Thu, 21 Aug 2014 16:44:41 +0200
8055098: WB API should be extended to provide information about size and age of object.
Summary: Extend the WhiteBox API to provide information about the size and age of objects. Further add a mechanism to trigger a young GC.
Reviewed-by: tschatzl, sjohanss
Contributed-by: Leonid Mesnik <leonid.mesnik@oracle.com>
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 */
25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
28 #include "classfile/javaClasses.hpp"
29 #include "gc_implementation/g1/heapRegionSet.hpp"
30 #include "gc_implementation/g1/g1RegionToSpaceMapper.hpp"
31 #include "gc_implementation/shared/gcId.hpp"
32 #include "utilities/taskqueue.hpp"
34 class G1CollectedHeap;
35 class CMBitMap;
36 class CMTask;
37 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
38 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
40 // Closure used by CM during concurrent reference discovery
41 // and reference processing (during remarking) to determine
42 // if a particular object is alive. It is primarily used
43 // to determine if referents of discovered reference objects
44 // are alive. An instance is also embedded into the
45 // reference processor as the _is_alive_non_header field
46 class G1CMIsAliveClosure: public BoolObjectClosure {
47 G1CollectedHeap* _g1;
48 public:
49 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
51 bool do_object_b(oop obj);
52 };
54 // A generic CM bit map. This is essentially a wrapper around the BitMap
55 // class, with one bit per (1<<_shifter) HeapWords.
57 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
58 protected:
59 HeapWord* _bmStartWord; // base address of range covered by map
60 size_t _bmWordSize; // map size (in #HeapWords covered)
61 const int _shifter; // map to char or bit
62 BitMap _bm; // the bit map itself
64 public:
65 // constructor
66 CMBitMapRO(int shifter);
68 enum { do_yield = true };
70 // inquiries
71 HeapWord* startWord() const { return _bmStartWord; }
72 size_t sizeInWords() const { return _bmWordSize; }
73 // the following is one past the last word in space
74 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
76 // read marks
78 bool isMarked(HeapWord* addr) const {
79 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
80 "outside underlying space?");
81 return _bm.at(heapWordToOffset(addr));
82 }
84 // iteration
85 inline bool iterate(BitMapClosure* cl, MemRegion mr);
86 inline bool iterate(BitMapClosure* cl);
88 // Return the address corresponding to the next marked bit at or after
89 // "addr", and before "limit", if "limit" is non-NULL. If there is no
90 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
91 HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
92 const HeapWord* limit = NULL) const;
93 // Return the address corresponding to the next unmarked bit at or after
94 // "addr", and before "limit", if "limit" is non-NULL. If there is no
95 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
96 HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr,
97 const HeapWord* limit = NULL) const;
99 // conversion utilities
100 HeapWord* offsetToHeapWord(size_t offset) const {
101 return _bmStartWord + (offset << _shifter);
102 }
103 size_t heapWordToOffset(const HeapWord* addr) const {
104 return pointer_delta(addr, _bmStartWord) >> _shifter;
105 }
106 int heapWordDiffToOffsetDiff(size_t diff) const;
108 // The argument addr should be the start address of a valid object
109 HeapWord* nextObject(HeapWord* addr) {
110 oop obj = (oop) addr;
111 HeapWord* res = addr + obj->size();
112 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
113 return res;
114 }
116 void print_on_error(outputStream* st, const char* prefix) const;
118 // debugging
119 NOT_PRODUCT(bool covers(MemRegion rs) const;)
120 };
122 class CMBitMapMappingChangedListener : public G1MappingChangedListener {
123 private:
124 CMBitMap* _bm;
125 public:
126 CMBitMapMappingChangedListener() : _bm(NULL) {}
128 void set_bitmap(CMBitMap* bm) { _bm = bm; }
130 virtual void on_commit(uint start_idx, size_t num_regions);
131 };
133 class CMBitMap : public CMBitMapRO {
134 private:
135 CMBitMapMappingChangedListener _listener;
137 public:
138 static size_t compute_size(size_t heap_size);
139 // Returns the amount of bytes on the heap between two marks in the bitmap.
140 static size_t mark_distance();
142 CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
144 // Initializes the underlying BitMap to cover the given area.
145 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
147 // Write marks.
148 inline void mark(HeapWord* addr);
149 inline void clear(HeapWord* addr);
150 inline bool parMark(HeapWord* addr);
151 inline bool parClear(HeapWord* addr);
153 void markRange(MemRegion mr);
154 void clearRange(MemRegion mr);
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);
165 // Clear the whole mark bitmap.
166 void clearAll();
167 };
169 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
170 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
171 VirtualSpace _virtual_space; // Underlying backing store for actual stack
172 ConcurrentMark* _cm;
173 oop* _base; // bottom of stack
174 jint _index; // one more than last occupied index
175 jint _capacity; // max #elements
176 jint _saved_index; // value of _index saved at start of GC
177 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
179 bool _overflow;
180 bool _should_expand;
181 DEBUG_ONLY(bool _drain_in_progress;)
182 DEBUG_ONLY(bool _drain_in_progress_yields;)
184 public:
185 CMMarkStack(ConcurrentMark* cm);
186 ~CMMarkStack();
188 #ifndef PRODUCT
189 jint max_depth() const {
190 return _max_depth;
191 }
192 #endif
194 bool allocate(size_t capacity);
196 oop pop() {
197 if (!isEmpty()) {
198 return _base[--_index] ;
199 }
200 return NULL;
201 }
203 // If overflow happens, don't do the push, and record the overflow.
204 // *Requires* that "ptr" is already marked.
205 void push(oop ptr) {
206 if (isFull()) {
207 // Record overflow.
208 _overflow = true;
209 return;
210 } else {
211 _base[_index++] = ptr;
212 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
213 }
214 }
215 // Non-block impl. Note: concurrency is allowed only with other
216 // "par_push" operations, not with "pop" or "drain". We would need
217 // parallel versions of them if such concurrency was desired.
218 void par_push(oop ptr);
220 // Pushes the first "n" elements of "ptr_arr" on the stack.
221 // Non-block impl. Note: concurrency is allowed only with other
222 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
223 void par_adjoin_arr(oop* ptr_arr, int n);
225 // Pushes the first "n" elements of "ptr_arr" on the stack.
226 // Locking impl: concurrency is allowed only with
227 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
228 // locking strategy.
229 void par_push_arr(oop* ptr_arr, int n);
231 // If returns false, the array was empty. Otherwise, removes up to "max"
232 // elements from the stack, and transfers them to "ptr_arr" in an
233 // unspecified order. The actual number transferred is given in "n" ("n
234 // == 0" is deliberately redundant with the return value.) Locking impl:
235 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
236 // operations, which use the same locking strategy.
237 bool par_pop_arr(oop* ptr_arr, int max, int* n);
239 // Drain the mark stack, applying the given closure to all fields of
240 // objects on the stack. (That is, continue until the stack is empty,
241 // even if closure applications add entries to the stack.) The "bm"
242 // argument, if non-null, may be used to verify that only marked objects
243 // are on the mark stack. If "yield_after" is "true", then the
244 // concurrent marker performing the drain offers to yield after
245 // processing each object. If a yield occurs, stops the drain operation
246 // and returns false. Otherwise, returns true.
247 template<class OopClosureClass>
248 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
250 bool isEmpty() { return _index == 0; }
251 bool isFull() { return _index == _capacity; }
252 int maxElems() { return _capacity; }
254 bool overflow() { return _overflow; }
255 void clear_overflow() { _overflow = false; }
257 bool should_expand() const { return _should_expand; }
258 void set_should_expand();
260 // Expand the stack, typically in response to an overflow condition
261 void expand();
263 int size() { return _index; }
265 void setEmpty() { _index = 0; clear_overflow(); }
267 // Record the current index.
268 void note_start_of_gc();
270 // Make sure that we have not added any entries to the stack during GC.
271 void note_end_of_gc();
273 // iterate over the oops in the mark stack, up to the bound recorded via
274 // the call above.
275 void oops_do(OopClosure* f);
276 };
278 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
279 private:
280 #ifndef PRODUCT
281 uintx _num_remaining;
282 bool _force;
283 #endif // !defined(PRODUCT)
285 public:
286 void init() PRODUCT_RETURN;
287 void update() PRODUCT_RETURN;
288 bool should_force() PRODUCT_RETURN_( return false; );
289 };
291 // this will enable a variety of different statistics per GC task
292 #define _MARKING_STATS_ 0
293 // this will enable the higher verbose levels
294 #define _MARKING_VERBOSE_ 0
296 #if _MARKING_STATS_
297 #define statsOnly(statement) \
298 do { \
299 statement ; \
300 } while (0)
301 #else // _MARKING_STATS_
302 #define statsOnly(statement) \
303 do { \
304 } while (0)
305 #endif // _MARKING_STATS_
307 typedef enum {
308 no_verbose = 0, // verbose turned off
309 stats_verbose, // only prints stats at the end of marking
310 low_verbose, // low verbose, mostly per region and per major event
311 medium_verbose, // a bit more detailed than low
312 high_verbose // per object verbose
313 } CMVerboseLevel;
315 class YoungList;
317 // Root Regions are regions that are not empty at the beginning of a
318 // marking cycle and which we might collect during an evacuation pause
319 // while the cycle is active. Given that, during evacuation pauses, we
320 // do not copy objects that are explicitly marked, what we have to do
321 // for the root regions is to scan them and mark all objects reachable
322 // from them. According to the SATB assumptions, we only need to visit
323 // each object once during marking. So, as long as we finish this scan
324 // before the next evacuation pause, we can copy the objects from the
325 // root regions without having to mark them or do anything else to them.
326 //
327 // Currently, we only support root region scanning once (at the start
328 // of the marking cycle) and the root regions are all the survivor
329 // regions populated during the initial-mark pause.
330 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
331 private:
332 YoungList* _young_list;
333 ConcurrentMark* _cm;
335 volatile bool _scan_in_progress;
336 volatile bool _should_abort;
337 HeapRegion* volatile _next_survivor;
339 public:
340 CMRootRegions();
341 // We actually do most of the initialization in this method.
342 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
344 // Reset the claiming / scanning of the root regions.
345 void prepare_for_scan();
347 // Forces get_next() to return NULL so that the iteration aborts early.
348 void abort() { _should_abort = true; }
350 // Return true if the CM thread are actively scanning root regions,
351 // false otherwise.
352 bool scan_in_progress() { return _scan_in_progress; }
354 // Claim the next root region to scan atomically, or return NULL if
355 // all have been claimed.
356 HeapRegion* claim_next();
358 // Flag that we're done with root region scanning and notify anyone
359 // who's waiting on it. If aborted is false, assume that all regions
360 // have been claimed.
361 void scan_finished();
363 // If CM threads are still scanning root regions, wait until they
364 // are done. Return true if we had to wait, false otherwise.
365 bool wait_until_scan_finished();
366 };
368 class ConcurrentMarkThread;
370 class ConcurrentMark: public CHeapObj<mtGC> {
371 friend class CMMarkStack;
372 friend class ConcurrentMarkThread;
373 friend class CMTask;
374 friend class CMBitMapClosure;
375 friend class CMGlobalObjectClosure;
376 friend class CMRemarkTask;
377 friend class CMConcurrentMarkingTask;
378 friend class G1ParNoteEndTask;
379 friend class CalcLiveObjectsClosure;
380 friend class G1CMRefProcTaskProxy;
381 friend class G1CMRefProcTaskExecutor;
382 friend class G1CMKeepAliveAndDrainClosure;
383 friend class G1CMDrainMarkingStackClosure;
385 protected:
386 ConcurrentMarkThread* _cmThread; // the thread doing the work
387 G1CollectedHeap* _g1h; // the heap.
388 uint _parallel_marking_threads; // the number of marking
389 // threads we're use
390 uint _max_parallel_marking_threads; // max number of marking
391 // threads we'll ever use
392 double _sleep_factor; // how much we have to sleep, with
393 // respect to the work we just did, to
394 // meet the marking overhead goal
395 double _marking_task_overhead; // marking target overhead for
396 // a single task
398 // same as the two above, but for the cleanup task
399 double _cleanup_sleep_factor;
400 double _cleanup_task_overhead;
402 FreeRegionList _cleanup_list;
404 // Concurrent marking support structures
405 CMBitMap _markBitMap1;
406 CMBitMap _markBitMap2;
407 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
408 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
410 BitMap _region_bm;
411 BitMap _card_bm;
413 // Heap bounds
414 HeapWord* _heap_start;
415 HeapWord* _heap_end;
417 // Root region tracking and claiming.
418 CMRootRegions _root_regions;
420 // For gray objects
421 CMMarkStack _markStack; // Grey objects behind global finger.
422 HeapWord* volatile _finger; // the global finger, region aligned,
423 // always points to the end of the
424 // last claimed region
426 // marking tasks
427 uint _max_worker_id;// maximum worker id
428 uint _active_tasks; // task num currently active
429 CMTask** _tasks; // task queue array (max_worker_id len)
430 CMTaskQueueSet* _task_queues; // task queue set
431 ParallelTaskTerminator _terminator; // for termination
433 // Two sync barriers that are used to synchronise tasks when an
434 // overflow occurs. The algorithm is the following. All tasks enter
435 // the first one to ensure that they have all stopped manipulating
436 // the global data structures. After they exit it, they re-initialise
437 // their data structures and task 0 re-initialises the global data
438 // structures. Then, they enter the second sync barrier. This
439 // ensure, that no task starts doing work before all data
440 // structures (local and global) have been re-initialised. When they
441 // exit it, they are free to start working again.
442 WorkGangBarrierSync _first_overflow_barrier_sync;
443 WorkGangBarrierSync _second_overflow_barrier_sync;
445 // this is set by any task, when an overflow on the global data
446 // structures is detected.
447 volatile bool _has_overflown;
448 // true: marking is concurrent, false: we're in remark
449 volatile bool _concurrent;
450 // set at the end of a Full GC so that marking aborts
451 volatile bool _has_aborted;
452 GCId _aborted_gc_id;
454 // used when remark aborts due to an overflow to indicate that
455 // another concurrent marking phase should start
456 volatile bool _restart_for_overflow;
458 // This is true from the very start of concurrent marking until the
459 // point when all the tasks complete their work. It is really used
460 // to determine the points between the end of concurrent marking and
461 // time of remark.
462 volatile bool _concurrent_marking_in_progress;
464 // verbose level
465 CMVerboseLevel _verbose_level;
467 // All of these times are in ms.
468 NumberSeq _init_times;
469 NumberSeq _remark_times;
470 NumberSeq _remark_mark_times;
471 NumberSeq _remark_weak_ref_times;
472 NumberSeq _cleanup_times;
473 double _total_counting_time;
474 double _total_rs_scrub_time;
476 double* _accum_task_vtime; // accumulated task vtime
478 FlexibleWorkGang* _parallel_workers;
480 ForceOverflowSettings _force_overflow_conc;
481 ForceOverflowSettings _force_overflow_stw;
483 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
484 void weakRefsWork(bool clear_all_soft_refs);
486 void swapMarkBitMaps();
488 // It resets the global marking data structures, as well as the
489 // task local ones; should be called during initial mark.
490 void reset();
492 // Resets all the marking data structures. Called when we have to restart
493 // marking or when marking completes (via set_non_marking_state below).
494 void reset_marking_state(bool clear_overflow = true);
496 // We do this after we're done with marking so that the marking data
497 // structures are initialised to a sensible and predictable state.
498 void set_non_marking_state();
500 // Called to indicate how many threads are currently active.
501 void set_concurrency(uint active_tasks);
503 // It should be called to indicate which phase we're in (concurrent
504 // mark or remark) and how many threads are currently active.
505 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
507 // prints all gathered CM-related statistics
508 void print_stats();
510 bool cleanup_list_is_empty() {
511 return _cleanup_list.is_empty();
512 }
514 // accessor methods
515 uint parallel_marking_threads() const { return _parallel_marking_threads; }
516 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
517 double sleep_factor() { return _sleep_factor; }
518 double marking_task_overhead() { return _marking_task_overhead;}
519 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
520 double cleanup_task_overhead() { return _cleanup_task_overhead;}
522 bool use_parallel_marking_threads() const {
523 assert(parallel_marking_threads() <=
524 max_parallel_marking_threads(), "sanity");
525 assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
526 parallel_marking_threads() > 0,
527 "parallel workers not set up correctly");
528 return _parallel_workers != NULL;
529 }
531 HeapWord* finger() { return _finger; }
532 bool concurrent() { return _concurrent; }
533 uint active_tasks() { return _active_tasks; }
534 ParallelTaskTerminator* terminator() { return &_terminator; }
536 // It claims the next available region to be scanned by a marking
537 // task/thread. It might return NULL if the next region is empty or
538 // we have run out of regions. In the latter case, out_of_regions()
539 // determines whether we've really run out of regions or the task
540 // should call claim_region() again. This might seem a bit
541 // awkward. Originally, the code was written so that claim_region()
542 // either successfully returned with a non-empty region or there
543 // were no more regions to be claimed. The problem with this was
544 // that, in certain circumstances, it iterated over large chunks of
545 // the heap finding only empty regions and, while it was working, it
546 // was preventing the calling task to call its regular clock
547 // method. So, this way, each task will spend very little time in
548 // claim_region() and is allowed to call the regular clock method
549 // frequently.
550 HeapRegion* claim_region(uint worker_id);
552 // It determines whether we've run out of regions to scan. Note that
553 // the finger can point past the heap end in case the heap was expanded
554 // to satisfy an allocation without doing a GC. This is fine, because all
555 // objects in those regions will be considered live anyway because of
556 // SATB guarantees (i.e. their TAMS will be equal to bottom).
557 bool out_of_regions() { return _finger >= _heap_end; }
559 // Returns the task with the given id
560 CMTask* task(int id) {
561 assert(0 <= id && id < (int) _active_tasks,
562 "task id not within active bounds");
563 return _tasks[id];
564 }
566 // Returns the task queue with the given id
567 CMTaskQueue* task_queue(int id) {
568 assert(0 <= id && id < (int) _active_tasks,
569 "task queue id not within active bounds");
570 return (CMTaskQueue*) _task_queues->queue(id);
571 }
573 // Returns the task queue set
574 CMTaskQueueSet* task_queues() { return _task_queues; }
576 // Access / manipulation of the overflow flag which is set to
577 // indicate that the global stack has overflown
578 bool has_overflown() { return _has_overflown; }
579 void set_has_overflown() { _has_overflown = true; }
580 void clear_has_overflown() { _has_overflown = false; }
581 bool restart_for_overflow() { return _restart_for_overflow; }
583 // Methods to enter the two overflow sync barriers
584 void enter_first_sync_barrier(uint worker_id);
585 void enter_second_sync_barrier(uint worker_id);
587 ForceOverflowSettings* force_overflow_conc() {
588 return &_force_overflow_conc;
589 }
591 ForceOverflowSettings* force_overflow_stw() {
592 return &_force_overflow_stw;
593 }
595 ForceOverflowSettings* force_overflow() {
596 if (concurrent()) {
597 return force_overflow_conc();
598 } else {
599 return force_overflow_stw();
600 }
601 }
603 // Live Data Counting data structures...
604 // These data structures are initialized at the start of
605 // marking. They are written to while marking is active.
606 // They are aggregated during remark; the aggregated values
607 // are then used to populate the _region_bm, _card_bm, and
608 // the total live bytes, which are then subsequently updated
609 // during cleanup.
611 // An array of bitmaps (one bit map per task). Each bitmap
612 // is used to record the cards spanned by the live objects
613 // marked by that task/worker.
614 BitMap* _count_card_bitmaps;
616 // Used to record the number of marked live bytes
617 // (for each region, by worker thread).
618 size_t** _count_marked_bytes;
620 // Card index of the bottom of the G1 heap. Used for biasing indices into
621 // the card bitmaps.
622 intptr_t _heap_bottom_card_num;
624 // Set to true when initialization is complete
625 bool _completed_initialization;
627 public:
628 // Manipulation of the global mark stack.
629 // Notice that the first mark_stack_push is CAS-based, whereas the
630 // two below are Mutex-based. This is OK since the first one is only
631 // called during evacuation pauses and doesn't compete with the
632 // other two (which are called by the marking tasks during
633 // concurrent marking or remark).
634 bool mark_stack_push(oop p) {
635 _markStack.par_push(p);
636 if (_markStack.overflow()) {
637 set_has_overflown();
638 return false;
639 }
640 return true;
641 }
642 bool mark_stack_push(oop* arr, int n) {
643 _markStack.par_push_arr(arr, n);
644 if (_markStack.overflow()) {
645 set_has_overflown();
646 return false;
647 }
648 return true;
649 }
650 void mark_stack_pop(oop* arr, int max, int* n) {
651 _markStack.par_pop_arr(arr, max, n);
652 }
653 size_t mark_stack_size() { return _markStack.size(); }
654 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
655 bool mark_stack_overflow() { return _markStack.overflow(); }
656 bool mark_stack_empty() { return _markStack.isEmpty(); }
658 CMRootRegions* root_regions() { return &_root_regions; }
660 bool concurrent_marking_in_progress() {
661 return _concurrent_marking_in_progress;
662 }
663 void set_concurrent_marking_in_progress() {
664 _concurrent_marking_in_progress = true;
665 }
666 void clear_concurrent_marking_in_progress() {
667 _concurrent_marking_in_progress = false;
668 }
670 void update_accum_task_vtime(int i, double vtime) {
671 _accum_task_vtime[i] += vtime;
672 }
674 double all_task_accum_vtime() {
675 double ret = 0.0;
676 for (uint i = 0; i < _max_worker_id; ++i)
677 ret += _accum_task_vtime[i];
678 return ret;
679 }
681 // Attempts to steal an object from the task queues of other tasks
682 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
683 return _task_queues->steal(worker_id, hash_seed, obj);
684 }
686 ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage);
687 ~ConcurrentMark();
689 ConcurrentMarkThread* cmThread() { return _cmThread; }
691 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
692 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
694 // Returns the number of GC threads to be used in a concurrent
695 // phase based on the number of GC threads being used in a STW
696 // phase.
697 uint scale_parallel_threads(uint n_par_threads);
699 // Calculates the number of GC threads to be used in a concurrent phase.
700 uint calc_parallel_marking_threads();
702 // The following three are interaction between CM and
703 // G1CollectedHeap
705 // This notifies CM that a root during initial-mark needs to be
706 // grayed. It is MT-safe. word_size is the size of the object in
707 // words. It is passed explicitly as sometimes we cannot calculate
708 // it from the given object because it might be in an inconsistent
709 // state (e.g., in to-space and being copied). So the caller is
710 // responsible for dealing with this issue (e.g., get the size from
711 // the from-space image when the to-space image might be
712 // inconsistent) and always passing the size. hr is the region that
713 // contains the object and it's passed optionally from callers who
714 // might already have it (no point in recalculating it).
715 inline void grayRoot(oop obj, size_t word_size,
716 uint worker_id, HeapRegion* hr = NULL);
718 // It iterates over the heap and for each object it comes across it
719 // will dump the contents of its reference fields, as well as
720 // liveness information for the object and its referents. The dump
721 // will be written to a file with the following name:
722 // G1PrintReachableBaseFile + "." + str.
723 // vo decides whether the prev (vo == UsePrevMarking), the next
724 // (vo == UseNextMarking) marking information, or the mark word
725 // (vo == UseMarkWord) will be used to determine the liveness of
726 // each object / referent.
727 // If all is true, all objects in the heap will be dumped, otherwise
728 // only the live ones. In the dump the following symbols / breviations
729 // are used:
730 // M : an explicitly live object (its bitmap bit is set)
731 // > : an implicitly live object (over tams)
732 // O : an object outside the G1 heap (typically: in the perm gen)
733 // NOT : a reference field whose referent is not live
734 // AND MARKED : indicates that an object is both explicitly and
735 // implicitly live (it should be one or the other, not both)
736 void print_reachable(const char* str,
737 VerifyOption vo, bool all) PRODUCT_RETURN;
739 // Clear the next marking bitmap (will be called concurrently).
740 void clearNextBitmap();
742 // Return whether the next mark bitmap has no marks set. To be used for assertions
743 // only. Will not yield to pause requests.
744 bool nextMarkBitmapIsClear();
746 // These two do the work that needs to be done before and after the
747 // initial root checkpoint. Since this checkpoint can be done at two
748 // different points (i.e. an explicit pause or piggy-backed on a
749 // young collection), then it's nice to be able to easily share the
750 // pre/post code. It might be the case that we can put everything in
751 // the post method. TP
752 void checkpointRootsInitialPre();
753 void checkpointRootsInitialPost();
755 // Scan all the root regions and mark everything reachable from
756 // them.
757 void scanRootRegions();
759 // Scan a single root region and mark everything reachable from it.
760 void scanRootRegion(HeapRegion* hr, uint worker_id);
762 // Do concurrent phase of marking, to a tentative transitive closure.
763 void markFromRoots();
765 void checkpointRootsFinal(bool clear_all_soft_refs);
766 void checkpointRootsFinalWork();
767 void cleanup();
768 void completeCleanup();
770 // Mark in the previous bitmap. NB: this is usually read-only, so use
771 // this carefully!
772 inline void markPrev(oop p);
774 // Clears marks for all objects in the given range, for the prev,
775 // next, or both bitmaps. NB: the previous bitmap is usually
776 // read-only, so use this carefully!
777 void clearRangePrevBitmap(MemRegion mr);
778 void clearRangeNextBitmap(MemRegion mr);
779 void clearRangeBothBitmaps(MemRegion mr);
781 // Notify data structures that a GC has started.
782 void note_start_of_gc() {
783 _markStack.note_start_of_gc();
784 }
786 // Notify data structures that a GC is finished.
787 void note_end_of_gc() {
788 _markStack.note_end_of_gc();
789 }
791 // Verify that there are no CSet oops on the stacks (taskqueues /
792 // global mark stack), enqueued SATB buffers, per-thread SATB
793 // buffers, and fingers (global / per-task). The boolean parameters
794 // decide which of the above data structures to verify. If marking
795 // is not in progress, it's a no-op.
796 void verify_no_cset_oops(bool verify_stacks,
797 bool verify_enqueued_buffers,
798 bool verify_thread_buffers,
799 bool verify_fingers) PRODUCT_RETURN;
801 bool isMarked(oop p) const {
802 assert(p != NULL && p->is_oop(), "expected an oop");
803 HeapWord* addr = (HeapWord*)p;
804 assert(addr >= _nextMarkBitMap->startWord() ||
805 addr < _nextMarkBitMap->endWord(), "in a region");
807 return _nextMarkBitMap->isMarked(addr);
808 }
810 inline bool not_yet_marked(oop p) const;
812 // XXX Debug code
813 bool containing_card_is_marked(void* p);
814 bool containing_cards_are_marked(void* start, void* last);
816 bool isPrevMarked(oop p) const {
817 assert(p != NULL && p->is_oop(), "expected an oop");
818 HeapWord* addr = (HeapWord*)p;
819 assert(addr >= _prevMarkBitMap->startWord() ||
820 addr < _prevMarkBitMap->endWord(), "in a region");
822 return _prevMarkBitMap->isMarked(addr);
823 }
825 inline bool do_yield_check(uint worker_i = 0);
827 // Called to abort the marking cycle after a Full GC takes palce.
828 void abort();
830 bool has_aborted() { return _has_aborted; }
832 const GCId& concurrent_gc_id();
834 // This prints the global/local fingers. It is used for debugging.
835 NOT_PRODUCT(void print_finger();)
837 void print_summary_info();
839 void print_worker_threads_on(outputStream* st) const;
841 void print_on_error(outputStream* st) const;
843 // The following indicate whether a given verbose level has been
844 // set. Notice that anything above stats is conditional to
845 // _MARKING_VERBOSE_ having been set to 1
846 bool verbose_stats() {
847 return _verbose_level >= stats_verbose;
848 }
849 bool verbose_low() {
850 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
851 }
852 bool verbose_medium() {
853 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
854 }
855 bool verbose_high() {
856 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
857 }
859 // Liveness counting
861 // Utility routine to set an exclusive range of cards on the given
862 // card liveness bitmap
863 inline void set_card_bitmap_range(BitMap* card_bm,
864 BitMap::idx_t start_idx,
865 BitMap::idx_t end_idx,
866 bool is_par);
868 // Returns the card number of the bottom of the G1 heap.
869 // Used in biasing indices into accounting card bitmaps.
870 intptr_t heap_bottom_card_num() const {
871 return _heap_bottom_card_num;
872 }
874 // Returns the card bitmap for a given task or worker id.
875 BitMap* count_card_bitmap_for(uint worker_id) {
876 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
877 assert(_count_card_bitmaps != NULL, "uninitialized");
878 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
879 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
880 return task_card_bm;
881 }
883 // Returns the array containing the marked bytes for each region,
884 // for the given worker or task id.
885 size_t* count_marked_bytes_array_for(uint worker_id) {
886 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
887 assert(_count_marked_bytes != NULL, "uninitialized");
888 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
889 assert(marked_bytes_array != NULL, "uninitialized");
890 return marked_bytes_array;
891 }
893 // Returns the index in the liveness accounting card table bitmap
894 // for the given address
895 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
897 // Counts the size of the given memory region in the the given
898 // marked_bytes array slot for the given HeapRegion.
899 // Sets the bits in the given card bitmap that are associated with the
900 // cards that are spanned by the memory region.
901 inline void count_region(MemRegion mr, HeapRegion* hr,
902 size_t* marked_bytes_array,
903 BitMap* task_card_bm);
905 // Counts the given memory region in the task/worker counting
906 // data structures for the given worker id.
907 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
909 // Counts the given memory region in the task/worker counting
910 // data structures for the given worker id.
911 inline void count_region(MemRegion mr, uint worker_id);
913 // Counts the given object in the given task/worker counting
914 // data structures.
915 inline void count_object(oop obj, HeapRegion* hr,
916 size_t* marked_bytes_array,
917 BitMap* task_card_bm);
919 // Counts the given object in the task/worker counting data
920 // structures for the given worker id.
921 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
923 // Attempts to mark the given object and, if successful, counts
924 // the object in the given task/worker counting structures.
925 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
926 size_t* marked_bytes_array,
927 BitMap* task_card_bm);
929 // Attempts to mark the given object and, if successful, counts
930 // the object in the task/worker counting structures for the
931 // given worker id.
932 inline bool par_mark_and_count(oop obj, size_t word_size,
933 HeapRegion* hr, uint worker_id);
935 // Attempts to mark the given object and, if successful, counts
936 // the object in the task/worker counting structures for the
937 // given worker id.
938 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
940 // Similar to the above routine but we don't know the heap region that
941 // contains the object to be marked/counted, which this routine looks up.
942 inline bool par_mark_and_count(oop obj, uint worker_id);
944 // Similar to the above routine but there are times when we cannot
945 // safely calculate the size of obj due to races and we, therefore,
946 // pass the size in as a parameter. It is the caller's reponsibility
947 // to ensure that the size passed in for obj is valid.
948 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
950 // Unconditionally mark the given object, and unconditinally count
951 // the object in the counting structures for worker id 0.
952 // Should *not* be called from parallel code.
953 inline bool mark_and_count(oop obj, HeapRegion* hr);
955 // Similar to the above routine but we don't know the heap region that
956 // contains the object to be marked/counted, which this routine looks up.
957 // Should *not* be called from parallel code.
958 inline bool mark_and_count(oop obj);
960 // Returns true if initialization was successfully completed.
961 bool completed_initialization() const {
962 return _completed_initialization;
963 }
965 protected:
966 // Clear all the per-task bitmaps and arrays used to store the
967 // counting data.
968 void clear_all_count_data();
970 // Aggregates the counting data for each worker/task
971 // that was constructed while marking. Also sets
972 // the amount of marked bytes for each region and
973 // the top at concurrent mark count.
974 void aggregate_count_data();
976 // Verification routine
977 void verify_count_data();
978 };
980 // A class representing a marking task.
981 class CMTask : public TerminatorTerminator {
982 private:
983 enum PrivateConstants {
984 // the regular clock call is called once the scanned words reaches
985 // this limit
986 words_scanned_period = 12*1024,
987 // the regular clock call is called once the number of visited
988 // references reaches this limit
989 refs_reached_period = 384,
990 // initial value for the hash seed, used in the work stealing code
991 init_hash_seed = 17,
992 // how many entries will be transferred between global stack and
993 // local queues
994 global_stack_transfer_size = 16
995 };
997 uint _worker_id;
998 G1CollectedHeap* _g1h;
999 ConcurrentMark* _cm;
1000 CMBitMap* _nextMarkBitMap;
1001 // the task queue of this task
1002 CMTaskQueue* _task_queue;
1003 private:
1004 // the task queue set---needed for stealing
1005 CMTaskQueueSet* _task_queues;
1006 // indicates whether the task has been claimed---this is only for
1007 // debugging purposes
1008 bool _claimed;
1010 // number of calls to this task
1011 int _calls;
1013 // when the virtual timer reaches this time, the marking step should
1014 // exit
1015 double _time_target_ms;
1016 // the start time of the current marking step
1017 double _start_time_ms;
1019 // the oop closure used for iterations over oops
1020 G1CMOopClosure* _cm_oop_closure;
1022 // the region this task is scanning, NULL if we're not scanning any
1023 HeapRegion* _curr_region;
1024 // the local finger of this task, NULL if we're not scanning a region
1025 HeapWord* _finger;
1026 // limit of the region this task is scanning, NULL if we're not scanning one
1027 HeapWord* _region_limit;
1029 // the number of words this task has scanned
1030 size_t _words_scanned;
1031 // When _words_scanned reaches this limit, the regular clock is
1032 // called. Notice that this might be decreased under certain
1033 // circumstances (i.e. when we believe that we did an expensive
1034 // operation).
1035 size_t _words_scanned_limit;
1036 // the initial value of _words_scanned_limit (i.e. what it was
1037 // before it was decreased).
1038 size_t _real_words_scanned_limit;
1040 // the number of references this task has visited
1041 size_t _refs_reached;
1042 // When _refs_reached reaches this limit, the regular clock is
1043 // called. Notice this this might be decreased under certain
1044 // circumstances (i.e. when we believe that we did an expensive
1045 // operation).
1046 size_t _refs_reached_limit;
1047 // the initial value of _refs_reached_limit (i.e. what it was before
1048 // it was decreased).
1049 size_t _real_refs_reached_limit;
1051 // used by the work stealing stuff
1052 int _hash_seed;
1053 // if this is true, then the task has aborted for some reason
1054 bool _has_aborted;
1055 // set when the task aborts because it has met its time quota
1056 bool _has_timed_out;
1057 // true when we're draining SATB buffers; this avoids the task
1058 // aborting due to SATB buffers being available (as we're already
1059 // dealing with them)
1060 bool _draining_satb_buffers;
1062 // number sequence of past step times
1063 NumberSeq _step_times_ms;
1064 // elapsed time of this task
1065 double _elapsed_time_ms;
1066 // termination time of this task
1067 double _termination_time_ms;
1068 // when this task got into the termination protocol
1069 double _termination_start_time_ms;
1071 // true when the task is during a concurrent phase, false when it is
1072 // in the remark phase (so, in the latter case, we do not have to
1073 // check all the things that we have to check during the concurrent
1074 // phase, i.e. SATB buffer availability...)
1075 bool _concurrent;
1077 TruncatedSeq _marking_step_diffs_ms;
1079 // Counting data structures. Embedding the task's marked_bytes_array
1080 // and card bitmap into the actual task saves having to go through
1081 // the ConcurrentMark object.
1082 size_t* _marked_bytes_array;
1083 BitMap* _card_bm;
1085 // LOTS of statistics related with this task
1086 #if _MARKING_STATS_
1087 NumberSeq _all_clock_intervals_ms;
1088 double _interval_start_time_ms;
1090 int _aborted;
1091 int _aborted_overflow;
1092 int _aborted_cm_aborted;
1093 int _aborted_yield;
1094 int _aborted_timed_out;
1095 int _aborted_satb;
1096 int _aborted_termination;
1098 int _steal_attempts;
1099 int _steals;
1101 int _clock_due_to_marking;
1102 int _clock_due_to_scanning;
1104 int _local_pushes;
1105 int _local_pops;
1106 int _local_max_size;
1107 int _objs_scanned;
1109 int _global_pushes;
1110 int _global_pops;
1111 int _global_max_size;
1113 int _global_transfers_to;
1114 int _global_transfers_from;
1116 int _regions_claimed;
1117 int _objs_found_on_bitmap;
1119 int _satb_buffers_processed;
1120 #endif // _MARKING_STATS_
1122 // it updates the local fields after this task has claimed
1123 // a new region to scan
1124 void setup_for_region(HeapRegion* hr);
1125 // it brings up-to-date the limit of the region
1126 void update_region_limit();
1128 // called when either the words scanned or the refs visited limit
1129 // has been reached
1130 void reached_limit();
1131 // recalculates the words scanned and refs visited limits
1132 void recalculate_limits();
1133 // decreases the words scanned and refs visited limits when we reach
1134 // an expensive operation
1135 void decrease_limits();
1136 // it checks whether the words scanned or refs visited reached their
1137 // respective limit and calls reached_limit() if they have
1138 void check_limits() {
1139 if (_words_scanned >= _words_scanned_limit ||
1140 _refs_reached >= _refs_reached_limit) {
1141 reached_limit();
1142 }
1143 }
1144 // this is supposed to be called regularly during a marking step as
1145 // it checks a bunch of conditions that might cause the marking step
1146 // to abort
1147 void regular_clock_call();
1148 bool concurrent() { return _concurrent; }
1150 public:
1151 // It resets the task; it should be called right at the beginning of
1152 // a marking phase.
1153 void reset(CMBitMap* _nextMarkBitMap);
1154 // it clears all the fields that correspond to a claimed region.
1155 void clear_region_fields();
1157 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1159 // The main method of this class which performs a marking step
1160 // trying not to exceed the given duration. However, it might exit
1161 // prematurely, according to some conditions (i.e. SATB buffers are
1162 // available for processing).
1163 void do_marking_step(double target_ms,
1164 bool do_termination,
1165 bool is_serial);
1167 // These two calls start and stop the timer
1168 void record_start_time() {
1169 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1170 }
1171 void record_end_time() {
1172 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1173 }
1175 // returns the worker ID associated with this task.
1176 uint worker_id() { return _worker_id; }
1178 // From TerminatorTerminator. It determines whether this task should
1179 // exit the termination protocol after it's entered it.
1180 virtual bool should_exit_termination();
1182 // Resets the local region fields after a task has finished scanning a
1183 // region; or when they have become stale as a result of the region
1184 // being evacuated.
1185 void giveup_current_region();
1187 HeapWord* finger() { return _finger; }
1189 bool has_aborted() { return _has_aborted; }
1190 void set_has_aborted() { _has_aborted = true; }
1191 void clear_has_aborted() { _has_aborted = false; }
1192 bool has_timed_out() { return _has_timed_out; }
1193 bool claimed() { return _claimed; }
1195 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1197 // It grays the object by marking it and, if necessary, pushing it
1198 // on the local queue
1199 inline void deal_with_reference(oop obj);
1201 // It scans an object and visits its children.
1202 void scan_object(oop obj);
1204 // It pushes an object on the local queue.
1205 inline void push(oop obj);
1207 // These two move entries to/from the global stack.
1208 void move_entries_to_global_stack();
1209 void get_entries_from_global_stack();
1211 // It pops and scans objects from the local queue. If partially is
1212 // true, then it stops when the queue size is of a given limit. If
1213 // partially is false, then it stops when the queue is empty.
1214 void drain_local_queue(bool partially);
1215 // It moves entries from the global stack to the local queue and
1216 // drains the local queue. If partially is true, then it stops when
1217 // both the global stack and the local queue reach a given size. If
1218 // partially if false, it tries to empty them totally.
1219 void drain_global_stack(bool partially);
1220 // It keeps picking SATB buffers and processing them until no SATB
1221 // buffers are available.
1222 void drain_satb_buffers();
1224 // moves the local finger to a new location
1225 inline void move_finger_to(HeapWord* new_finger) {
1226 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1227 _finger = new_finger;
1228 }
1230 CMTask(uint worker_id, ConcurrentMark *cm,
1231 size_t* marked_bytes, BitMap* card_bm,
1232 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1234 // it prints statistics associated with this task
1235 void print_stats();
1237 #if _MARKING_STATS_
1238 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1239 #endif // _MARKING_STATS_
1240 };
1242 // Class that's used to to print out per-region liveness
1243 // information. It's currently used at the end of marking and also
1244 // after we sort the old regions at the end of the cleanup operation.
1245 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1246 private:
1247 outputStream* _out;
1249 // Accumulators for these values.
1250 size_t _total_used_bytes;
1251 size_t _total_capacity_bytes;
1252 size_t _total_prev_live_bytes;
1253 size_t _total_next_live_bytes;
1255 // These are set up when we come across a "stars humongous" region
1256 // (as this is where most of this information is stored, not in the
1257 // subsequent "continues humongous" regions). After that, for every
1258 // region in a given humongous region series we deduce the right
1259 // values for it by simply subtracting the appropriate amount from
1260 // these fields. All these values should reach 0 after we've visited
1261 // the last region in the series.
1262 size_t _hum_used_bytes;
1263 size_t _hum_capacity_bytes;
1264 size_t _hum_prev_live_bytes;
1265 size_t _hum_next_live_bytes;
1267 // Accumulator for the remembered set size
1268 size_t _total_remset_bytes;
1270 // Accumulator for strong code roots memory size
1271 size_t _total_strong_code_roots_bytes;
1273 static double perc(size_t val, size_t total) {
1274 if (total == 0) {
1275 return 0.0;
1276 } else {
1277 return 100.0 * ((double) val / (double) total);
1278 }
1279 }
1281 static double bytes_to_mb(size_t val) {
1282 return (double) val / (double) M;
1283 }
1285 // See the .cpp file.
1286 size_t get_hum_bytes(size_t* hum_bytes);
1287 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1288 size_t* prev_live_bytes, size_t* next_live_bytes);
1290 public:
1291 // The header and footer are printed in the constructor and
1292 // destructor respectively.
1293 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1294 virtual bool doHeapRegion(HeapRegion* r);
1295 ~G1PrintRegionLivenessInfoClosure();
1296 };
1298 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP