behind global finger
418 HeapWord* volatile _finger; // The global finger, region aligned,
419 // always points to the end of the
420 // last claimed region
421
422 // Marking tasks
423 uint _max_worker_id;// Maximum worker id
424 uint _active_tasks; // Task num currently active
425 CMTask** _tasks; // Task queue array (max_worker_id len)
426 CMTaskQueueSet* _task_queues; // Task queue set
427 ParallelTaskTerminator _terminator; // For termination
428
429 // Two sync barriers that are used to synchronize tasks when an
430 // overflow occurs. The algorithm is the following. All tasks enter
431 // the first one to ensure that they have all stopped manipulating
432 // the global data structures. After they exit it, they re-initialize
433 // their data structures and task 0 re-initializes the global data
434 // structures. Then, they enter the second sync barrier. This
435 // ensure, that no task starts doing work before all data
436 // structures (local and global) have been re-initialized. When they
437 // exit it, they are free to start working again.
438 WorkGangBarrierSync _first_overflow_barrier_sync;
439 WorkGangBarrierSync _second_overflow_barrier_sync;
440
441 // This is set by any task, when an overflow on the global data
442 // structures is detected
443 volatile bool _has_overflown;
444 // True: marking is concurrent, false: we're in remark
445 volatile bool _concurrent;
446 // Set at the end of a Full GC so that marking aborts
447 volatile bool _has_aborted;
448 GCId _aborted_gc_id;
449
450 // Used when remark aborts due to an overflow to indicate that
451 // another concurrent marking phase should start
452 volatile bool _restart_for_overflow;
453
454 // This is true from the very start of concurrent marking until the
455 // point when all the tasks complete their work. It is really used
456 // to determine the points between the end of concurrent marking and
457 // time of remark.
458 volatile bool _concurrent_marking_in_progress;
459
460 // Verbose level
461 CMVerboseLevel _verbose_level;
462
463 // All of these times are in ms
464 NumberSeq _init_times;
465 NumberSeq _remark_times;
466 NumberSeq _remark_mark_times;
467 NumberSeq _remark_weak_ref_times;
468 NumberSeq _cleanup_times;
469 double _total_counting_time;
470 double _total_rs_scrub_time;
471
472 double* _accum_task_vtime; // Accumulated task vtime
473
474 FlexibleWorkGang* _parallel_workers;
475
476 ForceOverflowSettings _force_overflow_conc;
477 ForceOverflowSettings _force_overflow_stw;
478
479 void weakRefsWork(bool clear_all_soft_refs);
480
481 void swapMarkBitMaps();
482
483 // It resets the global marking data structures, as well as the
484 // task local ones; should be called during initial mark.
485 void reset();
486
487 // Resets all the marking data structures. Called when we have to restart
488 // marking or when marking completes (via set_non_marking_state below).
489 void reset_marking_state(bool clear_overflow = true);
490
491 // We do this after we're done with marking so that the marking data
492 // structures are initialized to a sensible and predictable state.
493 void set_non_marking_state();
494
495 // Called to indicate how many threads are currently active.
496 void set_concurrency(uint active_tasks);
497
498 // It should be called to indicate which phase we're in (concurrent
499 // mark or remark) and how many threads are currently active.
500 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
501
502 // Prints all gathered CM-related statistics
503 void print_stats();
504
505 bool cleanup_list_is_empty() {
506 return _cleanup_list.is_empty();
507 }
508
509 // Accessor methods
510 uint parallel_marking_threads() const { return _parallel_marking_threads; }
511 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
512 double sleep_factor() { return _sleep_factor; }
513 double marking_task_overhead() { return _marking_task_overhead;}
514 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
515 double cleanup_task_overhead() { return _cleanup_task_overhead;}
516
517 bool use_parallel_marking_threads() const {
518 assert(parallel_marking_threads() <=
519 max_parallel_marking_threads(), "sanity");
520 assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
521 parallel_marking_threads() > 0,
522 "parallel workers not set up correctly");
523 return _parallel_workers != NULL;
524 }
525
526 HeapWord* finger() { return _finger; }
527 bool concurrent() { return _concurrent; }
528 uint active_tasks() { return _active_tasks; }
529 ParallelTaskTerminator* terminator() { return &_terminator; }
530
531 // It claims the next available region to be scanned by a marking
532 // task/thread. It might return NULL if the next region is empty or
533 // we have run out of regions. In the latter case, out_of_regions()
534 // determines whether we've really run out of regions or the task
535 // should call claim_region() again. This might seem a bit
536 // awkward. Originally, the code was written so that claim_region()
537 // either successfully returned with a non-empty region or there
538 // were no more regions to be claimed. The problem with this was
539 // that, in certain circumstances, it iterated over large chunks of
540 // the heap finding only empty regions and, while it was working, it
541 // was preventing the calling task to call its regular clock
542 // method. So, this way, each task will spend very little time in
543 // claim_region() and is allowed to call the regular clock method
544 // frequently.
545 HeapRegion* claim_region(uint worker_id);
546
547 // It determines whether we've run out of regions to scan. Note that
548 // the finger can point past the heap end in case the heap was expanded
549 // to satisfy an allocation without doing a GC. This is fine, because all
550 // objects in those regions will be considered live anyway because of
551 // SATB guarantees (i.e. their TAMS will be equal to bottom).
552 bool out_of_regions() { return _finger >= _heap_end; }
553
554 // Returns the task with the given id
555 CMTask* task(int id) {
556 assert(0 <= id && id < (int) _active_tasks,
557 "task id not within active bounds");
558 return _tasks[id];
559 }
560
561 // Returns the task queue with the given id
562 CMTaskQueue* task_queue(int id) {
563 assert(0 <= id && id < (int) _active_tasks,
564 "task queue id not within active bounds");
565 return (CMTaskQueue*) _task_queues->queue(id);
566 }
567
568 // Returns the task queue set
569 CMTaskQueueSet* task_queues() { return _task_queues; }
570
571 // Access / manipulation of the overflow flag which is set to
572 // indicate that the global stack has overflown
573 bool has_overflown() { return _has_overflown; }
574 void set_has_overflown() { _has_overflown = true; }
575 void clear_has_overflown() { _has_overflown = false; }
576 bool restart_for_overflow() { return _restart_for_overflow; }
577
578 // Methods to enter the two overflow sync barriers
579 void enter_first_sync_barrier(uint worker_id);
580 void enter_second_sync_barrier(uint worker_id);
581
582 ForceOverflowSettings* force_overflow_conc() {
583 return &_force_overflow_conc;
584 }
585
586 ForceOverflowSettings* force_overflow_stw() {
587 return &_force_overflow_stw;
588 }
589
590 ForceOverflowSettings* force_overflow() {
591 if (concurrent()) {
592 return force_overflow_conc();
593 } else {
594 return force_overflow_stw();
595 }
596 }
597
598 // Live Data Counting data structures...
599 // These data structures are initialized at the start of
600 // marking. They are written to while marking is active.
601 // They are aggregated during remark; the aggregated values
602 // are then used to populate the _region_bm, _card_bm, and
603 // the total live bytes, which are then subsequently updated
604 // during cleanup.
605
606 // An array of bitmaps (one bit map per task). Each bitmap
607 // is used to record the cards spanned by the live objects
608 // marked by that task/worker.
609 BitMap* _count_card_bitmaps;
610
611 // Used to record the number of marked live bytes
612 // (for each region, by worker thread).
613 size_t** _count_marked_bytes;
614
615 // Card index of the bottom of the G1 heap. Used for biasing indices into
616 // the card bitmaps.
617 intptr_t _heap_bottom_card_num;
618
619 // Set to true when initialization is complete
620 bool _completed_initialization;
621
622 public:
623 // Manipulation of the global mark stack.
624 // Notice that the first mark_stack_push is CAS-based, whereas the
625 // two below are Mutex-based. This is OK since the first one is only
626 // called during evacuation pauses and doesn't compete with the
627 // other two (which are called by the marking tasks during
628 // concurrent marking or remark).
629 bool mark_stack_push(oop p) {
630 _markStack.par_push(p);
631 if (_markStack.overflow()) {
632 set_has_overflown();
633 return false;
634 }
635 return true;
636 }
637 bool mark_stack_push(oop* arr, int n) {
638 _markStack.par_push_arr(arr, n);
639 if (_markStack.overflow()) {
640 set_has_overflown();
641 return false;
642 }
643 return true;
644 }
645 void mark_stack_pop(oop* arr, int max, int* n) {
646 _markStack.par_pop_arr(arr, max, n);
647 }
648 size_t mark_stack_size() { return _markStack.size(); }
649 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
650 bool mark_stack_overflow() { return _markStack.overflow(); }
651 bool mark_stack_empty() { return _markStack.isEmpty(); }
652
653 CMRootRegions* root_regions() { return &_root_regions; }
654
655 bool concurrent_marking_in_progress() {
656 return _concurrent_marking_in_progress;
657 }
658 void set_concurrent_marking_in_progress() {
659 _concurrent_marking_in_progress = true;
660 }
661 void clear_concurrent_marking_in_progress() {
662 _concurrent_marking_in_progress = false;
663 }
664
665 void update_accum_task_vtime(int i, double vtime) {
666 _accum_task_vtime[i] += vtime;
667 }
668
669 double all_task_accum_vtime() {
670 double ret = 0.0;
671 for (uint i = 0; i < _max_worker_id; ++i)
672 ret += _accum_task_vtime[i];
673 return ret;
674 }
675
676 // Attempts to steal an object from the task queues of other tasks
677 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
678 return _task_queues->steal(worker_id, hash_seed, obj);
679 }
680
681 ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
682 ~ConcurrentMark();
683
684 ConcurrentMarkThread* cmThread() { return _cmThread; }
685
686 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
687 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
688
689 // Returns the number of GC threads to be used in a concurrent
690 // phase based on the number of GC threads being used in a STW
691 // phase.
692 uint scale_parallel_threads(uint n_par_threads);
693
694 // Calculates the number of GC threads to be used in a concurrent phase.
695 uint calc_parallel_marking_threads();
696
697 // The following three are interaction between CM and
698 // G1CollectedHeap
699
700 // This notifies CM that a root during initial-mark needs to be
701 // grayed. It is MT-safe. word_size is the size of the object in
702 // words. It is passed explicitly as sometimes we cannot calculate
703 // it from the given object because it might be in an inconsistent
704 // state (e.g., in to-space and being copied). So the caller is
705 // responsible for dealing with this issue (e.g., get the size from
706 // the from-space image when the to-space image might be
707 // inconsistent) and always passing the size. hr is the region that
708 // contains the object and it's passed optionally from callers who
709 // might already have it (no point in recalculating it).
710 inline void grayRoot(oop obj, size_t word_size,
711 uint worker_id, HeapRegion* hr = NULL);
712
713 // It iterates over the heap and for each object it comes across it
714 // will dump the contents of its reference fields, as well as
715 // liveness information for the object and its referents. The dump
716 // will be written to a file with the following name:
717 // G1PrintReachableBaseFile + "." + str.
718 // vo decides whether the prev (vo == UsePrevMarking), the next
719 // (vo == UseNextMarking) marking information, or the mark word
720 // (vo == UseMarkWord) will be used to determine the liveness of
721 // each object / referent.
722 // If all is true, all objects in the heap will be dumped, otherwise
723 // only the live ones. In the dump the following symbols / breviations
724 // are used:
725 // M : an explicitly live object (its bitmap bit is set)
726 // > : an implicitly live object (over tams)
727 // O : an object outside the G1 heap (typically: in the perm gen)
728 // NOT : a reference field whose referent is not live
729 // AND MARKED : indicates that an object is both explicitly and
730 // implicitly live (it should be one or the other, not both)
731 void print_reachable(const char* str,
732 VerifyOption vo, bool all) PRODUCT_RETURN;
733
734 // Clear the next marking bitmap (will be called concurrently).
735 void clearNextBitmap();
736
737 // These two do the work that needs to be done before and after the
738 // initial root checkpoint. Since this checkpoint can be done at two
739 // different points (i.e. an explicit pause or piggy-backed on a
740 // young collection), then it's nice to be able to easily share the
741 // pre/post code. It might be the case that we can put everything in
742 // the post method. TP
743 void checkpointRootsInitialPre();
744 void checkpointRootsInitialPost();
745
746 // Scan all the root regions and mark everything reachable from
747 // them.
748 void scanRootRegions();
749
750 // Scan a single root region and mark everything reachable from it.
751 void scanRootRegion(HeapRegion* hr, uint worker_id);
752
753 // Do concurrent phase of marking, to a tentative transitive closure.
754 void markFromRoots();
755
756 void checkpointRootsFinal(bool clear_all_soft_refs);
757 void checkpointRootsFinalWork();
758 void cleanup();
759 void completeCleanup();
760
761 // Mark in the previous bitmap. NB: this is usually read-only, so use
762 // this carefully!
763 inline void markPrev(oop p);
764
765 // Clears marks for all objects in the given range, for the prev,
766 // next, or both bitmaps. NB: the previous bitmap is usually
767 // read-only, so use this carefully!
768 void clearRangePrevBitmap(MemRegion mr);
769 void clearRangeNextBitmap(MemRegion mr);
770 void clearRangeBothBitmaps(MemRegion mr);
771
772 // Notify data structures that a GC has started.
773 void note_start_of_gc() {
774 _markStack.note_start_of_gc();
775 }
776
777 // Notify data structures that a GC is finished.
778 void note_end_of_gc() {
779 _markStack.note_end_of_gc();
780 }
781
782 // Verify that there are no CSet oops on the stacks (taskqueues /
783 // global mark stack), enqueued SATB buffers, per-thread SATB
784 // buffers, and fingers (global / per-task). The boolean parameters
785 // decide which of the above data structures to verify. If marking
786 // is not in progress, it's a no-op.
787 void verify_no_cset_oops(bool verify_stacks,
788 bool verify_enqueued_buffers,
789 bool verify_thread_buffers,
790 bool verify_fingers) PRODUCT_RETURN;
791
792 // It is called at the end of an evacuation pause during marking so
793 // that CM is notified of where the new end of the heap is. It
794 // doesn't do anything if concurrent_marking_in_progress() is false,
795 // unless the force parameter is true.
796 void update_g1_committed(bool force = false);
797
798 bool isMarked(oop p) const {
799 assert(p != NULL && p->is_oop(), "expected an oop");
800 HeapWord* addr = (HeapWord*)p;
801 assert(addr >= _nextMarkBitMap->startWord() ||
802 addr < _nextMarkBitMap->endWord(), "in a region");
803
804 return _nextMarkBitMap->isMarked(addr);
805 }
806
807 inline bool not_yet_marked(oop p) const;
808
809 // XXX Debug code
810 bool containing_card_is_marked(void* p);
811 bool containing_cards_are_marked(void* start, void* last);
812
813 bool isPrevMarked(oop p) const {
814 assert(p != NULL && p->is_oop(), "expected an oop");
815 HeapWord* addr = (HeapWord*)p;
816 assert(addr >= _prevMarkBitMap->startWord() ||
817 addr < _prevMarkBitMap->endWord(), "in a region");
818
819 return _prevMarkBitMap->isMarked(addr);
820 }
821
822 inline bool do_yield_check(uint worker_i = 0);
823
824 // Called to abort the marking cycle after a Full GC takes place.
825 void abort();
826
827 bool has_aborted() { return _has_aborted; }
828
829 const GCId& concurrent_gc_id();
830
831 // This prints the global/local fingers. It is used for debugging.
832 NOT_PRODUCT(void print_finger();)
833
834 void print_summary_info();
835
836 void print_worker_threads_on(outputStream* st) const;
837
838 void print_on_error(outputStream* st) const;
839
840 // The following indicate whether a given verbose level has been
841 // set. Notice that anything above stats is conditional to
842 // _MARKING_VERBOSE_ having been set to 1
843 bool verbose_stats() {
844 return _verbose_level >= stats_verbose;
845 }
846 bool verbose_low() {
847 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
848 }
849 bool verbose_medium() {
850 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
851 }
852 bool verbose_high() {
853 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
854 }
855
856 // Liveness counting
857
858 // Utility routine to set an exclusive range of cards on the given
859 // card liveness bitmap
860 inline void set_card_bitmap_range(BitMap* card_bm,
861 BitMap::idx_t start_idx,
862 BitMap::idx_t end_idx,
863 bool is_par);
864
865 // Returns the card number of the bottom of the G1 heap.
866 // Used in biasing indices into accounting card bitmaps.
867 intptr_t heap_bottom_card_num() const {
868 return _heap_bottom_card_num;
869 }
870
871 // Returns the card bitmap for a given task or worker id.
872 BitMap* count_card_bitmap_for(uint worker_id) {
873 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
874 assert(_count_card_bitmaps != NULL, "uninitialized");
875 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
876 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
877 return task_card_bm;
878 }
879
880 // Returns the array containing the marked bytes for each region,
881 // for the given worker or task id.
882 size_t* count_marked_bytes_array_for(uint worker_id) {
883 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
884 assert(_count_marked_bytes != NULL, "uninitialized");
885 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
886 assert(marked_bytes_array != NULL, "uninitialized");
887 return marked_bytes_array;
888 }
889
890 // Returns the index in the liveness accounting card table bitmap
891 // for the given address
892 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
893
894 // Counts the size of the given memory region in the the given
895 // marked_bytes array slot for the given HeapRegion.
896 // Sets the bits in the given card bitmap that are associated with the
897 // cards that are spanned by the memory region.
898 inline void count_region(MemRegion mr, HeapRegion* hr,
899 size_t* marked_bytes_array,
900 BitMap* task_card_bm);
901
902 // Counts the given memory region in the task/worker counting
903 // data structures for the given worker id.
904 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
905
906 // Counts the given memory region in the task/worker counting
907 // data structures for the given worker id.
908 inline void count_region(MemRegion mr, uint worker_id);
909
910 // Counts the given object in the given task/worker counting
911 // data structures.
912 inline void count_object(oop obj, HeapRegion* hr,
913 size_t* marked_bytes_array,
914 BitMap* task_card_bm);
915
916 // Counts the given object in the task/worker counting data
917 // structures for the given worker id.
918 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
919
920 // Attempts to mark the given object and, if successful, counts
921 // the object in the given task/worker counting structures.
922 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
923 size_t* marked_bytes_array,
924 BitMap* task_card_bm);
925
926 // Attempts to mark the given object and, if successful, counts
927 // the object in the task/worker counting structures for the
928 // given worker id.
929 inline bool par_mark_and_count(oop obj, size_t word_size,
930 HeapRegion* hr, uint worker_id);
931
932 // Attempts to mark the given object and, if successful, counts
933 // the object in the task/worker counting structures for the
934 // given worker id.
935 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
936
937 // Similar to the above routine but we don't know the heap region that
938 // contains the object to be marked/counted, which this routine looks up.
939 inline bool par_mark_and_count(oop obj, uint worker_id);
940
941 // Similar to the above routine but there are times when we cannot
942 // safely calculate the size of obj due to races and we, therefore,
943 // pass the size in as a parameter. It is the caller's responsibility
944 // to ensure that the size passed in for obj is valid.
945 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
946
947 // Unconditionally mark the given object, and unconditionally count
948 // the object in the counting structures for worker id 0.
949 // Should *not* be called from parallel code.
950 inline bool mark_and_count(oop obj, HeapRegion* hr);
951
952 // Similar to the above routine but we don't know the heap region that
953 // contains the object to be marked/counted, which this routine looks up.
954 // Should *not* be called from parallel code.
955 inline bool mark_and_count(oop obj);
956
957 // Returns true if initialization was successfully completed.
958 bool completed_initialization() const {
959 return _completed_initialization;
960 }
961
962 protected:
963 // Clear all the per-task bitmaps and arrays used to store the
964 // counting data.
965 void clear_all_count_data();
966
967 // Aggregates the counting data for each worker/task
968 // that was constructed while marking. Also sets
969 // the amount of marked bytes for each region and
970 // the top at concurrent mark count.
971 void aggregate_count_data();
972
973 // Verification routine
974 void verify_count_data();
975 };
976
977 // A class representing a marking task.
978 class CMTask : public TerminatorTerminator {
979 private:
980 enum PrivateConstants {
981 // the regular clock call is called once the scanned words reaches
982 // this limit
983 words_scanned_period = 12*1024,
984 // the regular clock call is called once the number of visited
985 // references reaches this limit
986 refs_reached_period = 384,
987 // initial value for the hash seed, used in the work stealing code
988 init_hash_seed = 17,
989 // how many entries will be transferred between global stack and
990 // local queues
991 global_stack_transfer_size = 16
992 };
993
994 uint _worker_id;
995 G1CollectedHeap* _g1h;
996 ConcurrentMark* _cm;
997 CMBitMap* _nextMarkBitMap;
998 // the task queue of this task
999 CMTaskQueue* _task_queue;
1000 private:
1001 // the task queue set---needed for stealing
1002 CMTaskQueueSet* _task_queues;
1003 // indicates whether the task has been claimed---this is only for
1004 // debugging purposes
1005 bool _claimed;
1006
1007 // number of calls to this task
1008 int _calls;
1009
1010 // when the virtual timer reaches this time, the marking step should
1011 // exit
1012 double _time_target_ms;
1013 // the start time of the current marking step
1014 double _start_time_ms;
1015
1016 // the oop closure used for iterations over oops
1017 G1CMOopClosure* _cm_oop_closure;
1018
1019 // the region this task is scanning, NULL if we're not scanning any
1020 HeapRegion* _curr_region;
1021 // the local finger of this task, NULL if we're not scanning a region
1022 HeapWord* _finger;
1023 // limit of the region this task is scanning, NULL if we're not scanning one
1024 HeapWord* _region_limit;
1025
1026 // the number of words this task has scanned
1027 size_t _words_scanned;
1028 // When _words_scanned reaches this limit, the regular clock is
1029 // called. Notice that this might be decreased under certain
1030 // circumstances (i.e. when we believe that we did an expensive
1031 // operation).
1032 size_t _words_scanned_limit;
1033 // the initial value of _words_scanned_limit (i.e. what it was
1034 // before it was decreased).
1035 size_t _real_words_scanned_limit;
1036
1037 // the number of references this task has visited
1038 size_t _refs_reached;
1039 // When _refs_reached reaches this limit, the regular clock is
1040 // called. Notice this this might be decreased under certain
1041 // circumstances (i.e. when we believe that we did an expensive
1042 // operation).
1043 size_t _refs_reached_limit;
1044 // the initial value of _refs_reached_limit (i.e. what it was before
1045 // it was decreased).
1046 size_t _real_refs_reached_limit;
1047
1048 // used by the work stealing stuff
1049 int _hash_seed;
1050 // if this is true, then the task has aborted for some reason
1051 bool _has_aborted;
1052 // set when the task aborts because it has met its time quota
1053 bool _has_timed_out;
1054 // true when we're draining SATB buffers; this avoids the task
1055 // aborting due to SATB buffers being available (as we're already
1056 // dealing with them)
1057 bool _draining_satb_buffers;
1058
1059 // number sequence of past step times
1060 NumberSeq _step_times_ms;
1061 // elapsed time of this task
1062 double _elapsed_time_ms;
1063 // termination time of this task
1064 double _termination_time_ms;
1065 // when this task got into the termination protocol
1066 double _termination_start_time_ms;
1067
1068 // true when the task is during a concurrent phase, false when it is
1069 // in the remark phase (so, in the latter case, we do not have to
1070 // check all the things that we have to check during the concurrent
1071 // phase, i.e. SATB buffer availability...)
1072 bool _concurrent;
1073
1074 TruncatedSeq _marking_step_diffs_ms;
1075
1076 // Counting data structures. Embedding the task's marked_bytes_array
1077 // and card bitmap into the actual task saves having to go through
1078 // the ConcurrentMark object.
1079 size_t* _marked_bytes_array;
1080 BitMap* _card_bm;
1081
1082 // LOTS of statistics related with this task
1083 #if _MARKING_STATS_
1084 NumberSeq _all_clock_intervals_ms;
1085 double _interval_start_time_ms;
1086
1087 int _aborted;
1088 int _aborted_overflow;
1089 int _aborted_cm_aborted;
1090 int _aborted_yield;
1091 int _aborted_timed_out;
1092 int _aborted_satb;
1093 int _aborted_termination;
1094
1095 int _steal_attempts;
1096 int _steals;
1097
1098 int _clock_due_to_marking;
1099 int _clock_due_to_scanning;
1100
1101 int _local_pushes;
1102 int _local_pops;
1103 int _local_max_size;
1104 int _objs_scanned;
1105
1106 int _global_pushes;
1107 int _global_pops;
1108 int _global_max_size;
1109
1110 int _global_transfers_to;
1111 int _global_transfers_from;
1112
1113 int _regions_claimed;
1114 int _objs_found_on_bitmap;
1115
1116 int _satb_buffers_processed;
1117 #endif // _MARKING_STATS_
1118
1119 // it updates the local fields after this task has claimed
1120 // a new region to scan
1121 void setup_for_region(HeapRegion* hr);
1122 // it brings up-to-date the limit of the region
1123 void update_region_limit();
1124
1125 // called when either the words scanned or the refs visited limit
1126 // has been reached
1127 void reached_limit();
1128 // recalculates the words scanned and refs visited limits
1129 void recalculate_limits();
1130 // decreases the words scanned and refs visited limits when we reach
1131 // an expensive operation
1132 void decrease_limits();
1133 // it checks whether the words scanned or refs visited reached their
1134 // respective limit and calls reached_limit() if they have
1135 void check_limits() {
1136 if (_words_scanned >= _words_scanned_limit ||
1137 _refs_reached >= _refs_reached_limit) {
1138 reached_limit();
1139 }
1140 }
1141 // this is supposed to be called regularly during a marking step as
1142 // it checks a bunch of conditions that might cause the marking step
1143 // to abort
1144 void regular_clock_call();
1145 bool concurrent() { return _concurrent; }
1146
1147 public:
1148 // It resets the task; it should be called right at the beginning of
1149 // a marking phase.
1150 void reset(CMBitMap* _nextMarkBitMap);
1151 // it clears all the fields that correspond to a claimed region.
1152 void clear_region_fields();
1153
1154 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1155
1156 // The main method of this class which performs a marking step
1157 // trying not to exceed the given duration. However, it might exit
1158 // prematurely, according to some conditions (i.e. SATB buffers are
1159 // available for processing).
1160 void do_marking_step(double target_ms,
1161 bool do_termination,
1162 bool is_serial);
1163
1164 // These two calls start and stop the timer
1165 void record_start_time() {
1166 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1167 }
1168 void record_end_time() {
1169 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1170 }
1171
1172 // returns the worker ID associated with this task.
1173 uint worker_id() { return _worker_id; }
1174
1175 // From TerminatorTerminator. It determines whether this task should
1176 // exit the termination protocol after it's entered it.
1177 virtual bool should_exit_termination();
1178
1179 // Resets the local region fields after a task has finished scanning a
1180 // region; or when they have become stale as a result of the region
1181 // being evacuated.
1182 void giveup_current_region();
1183
1184 HeapWord* finger() { return _finger; }
1185
1186 bool has_aborted() { return _has_aborted; }
1187 void set_has_aborted() { _has_aborted = true; }
1188 void clear_has_aborted() { _has_aborted = false; }
1189 bool has_timed_out() { return _has_timed_out; }
1190 bool claimed() { return _claimed; }
1191
1192 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1193
1194 // It grays the object by marking it and, if necessary, pushing it
1195 // on the local queue
1196 inline void deal_with_reference(oop obj);
1197
1198 // It scans an object and visits its children.
1199 void scan_object(oop obj);
1200
1201 // It pushes an object on the local queue.
1202 inline void push(oop obj);
1203
1204 // These two move entries to/from the global stack.
1205 void move_entries_to_global_stack();
1206 void get_entries_from_global_stack();
1207
1208 // It pops and scans objects from the local queue. If partially is
1209 // true, then it stops when the queue size is of a given limit. If
1210 // partially is false, then it stops when the queue is empty.
1211 void drain_local_queue(bool partially);
1212 // It moves entries from the global stack to the local queue and
1213 // drains the local queue. If partially is true, then it stops when
1214 // both the global stack and the local queue reach a given size. If
1215 // partially if false, it tries to empty them totally.
1216 void drain_global_stack(bool partially);
1217 // It keeps picking SATB buffers and processing them until no SATB
1218 // buffers are available.
1219 void drain_satb_buffers();
1220
1221 // moves the local finger to a new location
1222 inline void move_finger_to(HeapWord* new_finger) {
1223 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1224 _finger = new_finger;
1225 }
1226
1227 CMTask(uint worker_id, ConcurrentMark *cm,
1228 size_t* marked_bytes, BitMap* card_bm,
1229 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1230
1231 // it prints statistics associated with this task
1232 void print_stats();
1233
1234 #if _MARKING_STATS_
1235 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1236 #endif // _MARKING_STATS_
1237 };
1238
1239 // Class that's used to to print out per-region liveness
1240 // information. It's currently used at the end of marking and also
1241 // after we sort the old regions at the end of the cleanup operation.
1242 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1243 private:
1244 outputStream* _out;
1245
1246 // Accumulators for these values.
1247 size_t _total_used_bytes;
1248 size_t _total_capacity_bytes;
1249 size_t _total_prev_live_bytes;
1250 size_t _total_next_live_bytes;
1251
1252 // These are set up when we come across a "stars humongous" region
1253 // (as this is where most of this information is stored, not in the
1254 // subsequent "continues humongous" regions). After that, for every
1255 // region in a given humongous region series we deduce the right
1256 // values for it by simply subtracting the appropriate amount from
1257 // these fields. All these values should reach 0 after we've visited
1258 // the last region in the series.
1259 size_t _hum_used_bytes;
1260 size_t _hum_capacity_bytes;
1261 size_t _hum_prev_live_bytes;
1262 size_t _hum_next_live_bytes;
1263
1264 // Accumulator for the remembered set size
1265 size_t _total_remset_bytes;
1266
1267 // Accumulator for strong code roots memory size
1268 size_t _total_strong_code_roots_bytes;
1269
1270 static double perc(size_t val, size_t total) {
1271 if (total == 0) {
1272 return 0.0;
1273 } else {
1274 return 100.0 * ((double) val / (double) total);
1275 }
1276 }
1277
1278 static double bytes_to_mb(size_t val) {
1279 return (double) val / (double) M;
1280 }
1281
1282 // See the .cpp file.
1283 size_t get_hum_bytes(size_t* hum_bytes);
1284 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1285 size_t* prev_live_bytes, size_t* next_live_bytes);
1286
1287 public:
1288 // The header and footer are printed in the constructor and
1289 // destructor respectively.
1290 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1291 virtual bool doHeapRegion(HeapRegion* r);
1292 ~G1PrintRegionLivenessInfoClosure();
1293 };
1294
1295 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP