Wed, 10 Apr 2013 14:26:49 +0200
8011872: Include Bit Map addresses in the hs_err files
Reviewed-by: brutisso, jmasa
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 "gc_implementation/g1/heapRegionSets.hpp"
29 #include "utilities/taskqueue.hpp"
31 class G1CollectedHeap;
32 class CMTask;
33 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
34 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
36 // Closure used by CM during concurrent reference discovery
37 // and reference processing (during remarking) to determine
38 // if a particular object is alive. It is primarily used
39 // to determine if referents of discovered reference objects
40 // are alive. An instance is also embedded into the
41 // reference processor as the _is_alive_non_header field
42 class G1CMIsAliveClosure: public BoolObjectClosure {
43 G1CollectedHeap* _g1;
44 public:
45 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
47 void do_object(oop obj) {
48 ShouldNotCallThis();
49 }
50 bool do_object_b(oop obj);
51 };
53 // A generic CM bit map. This is essentially a wrapper around the BitMap
54 // class, with one bit per (1<<_shifter) HeapWords.
56 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
57 protected:
58 HeapWord* _bmStartWord; // base address of range covered by map
59 size_t _bmWordSize; // map size (in #HeapWords covered)
60 const int _shifter; // map to char or bit
61 VirtualSpace _virtual_space; // underlying the bit map
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(HeapWord* addr,
92 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(HeapWord* addr,
97 HeapWord* limit = NULL) const;
99 // conversion utilities
100 HeapWord* offsetToHeapWord(size_t offset) const {
101 return _bmStartWord + (offset << _shifter);
102 }
103 size_t heapWordToOffset(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(ReservedSpace rs) const;)
120 };
122 class CMBitMap : public CMBitMapRO {
124 public:
125 // constructor
126 CMBitMap(int shifter) :
127 CMBitMapRO(shifter) {}
129 // Allocates the back store for the marking bitmap
130 bool allocate(ReservedSpace heap_rs);
132 // write marks
133 void mark(HeapWord* addr) {
134 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
135 "outside underlying space?");
136 _bm.set_bit(heapWordToOffset(addr));
137 }
138 void clear(HeapWord* addr) {
139 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
140 "outside underlying space?");
141 _bm.clear_bit(heapWordToOffset(addr));
142 }
143 bool parMark(HeapWord* addr) {
144 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
145 "outside underlying space?");
146 return _bm.par_set_bit(heapWordToOffset(addr));
147 }
148 bool parClear(HeapWord* addr) {
149 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
150 "outside underlying space?");
151 return _bm.par_clear_bit(heapWordToOffset(addr));
152 }
153 void markRange(MemRegion mr);
154 void clearAll();
155 void clearRange(MemRegion mr);
157 // Starting at the bit corresponding to "addr" (inclusive), find the next
158 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
159 // the end of this run (stopping at "end_addr"). Return the MemRegion
160 // covering from the start of the region corresponding to the first bit
161 // of the run to the end of the region corresponding to the last bit of
162 // the run. If there is no "1" bit at or after "addr", return an empty
163 // MemRegion.
164 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
165 };
167 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
168 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
169 VirtualSpace _virtual_space; // Underlying backing store for actual stack
170 ConcurrentMark* _cm;
171 oop* _base; // bottom of stack
172 jint _index; // one more than last occupied index
173 jint _capacity; // max #elements
174 jint _saved_index; // value of _index saved at start of GC
175 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
177 bool _overflow;
178 bool _should_expand;
179 DEBUG_ONLY(bool _drain_in_progress;)
180 DEBUG_ONLY(bool _drain_in_progress_yields;)
182 public:
183 CMMarkStack(ConcurrentMark* cm);
184 ~CMMarkStack();
186 #ifndef PRODUCT
187 jint max_depth() const {
188 return _max_depth;
189 }
190 #endif
192 bool allocate(size_t capacity);
194 oop pop() {
195 if (!isEmpty()) {
196 return _base[--_index] ;
197 }
198 return NULL;
199 }
201 // If overflow happens, don't do the push, and record the overflow.
202 // *Requires* that "ptr" is already marked.
203 void push(oop ptr) {
204 if (isFull()) {
205 // Record overflow.
206 _overflow = true;
207 return;
208 } else {
209 _base[_index++] = ptr;
210 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
211 }
212 }
213 // Non-block impl. Note: concurrency is allowed only with other
214 // "par_push" operations, not with "pop" or "drain". We would need
215 // parallel versions of them if such concurrency was desired.
216 void par_push(oop ptr);
218 // Pushes the first "n" elements of "ptr_arr" on the stack.
219 // Non-block impl. Note: concurrency is allowed only with other
220 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
221 void par_adjoin_arr(oop* ptr_arr, int n);
223 // Pushes the first "n" elements of "ptr_arr" on the stack.
224 // Locking impl: concurrency is allowed only with
225 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
226 // locking strategy.
227 void par_push_arr(oop* ptr_arr, int n);
229 // If returns false, the array was empty. Otherwise, removes up to "max"
230 // elements from the stack, and transfers them to "ptr_arr" in an
231 // unspecified order. The actual number transferred is given in "n" ("n
232 // == 0" is deliberately redundant with the return value.) Locking impl:
233 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
234 // operations, which use the same locking strategy.
235 bool par_pop_arr(oop* ptr_arr, int max, int* n);
237 // Drain the mark stack, applying the given closure to all fields of
238 // objects on the stack. (That is, continue until the stack is empty,
239 // even if closure applications add entries to the stack.) The "bm"
240 // argument, if non-null, may be used to verify that only marked objects
241 // are on the mark stack. If "yield_after" is "true", then the
242 // concurrent marker performing the drain offers to yield after
243 // processing each object. If a yield occurs, stops the drain operation
244 // and returns false. Otherwise, returns true.
245 template<class OopClosureClass>
246 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
248 bool isEmpty() { return _index == 0; }
249 bool isFull() { return _index == _capacity; }
250 int maxElems() { return _capacity; }
252 bool overflow() { return _overflow; }
253 void clear_overflow() { _overflow = false; }
255 bool should_expand() const { return _should_expand; }
256 void set_should_expand();
258 // Expand the stack, typically in response to an overflow condition
259 void expand();
261 int size() { return _index; }
263 void setEmpty() { _index = 0; clear_overflow(); }
265 // Record the current index.
266 void note_start_of_gc();
268 // Make sure that we have not added any entries to the stack during GC.
269 void note_end_of_gc();
271 // iterate over the oops in the mark stack, up to the bound recorded via
272 // the call above.
273 void oops_do(OopClosure* f);
274 };
276 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
277 private:
278 #ifndef PRODUCT
279 uintx _num_remaining;
280 bool _force;
281 #endif // !defined(PRODUCT)
283 public:
284 void init() PRODUCT_RETURN;
285 void update() PRODUCT_RETURN;
286 bool should_force() PRODUCT_RETURN_( return false; );
287 };
289 // this will enable a variety of different statistics per GC task
290 #define _MARKING_STATS_ 0
291 // this will enable the higher verbose levels
292 #define _MARKING_VERBOSE_ 0
294 #if _MARKING_STATS_
295 #define statsOnly(statement) \
296 do { \
297 statement ; \
298 } while (0)
299 #else // _MARKING_STATS_
300 #define statsOnly(statement) \
301 do { \
302 } while (0)
303 #endif // _MARKING_STATS_
305 typedef enum {
306 no_verbose = 0, // verbose turned off
307 stats_verbose, // only prints stats at the end of marking
308 low_verbose, // low verbose, mostly per region and per major event
309 medium_verbose, // a bit more detailed than low
310 high_verbose // per object verbose
311 } CMVerboseLevel;
313 class YoungList;
315 // Root Regions are regions that are not empty at the beginning of a
316 // marking cycle and which we might collect during an evacuation pause
317 // while the cycle is active. Given that, during evacuation pauses, we
318 // do not copy objects that are explicitly marked, what we have to do
319 // for the root regions is to scan them and mark all objects reachable
320 // from them. According to the SATB assumptions, we only need to visit
321 // each object once during marking. So, as long as we finish this scan
322 // before the next evacuation pause, we can copy the objects from the
323 // root regions without having to mark them or do anything else to them.
324 //
325 // Currently, we only support root region scanning once (at the start
326 // of the marking cycle) and the root regions are all the survivor
327 // regions populated during the initial-mark pause.
328 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
329 private:
330 YoungList* _young_list;
331 ConcurrentMark* _cm;
333 volatile bool _scan_in_progress;
334 volatile bool _should_abort;
335 HeapRegion* volatile _next_survivor;
337 public:
338 CMRootRegions();
339 // We actually do most of the initialization in this method.
340 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
342 // Reset the claiming / scanning of the root regions.
343 void prepare_for_scan();
345 // Forces get_next() to return NULL so that the iteration aborts early.
346 void abort() { _should_abort = true; }
348 // Return true if the CM thread are actively scanning root regions,
349 // false otherwise.
350 bool scan_in_progress() { return _scan_in_progress; }
352 // Claim the next root region to scan atomically, or return NULL if
353 // all have been claimed.
354 HeapRegion* claim_next();
356 // Flag that we're done with root region scanning and notify anyone
357 // who's waiting on it. If aborted is false, assume that all regions
358 // have been claimed.
359 void scan_finished();
361 // If CM threads are still scanning root regions, wait until they
362 // are done. Return true if we had to wait, false otherwise.
363 bool wait_until_scan_finished();
364 };
366 class ConcurrentMarkThread;
368 class ConcurrentMark: public CHeapObj<mtGC> {
369 friend class CMMarkStack;
370 friend class ConcurrentMarkThread;
371 friend class CMTask;
372 friend class CMBitMapClosure;
373 friend class CMGlobalObjectClosure;
374 friend class CMRemarkTask;
375 friend class CMConcurrentMarkingTask;
376 friend class G1ParNoteEndTask;
377 friend class CalcLiveObjectsClosure;
378 friend class G1CMRefProcTaskProxy;
379 friend class G1CMRefProcTaskExecutor;
380 friend class G1CMKeepAliveAndDrainClosure;
381 friend class G1CMDrainMarkingStackClosure;
383 protected:
384 ConcurrentMarkThread* _cmThread; // the thread doing the work
385 G1CollectedHeap* _g1h; // the heap.
386 uint _parallel_marking_threads; // the number of marking
387 // threads we're use
388 uint _max_parallel_marking_threads; // max number of marking
389 // threads we'll ever use
390 double _sleep_factor; // how much we have to sleep, with
391 // respect to the work we just did, to
392 // meet the marking overhead goal
393 double _marking_task_overhead; // marking target overhead for
394 // a single task
396 // same as the two above, but for the cleanup task
397 double _cleanup_sleep_factor;
398 double _cleanup_task_overhead;
400 FreeRegionList _cleanup_list;
402 // Concurrent marking support structures
403 CMBitMap _markBitMap1;
404 CMBitMap _markBitMap2;
405 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
406 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
408 BitMap _region_bm;
409 BitMap _card_bm;
411 // Heap bounds
412 HeapWord* _heap_start;
413 HeapWord* _heap_end;
415 // Root region tracking and claiming.
416 CMRootRegions _root_regions;
418 // For gray objects
419 CMMarkStack _markStack; // Grey objects behind global finger.
420 HeapWord* volatile _finger; // the global finger, region aligned,
421 // always points to the end of the
422 // last claimed region
424 // marking tasks
425 uint _max_worker_id;// maximum worker id
426 uint _active_tasks; // task num currently active
427 CMTask** _tasks; // task queue array (max_worker_id len)
428 CMTaskQueueSet* _task_queues; // task queue set
429 ParallelTaskTerminator _terminator; // for termination
431 // Two sync barriers that are used to synchronise tasks when an
432 // overflow occurs. The algorithm is the following. All tasks enter
433 // the first one to ensure that they have all stopped manipulating
434 // the global data structures. After they exit it, they re-initialise
435 // their data structures and task 0 re-initialises the global data
436 // structures. Then, they enter the second sync barrier. This
437 // ensure, that no task starts doing work before all data
438 // structures (local and global) have been re-initialised. When they
439 // exit it, they are free to start working again.
440 WorkGangBarrierSync _first_overflow_barrier_sync;
441 WorkGangBarrierSync _second_overflow_barrier_sync;
443 // this is set by any task, when an overflow on the global data
444 // structures is detected.
445 volatile bool _has_overflown;
446 // true: marking is concurrent, false: we're in remark
447 volatile bool _concurrent;
448 // set at the end of a Full GC so that marking aborts
449 volatile bool _has_aborted;
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;
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;
461 // verbose level
462 CMVerboseLevel _verbose_level;
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;
473 double* _accum_task_vtime; // accumulated task vtime
475 FlexibleWorkGang* _parallel_workers;
477 ForceOverflowSettings _force_overflow_conc;
478 ForceOverflowSettings _force_overflow_stw;
480 void weakRefsWork(bool clear_all_soft_refs);
482 void swapMarkBitMaps();
484 // It resets the global marking data structures, as well as the
485 // task local ones; should be called during initial mark.
486 void reset();
488 // Resets all the marking data structures. Called when we have to restart
489 // marking or when marking completes (via set_non_marking_state below).
490 void reset_marking_state(bool clear_overflow = true);
492 // We do this after we're done with marking so that the marking data
493 // structures are initialised to a sensible and predictable state.
494 void set_non_marking_state();
496 // Called to indicate how many threads are currently active.
497 void set_concurrency(uint active_tasks);
499 // It should be called to indicate which phase we're in (concurrent
500 // mark or remark) and how many threads are currently active.
501 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
503 // prints all gathered CM-related statistics
504 void print_stats();
506 bool cleanup_list_is_empty() {
507 return _cleanup_list.is_empty();
508 }
510 // accessor methods
511 uint parallel_marking_threads() const { return _parallel_marking_threads; }
512 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
513 double sleep_factor() { return _sleep_factor; }
514 double marking_task_overhead() { return _marking_task_overhead;}
515 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
516 double cleanup_task_overhead() { return _cleanup_task_overhead;}
518 bool use_parallel_marking_threads() const {
519 assert(parallel_marking_threads() <=
520 max_parallel_marking_threads(), "sanity");
521 assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
522 parallel_marking_threads() > 0,
523 "parallel workers not set up correctly");
524 return _parallel_workers != NULL;
525 }
527 HeapWord* finger() { return _finger; }
528 bool concurrent() { return _concurrent; }
529 uint active_tasks() { return _active_tasks; }
530 ParallelTaskTerminator* terminator() { return &_terminator; }
532 // It claims the next available region to be scanned by a marking
533 // task/thread. It might return NULL if the next region is empty or
534 // we have run out of regions. In the latter case, out_of_regions()
535 // determines whether we've really run out of regions or the task
536 // should call claim_region() again. This might seem a bit
537 // awkward. Originally, the code was written so that claim_region()
538 // either successfully returned with a non-empty region or there
539 // were no more regions to be claimed. The problem with this was
540 // that, in certain circumstances, it iterated over large chunks of
541 // the heap finding only empty regions and, while it was working, it
542 // was preventing the calling task to call its regular clock
543 // method. So, this way, each task will spend very little time in
544 // claim_region() and is allowed to call the regular clock method
545 // frequently.
546 HeapRegion* claim_region(uint worker_id);
548 // It determines whether we've run out of regions to scan.
549 bool out_of_regions() { return _finger == _heap_end; }
551 // Returns the task with the given id
552 CMTask* task(int id) {
553 assert(0 <= id && id < (int) _active_tasks,
554 "task id not within active bounds");
555 return _tasks[id];
556 }
558 // Returns the task queue with the given id
559 CMTaskQueue* task_queue(int id) {
560 assert(0 <= id && id < (int) _active_tasks,
561 "task queue id not within active bounds");
562 return (CMTaskQueue*) _task_queues->queue(id);
563 }
565 // Returns the task queue set
566 CMTaskQueueSet* task_queues() { return _task_queues; }
568 // Access / manipulation of the overflow flag which is set to
569 // indicate that the global stack has overflown
570 bool has_overflown() { return _has_overflown; }
571 void set_has_overflown() { _has_overflown = true; }
572 void clear_has_overflown() { _has_overflown = false; }
573 bool restart_for_overflow() { return _restart_for_overflow; }
575 bool has_aborted() { return _has_aborted; }
577 // Methods to enter the two overflow sync barriers
578 void enter_first_sync_barrier(uint worker_id);
579 void enter_second_sync_barrier(uint worker_id);
581 ForceOverflowSettings* force_overflow_conc() {
582 return &_force_overflow_conc;
583 }
585 ForceOverflowSettings* force_overflow_stw() {
586 return &_force_overflow_stw;
587 }
589 ForceOverflowSettings* force_overflow() {
590 if (concurrent()) {
591 return force_overflow_conc();
592 } else {
593 return force_overflow_stw();
594 }
595 }
597 // Live Data Counting data structures...
598 // These data structures are initialized at the start of
599 // marking. They are written to while marking is active.
600 // They are aggregated during remark; the aggregated values
601 // are then used to populate the _region_bm, _card_bm, and
602 // the total live bytes, which are then subsequently updated
603 // during cleanup.
605 // An array of bitmaps (one bit map per task). Each bitmap
606 // is used to record the cards spanned by the live objects
607 // marked by that task/worker.
608 BitMap* _count_card_bitmaps;
610 // Used to record the number of marked live bytes
611 // (for each region, by worker thread).
612 size_t** _count_marked_bytes;
614 // Card index of the bottom of the G1 heap. Used for biasing indices into
615 // the card bitmaps.
616 intptr_t _heap_bottom_card_num;
618 // Set to true when initialization is complete
619 bool _completed_initialization;
621 public:
622 // Manipulation of the global mark stack.
623 // Notice that the first mark_stack_push is CAS-based, whereas the
624 // two below are Mutex-based. This is OK since the first one is only
625 // called during evacuation pauses and doesn't compete with the
626 // other two (which are called by the marking tasks during
627 // concurrent marking or remark).
628 bool mark_stack_push(oop p) {
629 _markStack.par_push(p);
630 if (_markStack.overflow()) {
631 set_has_overflown();
632 return false;
633 }
634 return true;
635 }
636 bool mark_stack_push(oop* arr, int n) {
637 _markStack.par_push_arr(arr, n);
638 if (_markStack.overflow()) {
639 set_has_overflown();
640 return false;
641 }
642 return true;
643 }
644 void mark_stack_pop(oop* arr, int max, int* n) {
645 _markStack.par_pop_arr(arr, max, n);
646 }
647 size_t mark_stack_size() { return _markStack.size(); }
648 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
649 bool mark_stack_overflow() { return _markStack.overflow(); }
650 bool mark_stack_empty() { return _markStack.isEmpty(); }
652 CMRootRegions* root_regions() { return &_root_regions; }
654 bool concurrent_marking_in_progress() {
655 return _concurrent_marking_in_progress;
656 }
657 void set_concurrent_marking_in_progress() {
658 _concurrent_marking_in_progress = true;
659 }
660 void clear_concurrent_marking_in_progress() {
661 _concurrent_marking_in_progress = false;
662 }
664 void update_accum_task_vtime(int i, double vtime) {
665 _accum_task_vtime[i] += vtime;
666 }
668 double all_task_accum_vtime() {
669 double ret = 0.0;
670 for (uint i = 0; i < _max_worker_id; ++i)
671 ret += _accum_task_vtime[i];
672 return ret;
673 }
675 // Attempts to steal an object from the task queues of other tasks
676 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
677 return _task_queues->steal(worker_id, hash_seed, obj);
678 }
680 ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
681 ~ConcurrentMark();
683 ConcurrentMarkThread* cmThread() { return _cmThread; }
685 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
686 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
688 // Returns the number of GC threads to be used in a concurrent
689 // phase based on the number of GC threads being used in a STW
690 // phase.
691 uint scale_parallel_threads(uint n_par_threads);
693 // Calculates the number of GC threads to be used in a concurrent phase.
694 uint calc_parallel_marking_threads();
696 // The following three are interaction between CM and
697 // G1CollectedHeap
699 // This notifies CM that a root during initial-mark needs to be
700 // grayed. It is MT-safe. word_size is the size of the object in
701 // words. It is passed explicitly as sometimes we cannot calculate
702 // it from the given object because it might be in an inconsistent
703 // state (e.g., in to-space and being copied). So the caller is
704 // responsible for dealing with this issue (e.g., get the size from
705 // the from-space image when the to-space image might be
706 // inconsistent) and always passing the size. hr is the region that
707 // contains the object and it's passed optionally from callers who
708 // might already have it (no point in recalculating it).
709 inline void grayRoot(oop obj, size_t word_size,
710 uint worker_id, HeapRegion* hr = NULL);
712 // It iterates over the heap and for each object it comes across it
713 // will dump the contents of its reference fields, as well as
714 // liveness information for the object and its referents. The dump
715 // will be written to a file with the following name:
716 // G1PrintReachableBaseFile + "." + str.
717 // vo decides whether the prev (vo == UsePrevMarking), the next
718 // (vo == UseNextMarking) marking information, or the mark word
719 // (vo == UseMarkWord) will be used to determine the liveness of
720 // each object / referent.
721 // If all is true, all objects in the heap will be dumped, otherwise
722 // only the live ones. In the dump the following symbols / breviations
723 // are used:
724 // M : an explicitly live object (its bitmap bit is set)
725 // > : an implicitly live object (over tams)
726 // O : an object outside the G1 heap (typically: in the perm gen)
727 // NOT : a reference field whose referent is not live
728 // AND MARKED : indicates that an object is both explicitly and
729 // implicitly live (it should be one or the other, not both)
730 void print_reachable(const char* str,
731 VerifyOption vo, bool all) PRODUCT_RETURN;
733 // Clear the next marking bitmap (will be called concurrently).
734 void clearNextBitmap();
736 // These two do the work that needs to be done before and after the
737 // initial root checkpoint. Since this checkpoint can be done at two
738 // different points (i.e. an explicit pause or piggy-backed on a
739 // young collection), then it's nice to be able to easily share the
740 // pre/post code. It might be the case that we can put everything in
741 // the post method. TP
742 void checkpointRootsInitialPre();
743 void checkpointRootsInitialPost();
745 // Scan all the root regions and mark everything reachable from
746 // them.
747 void scanRootRegions();
749 // Scan a single root region and mark everything reachable from it.
750 void scanRootRegion(HeapRegion* hr, uint worker_id);
752 // Do concurrent phase of marking, to a tentative transitive closure.
753 void markFromRoots();
755 void checkpointRootsFinal(bool clear_all_soft_refs);
756 void checkpointRootsFinalWork();
757 void cleanup();
758 void completeCleanup();
760 // Mark in the previous bitmap. NB: this is usually read-only, so use
761 // this carefully!
762 inline void markPrev(oop p);
764 // Clears marks for all objects in the given range, for the prev,
765 // next, or both bitmaps. NB: the previous bitmap is usually
766 // read-only, so use this carefully!
767 void clearRangePrevBitmap(MemRegion mr);
768 void clearRangeNextBitmap(MemRegion mr);
769 void clearRangeBothBitmaps(MemRegion mr);
771 // Notify data structures that a GC has started.
772 void note_start_of_gc() {
773 _markStack.note_start_of_gc();
774 }
776 // Notify data structures that a GC is finished.
777 void note_end_of_gc() {
778 _markStack.note_end_of_gc();
779 }
781 // Verify that there are no CSet oops on the stacks (taskqueues /
782 // global mark stack), enqueued SATB buffers, per-thread SATB
783 // buffers, and fingers (global / per-task). The boolean parameters
784 // decide which of the above data structures to verify. If marking
785 // is not in progress, it's a no-op.
786 void verify_no_cset_oops(bool verify_stacks,
787 bool verify_enqueued_buffers,
788 bool verify_thread_buffers,
789 bool verify_fingers) PRODUCT_RETURN;
791 // It is called at the end of an evacuation pause during marking so
792 // that CM is notified of where the new end of the heap is. It
793 // doesn't do anything if concurrent_marking_in_progress() is false,
794 // unless the force parameter is true.
795 void update_g1_committed(bool force = false);
797 bool isMarked(oop p) const {
798 assert(p != NULL && p->is_oop(), "expected an oop");
799 HeapWord* addr = (HeapWord*)p;
800 assert(addr >= _nextMarkBitMap->startWord() ||
801 addr < _nextMarkBitMap->endWord(), "in a region");
803 return _nextMarkBitMap->isMarked(addr);
804 }
806 inline bool not_yet_marked(oop p) const;
808 // XXX Debug code
809 bool containing_card_is_marked(void* p);
810 bool containing_cards_are_marked(void* start, void* last);
812 bool isPrevMarked(oop p) const {
813 assert(p != NULL && p->is_oop(), "expected an oop");
814 HeapWord* addr = (HeapWord*)p;
815 assert(addr >= _prevMarkBitMap->startWord() ||
816 addr < _prevMarkBitMap->endWord(), "in a region");
818 return _prevMarkBitMap->isMarked(addr);
819 }
821 inline bool do_yield_check(uint worker_i = 0);
822 inline bool should_yield();
824 // Called to abort the marking cycle after a Full GC takes palce.
825 void abort();
827 // This prints the global/local fingers. It is used for debugging.
828 NOT_PRODUCT(void print_finger();)
830 void print_summary_info();
832 void print_worker_threads_on(outputStream* st) const;
834 void print_on_error(outputStream* st) const;
836 // The following indicate whether a given verbose level has been
837 // set. Notice that anything above stats is conditional to
838 // _MARKING_VERBOSE_ having been set to 1
839 bool verbose_stats() {
840 return _verbose_level >= stats_verbose;
841 }
842 bool verbose_low() {
843 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
844 }
845 bool verbose_medium() {
846 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
847 }
848 bool verbose_high() {
849 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
850 }
852 // Liveness counting
854 // Utility routine to set an exclusive range of cards on the given
855 // card liveness bitmap
856 inline void set_card_bitmap_range(BitMap* card_bm,
857 BitMap::idx_t start_idx,
858 BitMap::idx_t end_idx,
859 bool is_par);
861 // Returns the card number of the bottom of the G1 heap.
862 // Used in biasing indices into accounting card bitmaps.
863 intptr_t heap_bottom_card_num() const {
864 return _heap_bottom_card_num;
865 }
867 // Returns the card bitmap for a given task or worker id.
868 BitMap* count_card_bitmap_for(uint worker_id) {
869 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
870 assert(_count_card_bitmaps != NULL, "uninitialized");
871 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
872 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
873 return task_card_bm;
874 }
876 // Returns the array containing the marked bytes for each region,
877 // for the given worker or task id.
878 size_t* count_marked_bytes_array_for(uint worker_id) {
879 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
880 assert(_count_marked_bytes != NULL, "uninitialized");
881 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
882 assert(marked_bytes_array != NULL, "uninitialized");
883 return marked_bytes_array;
884 }
886 // Returns the index in the liveness accounting card table bitmap
887 // for the given address
888 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
890 // Counts the size of the given memory region in the the given
891 // marked_bytes array slot for the given HeapRegion.
892 // Sets the bits in the given card bitmap that are associated with the
893 // cards that are spanned by the memory region.
894 inline void count_region(MemRegion mr, HeapRegion* hr,
895 size_t* marked_bytes_array,
896 BitMap* task_card_bm);
898 // Counts the given memory region in the task/worker counting
899 // data structures for the given worker id.
900 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
902 // Counts the given memory region in the task/worker counting
903 // data structures for the given worker id.
904 inline void count_region(MemRegion mr, uint worker_id);
906 // Counts the given object in the given task/worker counting
907 // data structures.
908 inline void count_object(oop obj, HeapRegion* hr,
909 size_t* marked_bytes_array,
910 BitMap* task_card_bm);
912 // Counts the given object in the task/worker counting data
913 // structures for the given worker id.
914 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
916 // Attempts to mark the given object and, if successful, counts
917 // the object in the given task/worker counting structures.
918 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
919 size_t* marked_bytes_array,
920 BitMap* task_card_bm);
922 // Attempts to mark the given object and, if successful, counts
923 // the object in the task/worker counting structures for the
924 // given worker id.
925 inline bool par_mark_and_count(oop obj, size_t word_size,
926 HeapRegion* hr, uint worker_id);
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, HeapRegion* hr, uint worker_id);
933 // Similar to the above routine but we don't know the heap region that
934 // contains the object to be marked/counted, which this routine looks up.
935 inline bool par_mark_and_count(oop obj, uint worker_id);
937 // Similar to the above routine but there are times when we cannot
938 // safely calculate the size of obj due to races and we, therefore,
939 // pass the size in as a parameter. It is the caller's reponsibility
940 // to ensure that the size passed in for obj is valid.
941 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
943 // Unconditionally mark the given object, and unconditinally count
944 // the object in the counting structures for worker id 0.
945 // Should *not* be called from parallel code.
946 inline bool mark_and_count(oop obj, HeapRegion* hr);
948 // Similar to the above routine but we don't know the heap region that
949 // contains the object to be marked/counted, which this routine looks up.
950 // Should *not* be called from parallel code.
951 inline bool mark_and_count(oop obj);
953 // Returns true if initialization was successfully completed.
954 bool completed_initialization() const {
955 return _completed_initialization;
956 }
958 protected:
959 // Clear all the per-task bitmaps and arrays used to store the
960 // counting data.
961 void clear_all_count_data();
963 // Aggregates the counting data for each worker/task
964 // that was constructed while marking. Also sets
965 // the amount of marked bytes for each region and
966 // the top at concurrent mark count.
967 void aggregate_count_data();
969 // Verification routine
970 void verify_count_data();
971 };
973 // A class representing a marking task.
974 class CMTask : public TerminatorTerminator {
975 private:
976 enum PrivateConstants {
977 // the regular clock call is called once the scanned words reaches
978 // this limit
979 words_scanned_period = 12*1024,
980 // the regular clock call is called once the number of visited
981 // references reaches this limit
982 refs_reached_period = 384,
983 // initial value for the hash seed, used in the work stealing code
984 init_hash_seed = 17,
985 // how many entries will be transferred between global stack and
986 // local queues
987 global_stack_transfer_size = 16
988 };
990 uint _worker_id;
991 G1CollectedHeap* _g1h;
992 ConcurrentMark* _cm;
993 CMBitMap* _nextMarkBitMap;
994 // the task queue of this task
995 CMTaskQueue* _task_queue;
996 private:
997 // the task queue set---needed for stealing
998 CMTaskQueueSet* _task_queues;
999 // indicates whether the task has been claimed---this is only for
1000 // debugging purposes
1001 bool _claimed;
1003 // number of calls to this task
1004 int _calls;
1006 // when the virtual timer reaches this time, the marking step should
1007 // exit
1008 double _time_target_ms;
1009 // the start time of the current marking step
1010 double _start_time_ms;
1012 // the oop closure used for iterations over oops
1013 G1CMOopClosure* _cm_oop_closure;
1015 // the region this task is scanning, NULL if we're not scanning any
1016 HeapRegion* _curr_region;
1017 // the local finger of this task, NULL if we're not scanning a region
1018 HeapWord* _finger;
1019 // limit of the region this task is scanning, NULL if we're not scanning one
1020 HeapWord* _region_limit;
1022 // the number of words this task has scanned
1023 size_t _words_scanned;
1024 // When _words_scanned reaches this limit, the regular clock is
1025 // called. Notice that this might be decreased under certain
1026 // circumstances (i.e. when we believe that we did an expensive
1027 // operation).
1028 size_t _words_scanned_limit;
1029 // the initial value of _words_scanned_limit (i.e. what it was
1030 // before it was decreased).
1031 size_t _real_words_scanned_limit;
1033 // the number of references this task has visited
1034 size_t _refs_reached;
1035 // When _refs_reached reaches this limit, the regular clock is
1036 // called. Notice this this might be decreased under certain
1037 // circumstances (i.e. when we believe that we did an expensive
1038 // operation).
1039 size_t _refs_reached_limit;
1040 // the initial value of _refs_reached_limit (i.e. what it was before
1041 // it was decreased).
1042 size_t _real_refs_reached_limit;
1044 // used by the work stealing stuff
1045 int _hash_seed;
1046 // if this is true, then the task has aborted for some reason
1047 bool _has_aborted;
1048 // set when the task aborts because it has met its time quota
1049 bool _has_timed_out;
1050 // true when we're draining SATB buffers; this avoids the task
1051 // aborting due to SATB buffers being available (as we're already
1052 // dealing with them)
1053 bool _draining_satb_buffers;
1055 // number sequence of past step times
1056 NumberSeq _step_times_ms;
1057 // elapsed time of this task
1058 double _elapsed_time_ms;
1059 // termination time of this task
1060 double _termination_time_ms;
1061 // when this task got into the termination protocol
1062 double _termination_start_time_ms;
1064 // true when the task is during a concurrent phase, false when it is
1065 // in the remark phase (so, in the latter case, we do not have to
1066 // check all the things that we have to check during the concurrent
1067 // phase, i.e. SATB buffer availability...)
1068 bool _concurrent;
1070 TruncatedSeq _marking_step_diffs_ms;
1072 // Counting data structures. Embedding the task's marked_bytes_array
1073 // and card bitmap into the actual task saves having to go through
1074 // the ConcurrentMark object.
1075 size_t* _marked_bytes_array;
1076 BitMap* _card_bm;
1078 // LOTS of statistics related with this task
1079 #if _MARKING_STATS_
1080 NumberSeq _all_clock_intervals_ms;
1081 double _interval_start_time_ms;
1083 int _aborted;
1084 int _aborted_overflow;
1085 int _aborted_cm_aborted;
1086 int _aborted_yield;
1087 int _aborted_timed_out;
1088 int _aborted_satb;
1089 int _aborted_termination;
1091 int _steal_attempts;
1092 int _steals;
1094 int _clock_due_to_marking;
1095 int _clock_due_to_scanning;
1097 int _local_pushes;
1098 int _local_pops;
1099 int _local_max_size;
1100 int _objs_scanned;
1102 int _global_pushes;
1103 int _global_pops;
1104 int _global_max_size;
1106 int _global_transfers_to;
1107 int _global_transfers_from;
1109 int _regions_claimed;
1110 int _objs_found_on_bitmap;
1112 int _satb_buffers_processed;
1113 #endif // _MARKING_STATS_
1115 // it updates the local fields after this task has claimed
1116 // a new region to scan
1117 void setup_for_region(HeapRegion* hr);
1118 // it brings up-to-date the limit of the region
1119 void update_region_limit();
1121 // called when either the words scanned or the refs visited limit
1122 // has been reached
1123 void reached_limit();
1124 // recalculates the words scanned and refs visited limits
1125 void recalculate_limits();
1126 // decreases the words scanned and refs visited limits when we reach
1127 // an expensive operation
1128 void decrease_limits();
1129 // it checks whether the words scanned or refs visited reached their
1130 // respective limit and calls reached_limit() if they have
1131 void check_limits() {
1132 if (_words_scanned >= _words_scanned_limit ||
1133 _refs_reached >= _refs_reached_limit) {
1134 reached_limit();
1135 }
1136 }
1137 // this is supposed to be called regularly during a marking step as
1138 // it checks a bunch of conditions that might cause the marking step
1139 // to abort
1140 void regular_clock_call();
1141 bool concurrent() { return _concurrent; }
1143 public:
1144 // It resets the task; it should be called right at the beginning of
1145 // a marking phase.
1146 void reset(CMBitMap* _nextMarkBitMap);
1147 // it clears all the fields that correspond to a claimed region.
1148 void clear_region_fields();
1150 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1152 // The main method of this class which performs a marking step
1153 // trying not to exceed the given duration. However, it might exit
1154 // prematurely, according to some conditions (i.e. SATB buffers are
1155 // available for processing).
1156 void do_marking_step(double target_ms,
1157 bool do_termination,
1158 bool is_serial);
1160 // These two calls start and stop the timer
1161 void record_start_time() {
1162 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1163 }
1164 void record_end_time() {
1165 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1166 }
1168 // returns the worker ID associated with this task.
1169 uint worker_id() { return _worker_id; }
1171 // From TerminatorTerminator. It determines whether this task should
1172 // exit the termination protocol after it's entered it.
1173 virtual bool should_exit_termination();
1175 // Resets the local region fields after a task has finished scanning a
1176 // region; or when they have become stale as a result of the region
1177 // being evacuated.
1178 void giveup_current_region();
1180 HeapWord* finger() { return _finger; }
1182 bool has_aborted() { return _has_aborted; }
1183 void set_has_aborted() { _has_aborted = true; }
1184 void clear_has_aborted() { _has_aborted = false; }
1185 bool has_timed_out() { return _has_timed_out; }
1186 bool claimed() { return _claimed; }
1188 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1190 // It grays the object by marking it and, if necessary, pushing it
1191 // on the local queue
1192 inline void deal_with_reference(oop obj);
1194 // It scans an object and visits its children.
1195 void scan_object(oop obj);
1197 // It pushes an object on the local queue.
1198 inline void push(oop obj);
1200 // These two move entries to/from the global stack.
1201 void move_entries_to_global_stack();
1202 void get_entries_from_global_stack();
1204 // It pops and scans objects from the local queue. If partially is
1205 // true, then it stops when the queue size is of a given limit. If
1206 // partially is false, then it stops when the queue is empty.
1207 void drain_local_queue(bool partially);
1208 // It moves entries from the global stack to the local queue and
1209 // drains the local queue. If partially is true, then it stops when
1210 // both the global stack and the local queue reach a given size. If
1211 // partially if false, it tries to empty them totally.
1212 void drain_global_stack(bool partially);
1213 // It keeps picking SATB buffers and processing them until no SATB
1214 // buffers are available.
1215 void drain_satb_buffers();
1217 // moves the local finger to a new location
1218 inline void move_finger_to(HeapWord* new_finger) {
1219 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1220 _finger = new_finger;
1221 }
1223 CMTask(uint worker_id, ConcurrentMark *cm,
1224 size_t* marked_bytes, BitMap* card_bm,
1225 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1227 // it prints statistics associated with this task
1228 void print_stats();
1230 #if _MARKING_STATS_
1231 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1232 #endif // _MARKING_STATS_
1233 };
1235 // Class that's used to to print out per-region liveness
1236 // information. It's currently used at the end of marking and also
1237 // after we sort the old regions at the end of the cleanup operation.
1238 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1239 private:
1240 outputStream* _out;
1242 // Accumulators for these values.
1243 size_t _total_used_bytes;
1244 size_t _total_capacity_bytes;
1245 size_t _total_prev_live_bytes;
1246 size_t _total_next_live_bytes;
1248 // These are set up when we come across a "stars humongous" region
1249 // (as this is where most of this information is stored, not in the
1250 // subsequent "continues humongous" regions). After that, for every
1251 // region in a given humongous region series we deduce the right
1252 // values for it by simply subtracting the appropriate amount from
1253 // these fields. All these values should reach 0 after we've visited
1254 // the last region in the series.
1255 size_t _hum_used_bytes;
1256 size_t _hum_capacity_bytes;
1257 size_t _hum_prev_live_bytes;
1258 size_t _hum_next_live_bytes;
1260 static double perc(size_t val, size_t total) {
1261 if (total == 0) {
1262 return 0.0;
1263 } else {
1264 return 100.0 * ((double) val / (double) total);
1265 }
1266 }
1268 static double bytes_to_mb(size_t val) {
1269 return (double) val / (double) M;
1270 }
1272 // See the .cpp file.
1273 size_t get_hum_bytes(size_t* hum_bytes);
1274 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1275 size_t* prev_live_bytes, size_t* next_live_bytes);
1277 public:
1278 // The header and footer are printed in the constructor and
1279 // destructor respectively.
1280 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1281 virtual bool doHeapRegion(HeapRegion* r);
1282 ~G1PrintRegionLivenessInfoClosure();
1283 };
1285 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP