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 "classfile/javaClasses.hpp"
29 #include "gc_implementation/g1/heapRegionSet.hpp"
30 #include "gc_implementation/shared/gcId.hpp"
31 #include "utilities/taskqueue.hpp"
32
33 class G1CollectedHeap;
34 class CMTask;
35 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
36 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
37
38 // Closure used by CM during concurrent reference discovery
39 // and reference processing (during remarking) to determine
40 // if a particular object is alive. It is primarily used
41 // to determine if referents of discovered reference objects
42 // are alive. An instance is also embedded into the
43 // reference processor as the _is_alive_non_header field
44 class G1CMIsAliveClosure: public BoolObjectClosure {
45 G1CollectedHeap* _g1;
46 public:
47 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
48
49 bool do_object_b(oop obj);
50 };
51
52 // A generic CM bit map. This is essentially a wrapper around the BitMap
53 // class, with one bit per (1<<_shifter) HeapWords.
54
55 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
56 protected:
57 HeapWord* _bmStartWord; // base address of range covered by map
58 size_t _bmWordSize; // map size (in #HeapWords covered)
59 const int _shifter; // map to char or bit
60 VirtualSpace _virtual_space; // underlying the bit map
61 BitMap _bm; // the bit map itself
62
63 public:
64 // constructor
65 CMBitMapRO(int shifter);
66
67 enum { do_yield = true };
68
69 // inquiries
70 HeapWord* startWord() const { return _bmStartWord; }
71 size_t sizeInWords() const { return _bmWordSize; }
72 // the following is one past the last word in space
73 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
74
75 // read marks
76
77 bool isMarked(HeapWord* addr) const {
78 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
79 "outside underlying space?");
80 return _bm.at(heapWordToOffset(addr));
81 }
82
83 // iteration
84 inline bool iterate(BitMapClosure* cl, MemRegion mr);
85 inline bool iterate(BitMapClosure* cl);
86
87 // Return the address corresponding to the next marked bit at or after
88 // "addr", and before "limit", if "limit" is non-NULL. If there is no
89 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
90 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
91 const HeapWord* limit = NULL) const;
92 // Return the address corresponding to the next unmarked bit at or after
93 // "addr", and before "limit", if "limit" is non-NULL. If there is no
94 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
95 HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr,
96 const HeapWord* limit = NULL) const;
97
98 // conversion utilities
99 HeapWord* offsetToHeapWord(size_t offset) const {
100 return _bmStartWord + (offset << _shifter);
101 }
102 size_t heapWordToOffset(const HeapWord* addr) const {
103 return pointer_delta(addr, _bmStartWord) >> _shifter;
104 }
105 int heapWordDiffToOffsetDiff(size_t diff) const;
106
107 // The argument addr should be the start address of a valid object
108 HeapWord* nextObject(HeapWord* addr) {
109 oop obj = (oop) addr;
110 HeapWord* res = addr + obj->size();
111 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
112 return res;
113 }
114
115 void print_on_error(outputStream* st, const char* prefix) const;
116
117 // debugging
118 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
119 };
120
121 class CMBitMap : public CMBitMapRO {
122
123 public:
124 // constructor
125 CMBitMap(int shifter) :
126 CMBitMapRO(shifter) {}
127
128 // Allocates the back store for the marking bitmap
129 bool allocate(ReservedSpace heap_rs);
130
131 // write marks
132 void mark(HeapWord* addr) {
133 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
134 "outside underlying space?");
135 _bm.set_bit(heapWordToOffset(addr));
136 }
137 void clear(HeapWord* addr) {
138 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
139 "outside underlying space?");
140 _bm.clear_bit(heapWordToOffset(addr));
141 }
142 bool parMark(HeapWord* addr) {
143 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
144 "outside underlying space?");
145 return _bm.par_set_bit(heapWordToOffset(addr));
146 }
147 bool parClear(HeapWord* addr) {
148 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
149 "outside underlying space?");
150 return _bm.par_clear_bit(heapWordToOffset(addr));
151 }
152 void markRange(MemRegion mr);
153 void clearAll();
154 void clearRange(MemRegion mr);
155
156 // Starting at the bit corresponding to "addr" (inclusive), find the next
157 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
158 // the end of this run (stopping at "end_addr"). Return the MemRegion
159 // covering from the start of the region corresponding to the first bit
160 // of the run to the end of the region corresponding to the last bit of
161 // the run. If there is no "1" bit at or after "addr", return an empty
162 // MemRegion.
163 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
164 };
165
166 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
167 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
168 VirtualSpace _virtual_space; // Underlying backing store for actual stack
169 ConcurrentMark* _cm;
170 oop* _base; // bottom of stack
171 jint _index; // one more than last occupied index
172 jint _capacity; // max #elements
173 jint _saved_index; // value of _index saved at start of GC
174 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
175
176 bool _overflow;
177 bool _should_expand;
178 DEBUG_ONLY(bool _drain_in_progress;)
179 DEBUG_ONLY(bool _drain_in_progress_yields;)
180
181 public:
182 CMMarkStack(ConcurrentMark* cm);
183 ~CMMarkStack();
184
185 #ifndef PRODUCT
186 jint max_depth() const {
187 return _max_depth;
188 }
189 #endif
190
191 bool allocate(size_t capacity);
192
193 oop pop() {
194 if (!isEmpty()) {
195 return _base[--_index] ;
196 }
197 return NULL;
198 }
199
200 // If overflow happens, don't do the push, and record the overflow.
201 // *Requires* that "ptr" is already marked.
202 void push(oop ptr) {
203 if (isFull()) {
204 // Record overflow.
205 _overflow = true;
206 return;
207 } else {
208 _base[_index++] = ptr;
209 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
210 }
211 }
212 // Non-block impl. Note: concurrency is allowed only with other
213 // "par_push" operations, not with "pop" or "drain". We would need
214 // parallel versions of them if such concurrency was desired.
215 void par_push(oop ptr);
216
217 // Pushes the first "n" elements of "ptr_arr" on the stack.
218 // Non-block impl. Note: concurrency is allowed only with other
219 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
220 void par_adjoin_arr(oop* ptr_arr, int n);
221
222 // Pushes the first "n" elements of "ptr_arr" on the stack.
223 // Locking impl: concurrency is allowed only with
224 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
225 // locking strategy.
226 void par_push_arr(oop* ptr_arr, int n);
227
228 // If returns false, the array was empty. Otherwise, removes up to "max"
229 // elements from the stack, and transfers them to "ptr_arr" in an
230 // unspecified order. The actual number transferred is given in "n" ("n
231 // == 0" is deliberately redundant with the return value.) Locking impl:
232 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
233 // operations, which use the same locking strategy.
234 bool par_pop_arr(oop* ptr_arr, int max, int* n);
235
236 // Drain the mark stack, applying the given closure to all fields of
237 // objects on the stack. (That is, continue until the stack is empty,
238 // even if closure applications add entries to the stack.) The "bm"
239 // argument, if non-null, may be used to verify that only marked objects
240 // are on the mark stack. If "yield_after" is "true", then the
241 // concurrent marker performing the drain offers to yield after
242 // processing each object. If a yield occurs, stops the drain operation
243 // and returns false. Otherwise, returns true.
244 template<class OopClosureClass>
245 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
246
247 bool isEmpty() { return _index == 0; }
248 bool isFull() { return _index == _capacity; }
249 int maxElems() { return _capacity; }
250
251 bool overflow() { return _overflow; }
252 void clear_overflow() { _overflow = false; }
253
254 bool should_expand() const { return _should_expand; }
255 void set_should_expand();
256
257 // Expand the stack, typically in response to an overflow condition
258 void expand();
259
260 int size() { return _index; }
261
262 void setEmpty() { _index = 0; clear_overflow(); }
263
264 // Record the current index.
265 void note_start_of_gc();
266
267 // Make sure that we have not added any entries to the stack during GC.
268 void note_end_of_gc();
269
270 // iterate over the oops in the mark stack, up to the bound recorded via
271 // the call above.
272 void oops_do(OopClosure* f);
273 };
274
275 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
276 private:
277 #ifndef PRODUCT
278 uintx _num_remaining;
279 bool _force;
280 #endif // !defined(PRODUCT)
281
282 public:
283 void init() PRODUCT_RETURN;
284 void update() PRODUCT_RETURN;
285 bool should_force() PRODUCT_RETURN_( return false; );
286 };
287
288 // this will enable a variety of different statistics per GC task
289 #define _MARKING_STATS_ 0
290 // this will enable the higher verbose levels
291 #define _MARKING_VERBOSE_ 0
292
293 #if _MARKING_STATS_
294 #define statsOnly(statement) \
295 do { \
296 statement ; \
297 } while (0)
298 #else // _MARKING_STATS_
299 #define statsOnly(statement) \
300 do { \
301 } while (0)
302 #endif // _MARKING_STATS_
303
304 typedef enum {
305 no_verbose = 0, // verbose turned off
306 stats_verbose, // only prints stats at the end of marking
307 low_verbose, // low verbose, mostly per region and per major event
308 medium_verbose, // a bit more detailed than low
309 high_verbose // per object verbose
310 } CMVerboseLevel;
311
312 class YoungList;
313
314 // Root Regions are regions that are not empty at the beginning of a
315 // marking cycle and which we might collect during an evacuation pause
316 // while the cycle is active. Given that, during evacuation pauses, we
317 // do not copy objects that are explicitly marked, what we have to do
318 // for the root regions is to scan them and mark all objects reachable
319 // from them. According to the SATB assumptions, we only need to visit
320 // each object once during marking. So, as long as we finish this scan
321 // before the next evacuation pause, we can copy the objects from the
322 // root regions without having to mark them or do anything else to them.
323 //
324 // Currently, we only support root region scanning once (at the start
325 // of the marking cycle) and the root regions are all the survivor
326 // regions populated during the initial-mark pause.
327 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
328 private:
329 YoungList* _young_list;
330 ConcurrentMark* _cm;
331
332 volatile bool _scan_in_progress;
333 volatile bool _should_abort;
334 HeapRegion* volatile _next_survivor;
335
336 public:
337 CMRootRegions();
338 // We actually do most of the initialization in this method.
339 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
340
341 // Reset the claiming / scanning of the root regions.
342 void prepare_for_scan();
343
344 // Forces get_next() to return NULL so that the iteration aborts early.
345 void abort() { _should_abort = true; }
346
347 // Return true if the CM thread are actively scanning root regions,
348 // false otherwise.
349 bool scan_in_progress() { return _scan_in_progress; }
350
351 // Claim the next root region to scan atomically, or return NULL if
352 // all have been claimed.
353 HeapRegion* claim_next();
354
355 // Flag that we're done with root region scanning and notify anyone
356 // who's waiting on it. If aborted is false, assume that all regions
357 // have been claimed.
358 void scan_finished();
359
360 // If CM threads are still scanning root regions, wait until they
361 // are done. Return true if we had to wait, false otherwise.
362 bool wait_until_scan_finished();
363 };
364
365 class ConcurrentMarkThread;
366
367 class ConcurrentMark: public CHeapObj<mtGC> {
368 friend class CMMarkStack;
369 friend class ConcurrentMarkThread;
370 friend class CMTask;
371 friend class CMBitMapClosure;
372 friend class CMGlobalObjectClosure;
373 friend class CMRemarkTask;
374 friend class CMConcurrentMarkingTask;
375 friend class G1ParNoteEndTask;
376 friend class CalcLiveObjectsClosure;
377 friend class G1CMRefProcTaskProxy;
378 friend class G1CMRefProcTaskExecutor;
379 friend class G1CMKeepAliveAndDrainClosure;
380 friend class G1CMDrainMarkingStackClosure;
381
382 protected:
383 ConcurrentMarkThread* _cmThread; // The thread doing the work
384 G1CollectedHeap* _g1h; // The heap
385 uint _parallel_marking_threads; // The number of marking
386 // threads we're using
387 uint _max_parallel_marking_threads; // Max number of marking
388 // threads we'll ever use
389 double _sleep_factor; // How much we have to sleep, with
390 // respect to the work we just did, to
391 // meet the marking overhead goal
392 double _marking_task_overhead; // Marking target overhead for
393 // a single task
394
395 // Same as the two above, but for the cleanup task
396 double _cleanup_sleep_factor;
397 double _cleanup_task_overhead;
398
399 FreeRegionList _cleanup_list;
400
401 // Concurrent marking support structures
402 CMBitMap _markBitMap1;
403 CMBitMap _markBitMap2;
404 CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap
405 CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap
406
407 BitMap _region_bm;
408 BitMap _card_bm;
409
410 // Heap bounds
411 HeapWord* _heap_start;
412 HeapWord* _heap_end;
413
414 // Root region tracking and claiming
415 CMRootRegions _root_regions;
416
417 // For gray objects
418 CMMarkStack _markStack; // Grey objects behind global finger
419 HeapWord* volatile _finger; // The global finger, region aligned,
420 // always points to the end of the
421 // last claimed region
422
423 // Marking tasks
424 uint _max_worker_id;// Maximum worker id
425 uint _active_tasks; // Task num currently active
426 CMTask** _tasks; // Task queue array (max_worker_id len)
427 CMTaskQueueSet* _task_queues; // Task queue set
428 ParallelTaskTerminator _terminator; // For termination
429
430 // Two sync barriers that are used to synchronize tasks when an
431 // overflow occurs. The algorithm is the following. All tasks enter
432 // the first one to ensure that they have all stopped manipulating
433 // the global data structures. After they exit it, they re-initialize
434 // their data structures and task 0 re-initializes the global data
435 // structures. Then, they enter the second sync barrier. This
436 // ensure, that no task starts doing work before all data
437 // structures (local and global) have been re-initialized. When they
438 // exit it, they are free to start working again.
439 WorkGangBarrierSync _first_overflow_barrier_sync;
440 WorkGangBarrierSync _second_overflow_barrier_sync;
441
442 // This is set by any task, when an overflow on the global data
443 // structures is detected
444 volatile bool _has_overflown;
445 // True: marking is concurrent, false: we're in remark
446 volatile bool _concurrent;
447 // Set at the end of a Full GC so that marking aborts
448 volatile bool _has_aborted;
449 GCId _aborted_gc_id;
450
451 // Used when remark aborts due to an overflow to indicate that
452 // another concurrent marking phase should start
453 volatile bool _restart_for_overflow;
454
455 // This is true from the very start of concurrent marking until the
456 // point when all the tasks complete their work. It is really used
457 // to determine the points between the end of concurrent marking and
458 // time of remark.
459 volatile bool _concurrent_marking_in_progress;
460
461 // Verbose level
462 CMVerboseLevel _verbose_level;
463
464 // All of these times are in ms
465 NumberSeq _init_times;
466 NumberSeq _remark_times;
467 NumberSeq _remark_mark_times;
468 NumberSeq _remark_weak_ref_times;
469 NumberSeq _cleanup_times;
470 double _total_counting_time;
471 double _total_rs_scrub_time;
472
473 double* _accum_task_vtime; // Accumulated task vtime
474
475 FlexibleWorkGang* _parallel_workers;
476
477 ForceOverflowSettings _force_overflow_conc;
478 ForceOverflowSettings _force_overflow_stw;
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(bool clear_overflow = true);
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 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
517 double cleanup_task_overhead() { return _cleanup_task_overhead;}
518
519 bool use_parallel_marking_threads() const {
520 assert(parallel_marking_threads() <=
521 max_parallel_marking_threads(), "sanity");
522 assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
523 parallel_marking_threads() > 0,
524 "parallel workers not set up correctly");
525 return _parallel_workers != NULL;
526 }
527
528 HeapWord* finger() { return _finger; }
529 bool concurrent() { return _concurrent; }
530 uint active_tasks() { return _active_tasks; }
531 ParallelTaskTerminator* terminator() { return &_terminator; }
532
533 // It claims the next available region to be scanned by a marking
534 // task/thread. It might return NULL if the next region is empty or
535 // we have run out of regions. In the latter case, out_of_regions()
536 // determines whether we've really run out of regions or the task
537 // should call claim_region() again. This might seem a bit
538 // awkward. Originally, the code was written so that claim_region()
539 // either successfully returned with a non-empty region or there
540 // were no more regions to be claimed. The problem with this was
541 // that, in certain circumstances, it iterated over large chunks of
542 // the heap finding only empty regions and, while it was working, it
543 // was preventing the calling task to call its regular clock
544 // method. So, this way, each task will spend very little time in
545 // claim_region() and is allowed to call the regular clock method
546 // frequently.
547 HeapRegion* claim_region(uint worker_id);
548
549 // It determines whether we've run out of regions to scan. Note that
550 // the finger can point past the heap end in case the heap was expanded
551 // to satisfy an allocation without doing a GC. This is fine, because all
552 // objects in those regions will be considered live anyway because of
553 // SATB guarantees (i.e. their TAMS will be equal to bottom).
554 bool out_of_regions() { return _finger >= _heap_end; }
555
556 // Returns the task with the given id
557 CMTask* task(int id) {
558 assert(0 <= id && id < (int) _active_tasks,
559 "task id not within active bounds");
560 return _tasks[id];
561 }
562
563 // Returns the task queue with the given id
564 CMTaskQueue* task_queue(int id) {
565 assert(0 <= id && id < (int) _active_tasks,
566 "task queue id not within active bounds");
567 return (CMTaskQueue*) _task_queues->queue(id);
568 }
569
570 // Returns the task queue set
571 CMTaskQueueSet* task_queues() { return _task_queues; }
572
573 // Access / manipulation of the overflow flag which is set to
574 // indicate that the global stack has overflown
575 bool has_overflown() { return _has_overflown; }
576 void set_has_overflown() { _has_overflown = true; }
577 void clear_has_overflown() { _has_overflown = false; }
578 bool restart_for_overflow() { return _restart_for_overflow; }
579
580 // Methods to enter the two overflow sync barriers
581 void enter_first_sync_barrier(uint worker_id);
582 void enter_second_sync_barrier(uint worker_id);
583
584 ForceOverflowSettings* force_overflow_conc() {
585 return &_force_overflow_conc;
586 }
587
588 ForceOverflowSettings* force_overflow_stw() {
589 return &_force_overflow_stw;
590 }
591
592 ForceOverflowSettings* force_overflow() {
593 if (concurrent()) {
594 return force_overflow_conc();
595 } else {
596 return force_overflow_stw();
597 }
598 }
599
600 // Live Data Counting data structures...
601 // These data structures are initialized at the start of
602 // marking. They are written to while marking is active.
603 // They are aggregated during remark; the aggregated values
604 // are then used to populate the _region_bm, _card_bm, and
605 // the total live bytes, which are then subsequently updated
606 // during cleanup.
607
608 // An array of bitmaps (one bit map per task). Each bitmap
609 // is used to record the cards spanned by the live objects
610 // marked by that task/worker.
611 BitMap* _count_card_bitmaps;
612
613 // Used to record the number of marked live bytes
614 // (for each region, by worker thread).
615 size_t** _count_marked_bytes;
616
617 // Card index of the bottom of the G1 heap. Used for biasing indices into
618 // the card bitmaps.
619 intptr_t _heap_bottom_card_num;
620
621 // Set to true when initialization is complete
622 bool _completed_initialization;
623
624 public:
625 // Manipulation of the global mark stack.
626 // Notice that the first mark_stack_push is CAS-based, whereas the
627 // two below are Mutex-based. This is OK since the first one is only
628 // called during evacuation pauses and doesn't compete with the
629 // other two (which are called by the marking tasks during
630 // concurrent marking or remark).
631 bool mark_stack_push(oop p) {
632 _markStack.par_push(p);
633 if (_markStack.overflow()) {
634 set_has_overflown();
635 return false;
636 }
637 return true;
638 }
639 bool mark_stack_push(oop* arr, int n) {
640 _markStack.par_push_arr(arr, n);
641 if (_markStack.overflow()) {
642 set_has_overflown();
643 return false;
644 }
645 return true;
646 }
647 void mark_stack_pop(oop* arr, int max, int* n) {
648 _markStack.par_pop_arr(arr, max, n);
649 }
650 size_t mark_stack_size() { return _markStack.size(); }
651 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
652 bool mark_stack_overflow() { return _markStack.overflow(); }
653 bool mark_stack_empty() { return _markStack.isEmpty(); }
654
655 CMRootRegions* root_regions() { return &_root_regions; }
656
657 bool concurrent_marking_in_progress() {
658 return _concurrent_marking_in_progress;
659 }
660 void set_concurrent_marking_in_progress() {
661 _concurrent_marking_in_progress = true;
662 }
663 void clear_concurrent_marking_in_progress() {
664 _concurrent_marking_in_progress = false;
665 }
666
667 void update_accum_task_vtime(int i, double vtime) {
668 _accum_task_vtime[i] += vtime;
669 }
670
671 double all_task_accum_vtime() {
672 double ret = 0.0;
673 for (uint i = 0; i < _max_worker_id; ++i)
674 ret += _accum_task_vtime[i];
675 return ret;
676 }
677
678 // Attempts to steal an object from the task queues of other tasks
679 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
680 return _task_queues->steal(worker_id, hash_seed, obj);
681 }
682
683 ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
684 ~ConcurrentMark();
685
686 ConcurrentMarkThread* cmThread() { return _cmThread; }
687
688 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
689 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
690
691 // Returns the number of GC threads to be used in a concurrent
692 // phase based on the number of GC threads being used in a STW
693 // phase.
694 uint scale_parallel_threads(uint n_par_threads);
695
696 // Calculates the number of GC threads to be used in a concurrent phase.
697 uint calc_parallel_marking_threads();
698
699 // The following three are interaction between CM and
700 // G1CollectedHeap
701
702 // This notifies CM that a root during initial-mark needs to be
703 // grayed. It is MT-safe. word_size is the size of the object in
704 // words. It is passed explicitly as sometimes we cannot calculate
705 // it from the given object because it might be in an inconsistent
706 // state (e.g., in to-space and being copied). So the caller is
707 // responsible for dealing with this issue (e.g., get the size from
708 // the from-space image when the to-space image might be
709 // inconsistent) and always passing the size. hr is the region that
710 // contains the object and it's passed optionally from callers who
711 // might already have it (no point in recalculating it).
712 inline void grayRoot(oop obj, size_t word_size,
713 uint worker_id, HeapRegion* hr = NULL);
714
715 // It iterates over the heap and for each object it comes across it
716 // will dump the contents of its reference fields, as well as
717 // liveness information for the object and its referents. The dump
718 // will be written to a file with the following name:
719 // G1PrintReachableBaseFile + "." + str.
720 // vo decides whether the prev (vo == UsePrevMarking), the next
721 // (vo == UseNextMarking) marking information, or the mark word
722 // (vo == UseMarkWord) will be used to determine the liveness of
723 // each object / referent.
724 // If all is true, all objects in the heap will be dumped, otherwise
725 // only the live ones. In the dump the following symbols / breviations
726 // are used:
727 // M : an explicitly live object (its bitmap bit is set)
728 // > : an implicitly live object (over tams)
729 // O : an object outside the G1 heap (typically: in the perm gen)
730 // NOT : a reference field whose referent is not live
731 // AND MARKED : indicates that an object is both explicitly and
732 // implicitly live (it should be one or the other, not both)
733 void print_reachable(const char* str,
734 VerifyOption vo, bool all) PRODUCT_RETURN;
735
736 // Clear the next marking bitmap (will be called concurrently).
737 void clearNextBitmap();
738
739 // These two do the work that needs to be done before and after the
740 // initial root checkpoint. Since this checkpoint can be done at two
741 // different points (i.e. an explicit pause or piggy-backed on a
742 // young collection), then it's nice to be able to easily share the
743 // pre/post code. It might be the case that we can put everything in
744 // the post method. TP
745 void checkpointRootsInitialPre();
746 void checkpointRootsInitialPost();
747
748 // Scan all the root regions and mark everything reachable from
749 // them.
750 void scanRootRegions();
751
752 // Scan a single root region and mark everything reachable from it.
753 void scanRootRegion(HeapRegion* hr, uint worker_id);
754
755 // Do concurrent phase of marking, to a tentative transitive closure.
756 void markFromRoots();
757
758 void checkpointRootsFinal(bool clear_all_soft_refs);
759 void checkpointRootsFinalWork();
760 void cleanup();
761 void completeCleanup();
762
763 // Mark in the previous bitmap. NB: this is usually read-only, so use
764 // this carefully!
765 inline void markPrev(oop p);
766
767 // Clears marks for all objects in the given range, for the prev,
768 // next, or both bitmaps. NB: the previous bitmap is usually
769 // read-only, so use this carefully!
770 void clearRangePrevBitmap(MemRegion mr);
771 void clearRangeNextBitmap(MemRegion mr);
772 void clearRangeBothBitmaps(MemRegion mr);
773
774 // Notify data structures that a GC has started.
775 void note_start_of_gc() {
776 _markStack.note_start_of_gc();
777 }
778
779 // Notify data structures that a GC is finished.
780 void note_end_of_gc() {
781 _markStack.note_end_of_gc();
782 }
783
784 // Verify that there are no CSet oops on the stacks (taskqueues /
785 // global mark stack), enqueued SATB buffers, per-thread SATB
786 // buffers, and fingers (global / per-task). The boolean parameters
787 // decide which of the above data structures to verify. If marking
788 // is not in progress, it's a no-op.
789 void verify_no_cset_oops(bool verify_stacks,
790 bool verify_enqueued_buffers,
791 bool verify_thread_buffers,
792 bool verify_fingers) PRODUCT_RETURN;
793
794 // It is called at the end of an evacuation pause during marking so
795 // that CM is notified of where the new end of the heap is. It
796 // doesn't do anything if concurrent_marking_in_progress() is false,
797 // unless the force parameter is true.
798 void update_g1_committed(bool force = false);
799
800 bool isMarked(oop p) const {
801 assert(p != NULL && p->is_oop(), "expected an oop");
802 HeapWord* addr = (HeapWord*)p;
803 assert(addr >= _nextMarkBitMap->startWord() ||
804 addr < _nextMarkBitMap->endWord(), "in a region");
805
806 return _nextMarkBitMap->isMarked(addr);
807 }
808
809 inline bool not_yet_marked(oop p) const;
810
811 // XXX Debug code
812 bool containing_card_is_marked(void* p);
813 bool containing_cards_are_marked(void* start, void* last);
814
815 bool isPrevMarked(oop p) const {
816 assert(p != NULL && p->is_oop(), "expected an oop");
817 HeapWord* addr = (HeapWord*)p;
818 assert(addr >= _prevMarkBitMap->startWord() ||
819 addr < _prevMarkBitMap->endWord(), "in a region");
820
821 return _prevMarkBitMap->isMarked(addr);
822 }
823
824 inline bool do_yield_check(uint worker_i = 0);
825
826 // Called to abort the marking cycle after a Full GC takes place.
827 void abort();
828
829 bool has_aborted() { return _has_aborted; }
830
831 const GCId& concurrent_gc_id();
832
833 // This prints the global/local fingers. It is used for debugging.
834 NOT_PRODUCT(void print_finger();)
835
836 void print_summary_info();
837
838 void print_worker_threads_on(outputStream* st) const;
839
840 void print_on_error(outputStream* st) const;
841
842 // The following indicate whether a given verbose level has been
843 // set. Notice that anything above stats is conditional to
844 // _MARKING_VERBOSE_ having been set to 1
845 bool verbose_stats() {
846 return _verbose_level >= stats_verbose;
847 }
848 bool verbose_low() {
849 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
850 }
851 bool verbose_medium() {
852 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
853 }
854 bool verbose_high() {
855 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
856 }
857
858 // Liveness counting
859
860 // Utility routine to set an exclusive range of cards on the given
861 // card liveness bitmap
862 inline void set_card_bitmap_range(BitMap* card_bm,
863 BitMap::idx_t start_idx,
864 BitMap::idx_t end_idx,
865 bool is_par);
866
867 // Returns the card number of the bottom of the G1 heap.
868 // Used in biasing indices into accounting card bitmaps.
869 intptr_t heap_bottom_card_num() const {
870 return _heap_bottom_card_num;
871 }
872
873 // Returns the card bitmap for a given task or worker id.
874 BitMap* count_card_bitmap_for(uint worker_id) {
875 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
876 assert(_count_card_bitmaps != NULL, "uninitialized");
877 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
878 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
879 return task_card_bm;
880 }
881
882 // Returns the array containing the marked bytes for each region,
883 // for the given worker or task id.
884 size_t* count_marked_bytes_array_for(uint worker_id) {
885 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
886 assert(_count_marked_bytes != NULL, "uninitialized");
887 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
888 assert(marked_bytes_array != NULL, "uninitialized");
889 return marked_bytes_array;
890 }
891
892 // Returns the index in the liveness accounting card table bitmap
893 // for the given address
894 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
895
896 // Counts the size of the given memory region in the the given
897 // marked_bytes array slot for the given HeapRegion.
898 // Sets the bits in the given card bitmap that are associated with the
899 // cards that are spanned by the memory region.
900 inline void count_region(MemRegion mr, HeapRegion* hr,
901 size_t* marked_bytes_array,
902 BitMap* task_card_bm);
903
904 // Counts the given memory region in the task/worker counting
905 // data structures for the given worker id.
906 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
907
908 // Counts the given memory region in the task/worker counting
909 // data structures for the given worker id.
910 inline void count_region(MemRegion mr, uint worker_id);
911
912 // Counts the given object in the given task/worker counting
913 // data structures.
914 inline void count_object(oop obj, HeapRegion* hr,
915 size_t* marked_bytes_array,
916 BitMap* task_card_bm);
917
918 // Counts the given object in the task/worker counting data
919 // structures for the given worker id.
920 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
921
922 // Attempts to mark the given object and, if successful, counts
923 // the object in the given task/worker counting structures.
924 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
925 size_t* marked_bytes_array,
926 BitMap* task_card_bm);
927
928 // Attempts to mark the given object and, if successful, counts
929 // the object in the task/worker counting structures for the
930 // given worker id.
931 inline bool par_mark_and_count(oop obj, size_t word_size,
932 HeapRegion* hr, uint worker_id);
933
934 // Attempts to mark the given object and, if successful, counts
935 // the object in the task/worker counting structures for the
936 // given worker id.
937 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
938
939 // Similar to the above routine but we don't know the heap region that
940 // contains the object to be marked/counted, which this routine looks up.
941 inline bool par_mark_and_count(oop obj, uint worker_id);
942
943 // Similar to the above routine but there are times when we cannot
944 // safely calculate the size of obj due to races and we, therefore,
945 // pass the size in as a parameter. It is the caller's responsibility
946 // to ensure that the size passed in for obj is valid.
947 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
948
949 // Unconditionally mark the given object, and unconditionally count
950 // the object in the counting structures for worker id 0.
951 // Should *not* be called from parallel code.
952 inline bool mark_and_count(oop obj, HeapRegion* hr);
953
954 // Similar to the above routine but we don't know the heap region that
955 // contains the object to be marked/counted, which this routine looks up.
956 // Should *not* be called from parallel code.
957 inline bool mark_and_count(oop obj);
958
959 // Returns true if initialization was successfully completed.
960 bool completed_initialization() const {
961 return _completed_initialization;
962 }
963
964 protected:
965 // Clear all the per-task bitmaps and arrays used to store the
966 // counting data.
967 void clear_all_count_data();
968
969 // Aggregates the counting data for each worker/task
970 // that was constructed while marking. Also sets
971 // the amount of marked bytes for each region and
972 // the top at concurrent mark count.
973 void aggregate_count_data();
974
975 // Verification routine
976 void verify_count_data();
977 };
978
979 // A class representing a marking task.
980 class CMTask : public TerminatorTerminator {
981 private:
982 enum PrivateConstants {
983 // the regular clock call is called once the scanned words reaches
984 // this limit
985 words_scanned_period = 12*1024,
986 // the regular clock call is called once the number of visited
987 // references reaches this limit
988 refs_reached_period = 384,
989 // initial value for the hash seed, used in the work stealing code
990 init_hash_seed = 17,
991 // how many entries will be transferred between global stack and
992 // local queues
993 global_stack_transfer_size = 16
994 };
995
996 uint _worker_id;
997 G1CollectedHeap* _g1h;
998 ConcurrentMark* _cm;
999 CMBitMap* _nextMarkBitMap;
1000 // the task queue of this task
1001 CMTaskQueue* _task_queue;
1002 private:
1003 // the task queue set---needed for stealing
1004 CMTaskQueueSet* _task_queues;
1005 // indicates whether the task has been claimed---this is only for
1006 // debugging purposes
1007 bool _claimed;
1008
1009 // number of calls to this task
1010 int _calls;
1011
1012 // when the virtual timer reaches this time, the marking step should
1013 // exit
1014 double _time_target_ms;
1015 // the start time of the current marking step
1016 double _start_time_ms;
1017
1018 // the oop closure used for iterations over oops
1019 G1CMOopClosure* _cm_oop_closure;
1020
1021 // the region this task is scanning, NULL if we're not scanning any
1022 HeapRegion* _curr_region;
1023 // the local finger of this task, NULL if we're not scanning a region
1024 HeapWord* _finger;
1025 // limit of the region this task is scanning, NULL if we're not scanning one
1026 HeapWord* _region_limit;
1027
1028 // the number of words this task has scanned
1029 size_t _words_scanned;
1030 // When _words_scanned reaches this limit, the regular clock is
1031 // called. Notice that this might be decreased under certain
1032 // circumstances (i.e. when we believe that we did an expensive
1033 // operation).
1034 size_t _words_scanned_limit;
1035 // the initial value of _words_scanned_limit (i.e. what it was
1036 // before it was decreased).
1037 size_t _real_words_scanned_limit;
1038
1039 // the number of references this task has visited
1040 size_t _refs_reached;
1041 // When _refs_reached reaches this limit, the regular clock is
1042 // called. Notice this this might be decreased under certain
1043 // circumstances (i.e. when we believe that we did an expensive
1044 // operation).
1045 size_t _refs_reached_limit;
1046 // the initial value of _refs_reached_limit (i.e. what it was before
1047 // it was decreased).
1048 size_t _real_refs_reached_limit;
1049
1050 // used by the work stealing stuff
1051 int _hash_seed;
1052 // if this is true, then the task has aborted for some reason
1053 bool _has_aborted;
1054 // set when the task aborts because it has met its time quota
1055 bool _has_timed_out;
1056 // true when we're draining SATB buffers; this avoids the task
1057 // aborting due to SATB buffers being available (as we're already
1058 // dealing with them)
1059 bool _draining_satb_buffers;
1060
1061 // number sequence of past step times
1062 NumberSeq _step_times_ms;
1063 // elapsed time of this task
1064 double _elapsed_time_ms;
1065 // termination time of this task
1066 double _termination_time_ms;
1067 // when this task got into the termination protocol
1068 double _termination_start_time_ms;
1069
1070 // true when the task is during a concurrent phase, false when it is
1071 // in the remark phase (so, in the latter case, we do not have to
1072 // check all the things that we have to check during the concurrent
1073 // phase, i.e. SATB buffer availability...)
1074 bool _concurrent;
1075
1076 TruncatedSeq _marking_step_diffs_ms;
1077
1078 // Counting data structures. Embedding the task's marked_bytes_array
1079 // and card bitmap into the actual task saves having to go through
1080 // the ConcurrentMark object.
1081 size_t* _marked_bytes_array;
1082 BitMap* _card_bm;
1083
1084 // LOTS of statistics related with this task
1085 #if _MARKING_STATS_
1086 NumberSeq _all_clock_intervals_ms;
1087 double _interval_start_time_ms;
1088
1089 int _aborted;
1090 int _aborted_overflow;
1091 int _aborted_cm_aborted;
1092 int _aborted_yield;
1093 int _aborted_timed_out;
1094 int _aborted_satb;
1095 int _aborted_termination;
1096
1097 int _steal_attempts;
1098 int _steals;
1099
1100 int _clock_due_to_marking;
1101 int _clock_due_to_scanning;
1102
1103 int _local_pushes;
1104 int _local_pops;
1105 int _local_max_size;
1106 int _objs_scanned;
1107
1108 int _global_pushes;
1109 int _global_pops;
1110 int _global_max_size;
1111
1112 int _global_transfers_to;
1113 int _global_transfers_from;
1114
1115 int _regions_claimed;
1116 int _objs_found_on_bitmap;
1117
1118 int _satb_buffers_processed;
1119 #endif // _MARKING_STATS_
1120
1121 // it updates the local fields after this task has claimed
1122 // a new region to scan
1123 void setup_for_region(HeapRegion* hr);
1124 // it brings up-to-date the limit of the region
1125 void update_region_limit();
1126
1127 // called when either the words scanned or the refs visited limit
1128 // has been reached
1129 void reached_limit();
1130 // recalculates the words scanned and refs visited limits
1131 void recalculate_limits();
1132 // decreases the words scanned and refs visited limits when we reach
1133 // an expensive operation
1134 void decrease_limits();
1135 // it checks whether the words scanned or refs visited reached their
1136 // respective limit and calls reached_limit() if they have
1137 void check_limits() {
1138 if (_words_scanned >= _words_scanned_limit ||
1139 _refs_reached >= _refs_reached_limit) {
1140 reached_limit();
1141 }
1142 }
1143 // this is supposed to be called regularly during a marking step as
1144 // it checks a bunch of conditions that might cause the marking step
1145 // to abort
1146 void regular_clock_call();
1147 bool concurrent() { return _concurrent; }
1148
1149 public:
1150 // It resets the task; it should be called right at the beginning of
1151 // a marking phase.
1152 void reset(CMBitMap* _nextMarkBitMap);
1153 // it clears all the fields that correspond to a claimed region.
1154 void clear_region_fields();
1155
1156 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1157
1158 // The main method of this class which performs a marking step
1159 // trying not to exceed the given duration. However, it might exit
1160 // prematurely, according to some conditions (i.e. SATB buffers are
1161 // available for processing).
1162 void do_marking_step(double target_ms,
1163 bool do_termination,
1164 bool is_serial);
1165
1166 // These two calls start and stop the timer
1167 void record_start_time() {
1168 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1169 }
1170 void record_end_time() {
1171 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1172 }
1173
1174 // returns the worker ID associated with this task.
1175 uint worker_id() { return _worker_id; }
1176
1177 // From TerminatorTerminator. It determines whether this task should
1178 // exit the termination protocol after it's entered it.
1179 virtual bool should_exit_termination();
1180
1181 // Resets the local region fields after a task has finished scanning a
1182 // region; or when they have become stale as a result of the region
1183 // being evacuated.
1184 void giveup_current_region();
1185
1186 HeapWord* finger() { return _finger; }
1187
1188 bool has_aborted() { return _has_aborted; }
1189 void set_has_aborted() { _has_aborted = true; }
1190 void clear_has_aborted() { _has_aborted = false; }
1191 bool has_timed_out() { return _has_timed_out; }
1192 bool claimed() { return _claimed; }
1193
1194 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1195
1196 // It grays the object by marking it and, if necessary, pushing it
1197 // on the local queue
1198 inline void deal_with_reference(oop obj);
1199
1200 // It scans an object and visits its children.
1201 void scan_object(oop obj);
1202
1203 // It pushes an object on the local queue.
1204 inline void push(oop obj);
1205
1206 // These two move entries to/from the global stack.
1207 void move_entries_to_global_stack();
1208 void get_entries_from_global_stack();
1209
1210 // It pops and scans objects from the local queue. If partially is
1211 // true, then it stops when the queue size is of a given limit. If
1212 // partially is false, then it stops when the queue is empty.
1213 void drain_local_queue(bool partially);
1214 // It moves entries from the global stack to the local queue and
1215 // drains the local queue. If partially is true, then it stops when
1216 // both the global stack and the local queue reach a given size. If
1217 // partially if false, it tries to empty them totally.
1218 void drain_global_stack(bool partially);
1219 // It keeps picking SATB buffers and processing them until no SATB
1220 // buffers are available.
1221 void drain_satb_buffers();
1222
1223 // moves the local finger to a new location
1224 inline void move_finger_to(HeapWord* new_finger) {
1225 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1226 _finger = new_finger;
1227 }
1228
1229 CMTask(uint worker_id, ConcurrentMark *cm,
1230 size_t* marked_bytes, BitMap* card_bm,
1231 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1232
1233 // it prints statistics associated with this task
1234 void print_stats();
1235
1236 #if _MARKING_STATS_
1237 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1238 #endif // _MARKING_STATS_
1239 };
1240
1241 // Class that's used to to print out per-region liveness
1242 // information. It's currently used at the end of marking and also
1243 // after we sort the old regions at the end of the cleanup operation.
1244 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1245 private:
1246 outputStream* _out;
1247
1248 // Accumulators for these values.
1249 size_t _total_used_bytes;
1250 size_t _total_capacity_bytes;
1251 size_t _total_prev_live_bytes;
1252 size_t _total_next_live_bytes;
1253
1254 // These are set up when we come across a "stars humongous" region
1255 // (as this is where most of this information is stored, not in the
1256 // subsequent "continues humongous" regions). After that, for every
1257 // region in a given humongous region series we deduce the right
1258 // values for it by simply subtracting the appropriate amount from
1259 // these fields. All these values should reach 0 after we've visited
1260 // the last region in the series.
1261 size_t _hum_used_bytes;
1262 size_t _hum_capacity_bytes;
1263 size_t _hum_prev_live_bytes;
1264 size_t _hum_next_live_bytes;
1265
1266 // Accumulator for the remembered set size
1267 size_t _total_remset_bytes;
1268
1269 // Accumulator for strong code roots memory size
1270 size_t _total_strong_code_roots_bytes;
1271
1272 static double perc(size_t val, size_t total) {
1273 if (total == 0) {
1274 return 0.0;
1275 } else {
1276 return 100.0 * ((double) val / (double) total);
1277 }
1278 }
1279
1280 static double bytes_to_mb(size_t val) {
1281 return (double) val / (double) M;
1282 }
1283
1284 // See the .cpp file.
1285 size_t get_hum_bytes(size_t* hum_bytes);
1286 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1287 size_t* prev_live_bytes, size_t* next_live_bytes);
1288
1289 public:
1290 // The header and footer are printed in the constructor and
1291 // destructor respectively.
1292 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1293 virtual bool doHeapRegion(HeapRegion* r);
1294 ~G1PrintRegionLivenessInfoClosure();
1295 };
1296
1297 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP