Mon, 12 Mar 2012 14:59:00 -0700
7147724: G1: hang in SurrogateLockerThread::manipulatePLL
Summary: Attempting to initiate a marking cycle when allocating a humongous object can, if a marking cycle is successfully initiated by another thread, result in the allocating thread spinning until the marking cycle is complete. Eliminate a deadlock between the main ConcurrentMarkThread, the SurrogateLocker thread, the VM thread, and a mutator thread waiting on the SecondaryFreeList_lock (while free regions are going to become available) by not manipulating the pending list lock during the prologue and epilogue of the cleanup pause.
Reviewed-by: brutisso, jcoomes, tonyp
1 /*
2 * Copyright (c) 2001, 2012, 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> CMTaskQueue;
34 typedef GenericTaskQueueSet<CMTaskQueue> 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) :
46 _g1(g1)
47 {}
49 void do_object(oop obj) {
50 ShouldNotCallThis();
51 }
52 bool do_object_b(oop obj);
53 };
55 // A generic CM bit map. This is essentially a wrapper around the BitMap
56 // class, with one bit per (1<<_shifter) HeapWords.
58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
59 protected:
60 HeapWord* _bmStartWord; // base address of range covered by map
61 size_t _bmWordSize; // map size (in #HeapWords covered)
62 const int _shifter; // map to char or bit
63 VirtualSpace _virtual_space; // underlying the bit map
64 BitMap _bm; // the bit map itself
66 public:
67 // constructor
68 CMBitMapRO(ReservedSpace rs, int shifter);
70 enum { do_yield = true };
72 // inquiries
73 HeapWord* startWord() const { return _bmStartWord; }
74 size_t sizeInWords() const { return _bmWordSize; }
75 // the following is one past the last word in space
76 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
78 // read marks
80 bool isMarked(HeapWord* addr) const {
81 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
82 "outside underlying space?");
83 return _bm.at(heapWordToOffset(addr));
84 }
86 // iteration
87 inline bool iterate(BitMapClosure* cl, MemRegion mr);
88 inline bool iterate(BitMapClosure* cl);
90 // Return the address corresponding to the next marked bit at or after
91 // "addr", and before "limit", if "limit" is non-NULL. If there is no
92 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
93 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
94 HeapWord* limit = NULL) const;
95 // Return the address corresponding to the next unmarked bit at or after
96 // "addr", and before "limit", if "limit" is non-NULL. If there is no
97 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
98 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
99 HeapWord* limit = NULL) const;
101 // conversion utilities
102 // XXX Fix these so that offsets are size_t's...
103 HeapWord* offsetToHeapWord(size_t offset) const {
104 return _bmStartWord + (offset << _shifter);
105 }
106 size_t heapWordToOffset(HeapWord* addr) const {
107 return pointer_delta(addr, _bmStartWord) >> _shifter;
108 }
109 int heapWordDiffToOffsetDiff(size_t diff) const;
110 HeapWord* nextWord(HeapWord* addr) {
111 return offsetToHeapWord(heapWordToOffset(addr) + 1);
112 }
114 void mostly_disjoint_range_union(BitMap* from_bitmap,
115 size_t from_start_index,
116 HeapWord* to_start_word,
117 size_t word_num);
119 // debugging
120 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
121 };
123 class CMBitMap : public CMBitMapRO {
125 public:
126 // constructor
127 CMBitMap(ReservedSpace rs, int shifter) :
128 CMBitMapRO(rs, shifter) {}
130 // write marks
131 void mark(HeapWord* addr) {
132 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
133 "outside underlying space?");
134 _bm.set_bit(heapWordToOffset(addr));
135 }
136 void clear(HeapWord* addr) {
137 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
138 "outside underlying space?");
139 _bm.clear_bit(heapWordToOffset(addr));
140 }
141 bool parMark(HeapWord* addr) {
142 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
143 "outside underlying space?");
144 return _bm.par_set_bit(heapWordToOffset(addr));
145 }
146 bool parClear(HeapWord* addr) {
147 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
148 "outside underlying space?");
149 return _bm.par_clear_bit(heapWordToOffset(addr));
150 }
151 void markRange(MemRegion mr);
152 void clearAll();
153 void clearRange(MemRegion mr);
155 // Starting at the bit corresponding to "addr" (inclusive), find the next
156 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
157 // the end of this run (stopping at "end_addr"). Return the MemRegion
158 // covering from the start of the region corresponding to the first bit
159 // of the run to the end of the region corresponding to the last bit of
160 // the run. If there is no "1" bit at or after "addr", return an empty
161 // MemRegion.
162 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
163 };
165 // Represents a marking stack used by the CM collector.
166 // Ideally this should be GrowableArray<> just like MSC's marking stack(s).
167 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
168 ConcurrentMark* _cm;
169 oop* _base; // bottom of stack
170 jint _index; // one more than last occupied index
171 jint _capacity; // max #elements
172 jint _saved_index; // value of _index saved at start of GC
173 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
175 bool _overflow;
176 DEBUG_ONLY(bool _drain_in_progress;)
177 DEBUG_ONLY(bool _drain_in_progress_yields;)
179 public:
180 CMMarkStack(ConcurrentMark* cm);
181 ~CMMarkStack();
183 void allocate(size_t size);
185 oop pop() {
186 if (!isEmpty()) {
187 return _base[--_index] ;
188 }
189 return NULL;
190 }
192 // If overflow happens, don't do the push, and record the overflow.
193 // *Requires* that "ptr" is already marked.
194 void push(oop ptr) {
195 if (isFull()) {
196 // Record overflow.
197 _overflow = true;
198 return;
199 } else {
200 _base[_index++] = ptr;
201 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
202 }
203 }
204 // Non-block impl. Note: concurrency is allowed only with other
205 // "par_push" operations, not with "pop" or "drain". We would need
206 // parallel versions of them if such concurrency was desired.
207 void par_push(oop ptr);
209 // Pushes the first "n" elements of "ptr_arr" on the stack.
210 // Non-block impl. Note: concurrency is allowed only with other
211 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
212 void par_adjoin_arr(oop* ptr_arr, int n);
214 // Pushes the first "n" elements of "ptr_arr" on the stack.
215 // Locking impl: concurrency is allowed only with
216 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
217 // locking strategy.
218 void par_push_arr(oop* ptr_arr, int n);
220 // If returns false, the array was empty. Otherwise, removes up to "max"
221 // elements from the stack, and transfers them to "ptr_arr" in an
222 // unspecified order. The actual number transferred is given in "n" ("n
223 // == 0" is deliberately redundant with the return value.) Locking impl:
224 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
225 // operations, which use the same locking strategy.
226 bool par_pop_arr(oop* ptr_arr, int max, int* n);
228 // Drain the mark stack, applying the given closure to all fields of
229 // objects on the stack. (That is, continue until the stack is empty,
230 // even if closure applications add entries to the stack.) The "bm"
231 // argument, if non-null, may be used to verify that only marked objects
232 // are on the mark stack. If "yield_after" is "true", then the
233 // concurrent marker performing the drain offers to yield after
234 // processing each object. If a yield occurs, stops the drain operation
235 // and returns false. Otherwise, returns true.
236 template<class OopClosureClass>
237 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
239 bool isEmpty() { return _index == 0; }
240 bool isFull() { return _index == _capacity; }
241 int maxElems() { return _capacity; }
243 bool overflow() { return _overflow; }
244 void clear_overflow() { _overflow = false; }
246 int size() { return _index; }
248 void setEmpty() { _index = 0; clear_overflow(); }
250 // Record the current index.
251 void note_start_of_gc();
253 // Make sure that we have not added any entries to the stack during GC.
254 void note_end_of_gc();
256 // iterate over the oops in the mark stack, up to the bound recorded via
257 // the call above.
258 void oops_do(OopClosure* f);
259 };
261 class CMRegionStack VALUE_OBJ_CLASS_SPEC {
262 MemRegion* _base;
263 jint _capacity;
264 jint _index;
265 jint _oops_do_bound;
266 bool _overflow;
267 public:
268 CMRegionStack();
269 ~CMRegionStack();
270 void allocate(size_t size);
272 // This is lock-free; assumes that it will only be called in parallel
273 // with other "push" operations (no pops).
274 void push_lock_free(MemRegion mr);
276 // Lock-free; assumes that it will only be called in parallel
277 // with other "pop" operations (no pushes).
278 MemRegion pop_lock_free();
280 #if 0
281 // The routines that manipulate the region stack with a lock are
282 // not currently used. They should be retained, however, as a
283 // diagnostic aid.
285 // These two are the implementations that use a lock. They can be
286 // called concurrently with each other but they should not be called
287 // concurrently with the lock-free versions (push() / pop()).
288 void push_with_lock(MemRegion mr);
289 MemRegion pop_with_lock();
290 #endif
292 bool isEmpty() { return _index == 0; }
293 bool isFull() { return _index == _capacity; }
295 bool overflow() { return _overflow; }
296 void clear_overflow() { _overflow = false; }
298 int size() { return _index; }
300 // It iterates over the entries in the region stack and it
301 // invalidates (i.e. assigns MemRegion()) the ones that point to
302 // regions in the collection set.
303 bool invalidate_entries_into_cset();
305 // This gives an upper bound up to which the iteration in
306 // invalidate_entries_into_cset() will reach. This prevents
307 // newly-added entries to be unnecessarily scanned.
308 void set_oops_do_bound() {
309 _oops_do_bound = _index;
310 }
312 void setEmpty() { _index = 0; clear_overflow(); }
313 };
315 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
316 private:
317 #ifndef PRODUCT
318 uintx _num_remaining;
319 bool _force;
320 #endif // !defined(PRODUCT)
322 public:
323 void init() PRODUCT_RETURN;
324 void update() PRODUCT_RETURN;
325 bool should_force() PRODUCT_RETURN_( return false; );
326 };
328 // this will enable a variety of different statistics per GC task
329 #define _MARKING_STATS_ 0
330 // this will enable the higher verbose levels
331 #define _MARKING_VERBOSE_ 0
333 #if _MARKING_STATS_
334 #define statsOnly(statement) \
335 do { \
336 statement ; \
337 } while (0)
338 #else // _MARKING_STATS_
339 #define statsOnly(statement) \
340 do { \
341 } while (0)
342 #endif // _MARKING_STATS_
344 typedef enum {
345 no_verbose = 0, // verbose turned off
346 stats_verbose, // only prints stats at the end of marking
347 low_verbose, // low verbose, mostly per region and per major event
348 medium_verbose, // a bit more detailed than low
349 high_verbose // per object verbose
350 } CMVerboseLevel;
352 class YoungList;
354 // Root Regions are regions that are not empty at the beginning of a
355 // marking cycle and which we might collect during an evacuation pause
356 // while the cycle is active. Given that, during evacuation pauses, we
357 // do not copy objects that are explicitly marked, what we have to do
358 // for the root regions is to scan them and mark all objects reachable
359 // from them. According to the SATB assumptions, we only need to visit
360 // each object once during marking. So, as long as we finish this scan
361 // before the next evacuation pause, we can copy the objects from the
362 // root regions without having to mark them or do anything else to them.
363 //
364 // Currently, we only support root region scanning once (at the start
365 // of the marking cycle) and the root regions are all the survivor
366 // regions populated during the initial-mark pause.
367 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
368 private:
369 YoungList* _young_list;
370 ConcurrentMark* _cm;
372 volatile bool _scan_in_progress;
373 volatile bool _should_abort;
374 HeapRegion* volatile _next_survivor;
376 public:
377 CMRootRegions();
378 // We actually do most of the initialization in this method.
379 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
381 // Reset the claiming / scanning of the root regions.
382 void prepare_for_scan();
384 // Forces get_next() to return NULL so that the iteration aborts early.
385 void abort() { _should_abort = true; }
387 // Return true if the CM thread are actively scanning root regions,
388 // false otherwise.
389 bool scan_in_progress() { return _scan_in_progress; }
391 // Claim the next root region to scan atomically, or return NULL if
392 // all have been claimed.
393 HeapRegion* claim_next();
395 // Flag that we're done with root region scanning and notify anyone
396 // who's waiting on it. If aborted is false, assume that all regions
397 // have been claimed.
398 void scan_finished();
400 // If CM threads are still scanning root regions, wait until they
401 // are done. Return true if we had to wait, false otherwise.
402 bool wait_until_scan_finished();
403 };
405 class ConcurrentMarkThread;
407 class ConcurrentMark : public CHeapObj {
408 friend class ConcurrentMarkThread;
409 friend class CMTask;
410 friend class CMBitMapClosure;
411 friend class CSetMarkOopClosure;
412 friend class CMGlobalObjectClosure;
413 friend class CMRemarkTask;
414 friend class CMConcurrentMarkingTask;
415 friend class G1ParNoteEndTask;
416 friend class CalcLiveObjectsClosure;
417 friend class G1CMRefProcTaskProxy;
418 friend class G1CMRefProcTaskExecutor;
419 friend class G1CMParKeepAliveAndDrainClosure;
420 friend class G1CMParDrainMarkingStackClosure;
422 protected:
423 ConcurrentMarkThread* _cmThread; // the thread doing the work
424 G1CollectedHeap* _g1h; // the heap.
425 uint _parallel_marking_threads; // the number of marking
426 // threads we're use
427 uint _max_parallel_marking_threads; // max number of marking
428 // threads we'll ever use
429 double _sleep_factor; // how much we have to sleep, with
430 // respect to the work we just did, to
431 // meet the marking overhead goal
432 double _marking_task_overhead; // marking target overhead for
433 // a single task
435 // same as the two above, but for the cleanup task
436 double _cleanup_sleep_factor;
437 double _cleanup_task_overhead;
439 FreeRegionList _cleanup_list;
441 // Concurrent marking support structures
442 CMBitMap _markBitMap1;
443 CMBitMap _markBitMap2;
444 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
445 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
446 bool _at_least_one_mark_complete;
448 BitMap _region_bm;
449 BitMap _card_bm;
451 // Heap bounds
452 HeapWord* _heap_start;
453 HeapWord* _heap_end;
455 // Root region tracking and claiming.
456 CMRootRegions _root_regions;
458 // For gray objects
459 CMMarkStack _markStack; // Grey objects behind global finger.
460 CMRegionStack _regionStack; // Grey regions behind global finger.
461 HeapWord* volatile _finger; // the global finger, region aligned,
462 // always points to the end of the
463 // last claimed region
465 // marking tasks
466 uint _max_task_num; // maximum task number
467 uint _active_tasks; // task num currently active
468 CMTask** _tasks; // task queue array (max_task_num len)
469 CMTaskQueueSet* _task_queues; // task queue set
470 ParallelTaskTerminator _terminator; // for termination
472 // Two sync barriers that are used to synchronise tasks when an
473 // overflow occurs. The algorithm is the following. All tasks enter
474 // the first one to ensure that they have all stopped manipulating
475 // the global data structures. After they exit it, they re-initialise
476 // their data structures and task 0 re-initialises the global data
477 // structures. Then, they enter the second sync barrier. This
478 // ensure, that no task starts doing work before all data
479 // structures (local and global) have been re-initialised. When they
480 // exit it, they are free to start working again.
481 WorkGangBarrierSync _first_overflow_barrier_sync;
482 WorkGangBarrierSync _second_overflow_barrier_sync;
484 // this is set by any task, when an overflow on the global data
485 // structures is detected.
486 volatile bool _has_overflown;
487 // true: marking is concurrent, false: we're in remark
488 volatile bool _concurrent;
489 // set at the end of a Full GC so that marking aborts
490 volatile bool _has_aborted;
492 // used when remark aborts due to an overflow to indicate that
493 // another concurrent marking phase should start
494 volatile bool _restart_for_overflow;
496 // This is true from the very start of concurrent marking until the
497 // point when all the tasks complete their work. It is really used
498 // to determine the points between the end of concurrent marking and
499 // time of remark.
500 volatile bool _concurrent_marking_in_progress;
502 // verbose level
503 CMVerboseLevel _verbose_level;
505 // These two fields are used to implement the optimisation that
506 // avoids pushing objects on the global/region stack if there are
507 // no collection set regions above the lowest finger.
509 // This is the lowest finger (among the global and local fingers),
510 // which is calculated before a new collection set is chosen.
511 HeapWord* _min_finger;
512 // If this flag is true, objects/regions that are marked below the
513 // finger should be pushed on the stack(s). If this is flag is
514 // false, it is safe not to push them on the stack(s).
515 bool _should_gray_objects;
517 // All of these times are in ms.
518 NumberSeq _init_times;
519 NumberSeq _remark_times;
520 NumberSeq _remark_mark_times;
521 NumberSeq _remark_weak_ref_times;
522 NumberSeq _cleanup_times;
523 double _total_counting_time;
524 double _total_rs_scrub_time;
526 double* _accum_task_vtime; // accumulated task vtime
528 FlexibleWorkGang* _parallel_workers;
530 ForceOverflowSettings _force_overflow_conc;
531 ForceOverflowSettings _force_overflow_stw;
533 void weakRefsWork(bool clear_all_soft_refs);
535 void swapMarkBitMaps();
537 // It resets the global marking data structures, as well as the
538 // task local ones; should be called during initial mark.
539 void reset();
540 // It resets all the marking data structures.
541 void clear_marking_state(bool clear_overflow = true);
543 // It should be called to indicate which phase we're in (concurrent
544 // mark or remark) and how many threads are currently active.
545 void set_phase(uint active_tasks, bool concurrent);
546 // We do this after we're done with marking so that the marking data
547 // structures are initialised to a sensible and predictable state.
548 void set_non_marking_state();
550 // prints all gathered CM-related statistics
551 void print_stats();
553 bool cleanup_list_is_empty() {
554 return _cleanup_list.is_empty();
555 }
557 // accessor methods
558 uint parallel_marking_threads() { return _parallel_marking_threads; }
559 uint max_parallel_marking_threads() { return _max_parallel_marking_threads;}
560 double sleep_factor() { return _sleep_factor; }
561 double marking_task_overhead() { return _marking_task_overhead;}
562 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
563 double cleanup_task_overhead() { return _cleanup_task_overhead;}
565 HeapWord* finger() { return _finger; }
566 bool concurrent() { return _concurrent; }
567 uint active_tasks() { return _active_tasks; }
568 ParallelTaskTerminator* terminator() { return &_terminator; }
570 // It claims the next available region to be scanned by a marking
571 // task. It might return NULL if the next region is empty or we have
572 // run out of regions. In the latter case, out_of_regions()
573 // determines whether we've really run out of regions or the task
574 // should call claim_region() again. This might seem a bit
575 // awkward. Originally, the code was written so that claim_region()
576 // either successfully returned with a non-empty region or there
577 // were no more regions to be claimed. The problem with this was
578 // that, in certain circumstances, it iterated over large chunks of
579 // the heap finding only empty regions and, while it was working, it
580 // was preventing the calling task to call its regular clock
581 // method. So, this way, each task will spend very little time in
582 // claim_region() and is allowed to call the regular clock method
583 // frequently.
584 HeapRegion* claim_region(int task);
586 // It determines whether we've run out of regions to scan.
587 bool out_of_regions() { return _finger == _heap_end; }
589 // Returns the task with the given id
590 CMTask* task(int id) {
591 assert(0 <= id && id < (int) _active_tasks,
592 "task id not within active bounds");
593 return _tasks[id];
594 }
596 // Returns the task queue with the given id
597 CMTaskQueue* task_queue(int id) {
598 assert(0 <= id && id < (int) _active_tasks,
599 "task queue id not within active bounds");
600 return (CMTaskQueue*) _task_queues->queue(id);
601 }
603 // Returns the task queue set
604 CMTaskQueueSet* task_queues() { return _task_queues; }
606 // Access / manipulation of the overflow flag which is set to
607 // indicate that the global stack or region stack has overflown
608 bool has_overflown() { return _has_overflown; }
609 void set_has_overflown() { _has_overflown = true; }
610 void clear_has_overflown() { _has_overflown = false; }
611 bool restart_for_overflow() { return _restart_for_overflow; }
613 bool has_aborted() { return _has_aborted; }
615 // Methods to enter the two overflow sync barriers
616 void enter_first_sync_barrier(int task_num);
617 void enter_second_sync_barrier(int task_num);
619 ForceOverflowSettings* force_overflow_conc() {
620 return &_force_overflow_conc;
621 }
623 ForceOverflowSettings* force_overflow_stw() {
624 return &_force_overflow_stw;
625 }
627 ForceOverflowSettings* force_overflow() {
628 if (concurrent()) {
629 return force_overflow_conc();
630 } else {
631 return force_overflow_stw();
632 }
633 }
635 // Live Data Counting data structures...
636 // These data structures are initialized at the start of
637 // marking. They are written to while marking is active.
638 // They are aggregated during remark; the aggregated values
639 // are then used to populate the _region_bm, _card_bm, and
640 // the total live bytes, which are then subsequently updated
641 // during cleanup.
643 // An array of bitmaps (one bit map per task). Each bitmap
644 // is used to record the cards spanned by the live objects
645 // marked by that task/worker.
646 BitMap* _count_card_bitmaps;
648 // Used to record the number of marked live bytes
649 // (for each region, by worker thread).
650 size_t** _count_marked_bytes;
652 // Card index of the bottom of the G1 heap. Used for biasing indices into
653 // the card bitmaps.
654 intptr_t _heap_bottom_card_num;
656 public:
657 // Manipulation of the global mark stack.
658 // Notice that the first mark_stack_push is CAS-based, whereas the
659 // two below are Mutex-based. This is OK since the first one is only
660 // called during evacuation pauses and doesn't compete with the
661 // other two (which are called by the marking tasks during
662 // concurrent marking or remark).
663 bool mark_stack_push(oop p) {
664 _markStack.par_push(p);
665 if (_markStack.overflow()) {
666 set_has_overflown();
667 return false;
668 }
669 return true;
670 }
671 bool mark_stack_push(oop* arr, int n) {
672 _markStack.par_push_arr(arr, n);
673 if (_markStack.overflow()) {
674 set_has_overflown();
675 return false;
676 }
677 return true;
678 }
679 void mark_stack_pop(oop* arr, int max, int* n) {
680 _markStack.par_pop_arr(arr, max, n);
681 }
682 size_t mark_stack_size() { return _markStack.size(); }
683 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
684 bool mark_stack_overflow() { return _markStack.overflow(); }
685 bool mark_stack_empty() { return _markStack.isEmpty(); }
687 // (Lock-free) Manipulation of the region stack
688 bool region_stack_push_lock_free(MemRegion mr) {
689 // Currently we only call the lock-free version during evacuation
690 // pauses.
691 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
693 _regionStack.push_lock_free(mr);
694 if (_regionStack.overflow()) {
695 set_has_overflown();
696 return false;
697 }
698 return true;
699 }
701 // Lock-free version of region-stack pop. Should only be
702 // called in tandem with other lock-free pops.
703 MemRegion region_stack_pop_lock_free() {
704 return _regionStack.pop_lock_free();
705 }
707 #if 0
708 // The routines that manipulate the region stack with a lock are
709 // not currently used. They should be retained, however, as a
710 // diagnostic aid.
712 bool region_stack_push_with_lock(MemRegion mr) {
713 // Currently we only call the lock-based version during either
714 // concurrent marking or remark.
715 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
716 "if we are at a safepoint it should be the remark safepoint");
718 _regionStack.push_with_lock(mr);
719 if (_regionStack.overflow()) {
720 set_has_overflown();
721 return false;
722 }
723 return true;
724 }
726 MemRegion region_stack_pop_with_lock() {
727 // Currently we only call the lock-based version during either
728 // concurrent marking or remark.
729 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
730 "if we are at a safepoint it should be the remark safepoint");
732 return _regionStack.pop_with_lock();
733 }
734 #endif
736 int region_stack_size() { return _regionStack.size(); }
737 bool region_stack_overflow() { return _regionStack.overflow(); }
738 bool region_stack_empty() { return _regionStack.isEmpty(); }
740 // Iterate over any regions that were aborted while draining the
741 // region stack (any such regions are saved in the corresponding
742 // CMTask) and invalidate (i.e. assign to the empty MemRegion())
743 // any regions that point into the collection set.
744 bool invalidate_aborted_regions_in_cset();
746 // Returns true if there are any aborted memory regions.
747 bool has_aborted_regions();
749 CMRootRegions* root_regions() { return &_root_regions; }
751 bool concurrent_marking_in_progress() {
752 return _concurrent_marking_in_progress;
753 }
754 void set_concurrent_marking_in_progress() {
755 _concurrent_marking_in_progress = true;
756 }
757 void clear_concurrent_marking_in_progress() {
758 _concurrent_marking_in_progress = false;
759 }
761 void update_accum_task_vtime(int i, double vtime) {
762 _accum_task_vtime[i] += vtime;
763 }
765 double all_task_accum_vtime() {
766 double ret = 0.0;
767 for (int i = 0; i < (int)_max_task_num; ++i)
768 ret += _accum_task_vtime[i];
769 return ret;
770 }
772 // Attempts to steal an object from the task queues of other tasks
773 bool try_stealing(int task_num, int* hash_seed, oop& obj) {
774 return _task_queues->steal(task_num, hash_seed, obj);
775 }
777 // It grays an object by first marking it. Then, if it's behind the
778 // global finger, it also pushes it on the global stack.
779 void deal_with_reference(oop obj);
781 ConcurrentMark(ReservedSpace rs, int max_regions);
782 ~ConcurrentMark();
784 ConcurrentMarkThread* cmThread() { return _cmThread; }
786 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
787 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
789 // Returns the number of GC threads to be used in a concurrent
790 // phase based on the number of GC threads being used in a STW
791 // phase.
792 uint scale_parallel_threads(uint n_par_threads);
794 // Calculates the number of GC threads to be used in a concurrent phase.
795 uint calc_parallel_marking_threads();
797 // The following three are interaction between CM and
798 // G1CollectedHeap
800 // This notifies CM that a root during initial-mark needs to be
801 // grayed. It is MT-safe. word_size is the size of the object in
802 // words. It is passed explicitly as sometimes we cannot calculate
803 // it from the given object because it might be in an inconsistent
804 // state (e.g., in to-space and being copied). So the caller is
805 // responsible for dealing with this issue (e.g., get the size from
806 // the from-space image when the to-space image might be
807 // inconsistent) and always passing the size. hr is the region that
808 // contains the object and it's passed optionally from callers who
809 // might already have it (no point in recalculating it).
810 inline void grayRoot(oop obj, size_t word_size,
811 uint worker_id, HeapRegion* hr = NULL);
813 // It's used during evacuation pauses to gray a region, if
814 // necessary, and it's MT-safe. It assumes that the caller has
815 // marked any objects on that region. If _should_gray_objects is
816 // true and we're still doing concurrent marking, the region is
817 // pushed on the region stack, if it is located below the global
818 // finger, otherwise we do nothing.
819 void grayRegionIfNecessary(MemRegion mr);
821 // It's used during evacuation pauses to mark and, if necessary,
822 // gray a single object and it's MT-safe. It assumes the caller did
823 // not mark the object. If _should_gray_objects is true and we're
824 // still doing concurrent marking, the objects is pushed on the
825 // global stack, if it is located below the global finger, otherwise
826 // we do nothing.
827 void markAndGrayObjectIfNecessary(oop p);
829 // It iterates over the heap and for each object it comes across it
830 // will dump the contents of its reference fields, as well as
831 // liveness information for the object and its referents. The dump
832 // will be written to a file with the following name:
833 // G1PrintReachableBaseFile + "." + str.
834 // vo decides whether the prev (vo == UsePrevMarking), the next
835 // (vo == UseNextMarking) marking information, or the mark word
836 // (vo == UseMarkWord) will be used to determine the liveness of
837 // each object / referent.
838 // If all is true, all objects in the heap will be dumped, otherwise
839 // only the live ones. In the dump the following symbols / breviations
840 // are used:
841 // M : an explicitly live object (its bitmap bit is set)
842 // > : an implicitly live object (over tams)
843 // O : an object outside the G1 heap (typically: in the perm gen)
844 // NOT : a reference field whose referent is not live
845 // AND MARKED : indicates that an object is both explicitly and
846 // implicitly live (it should be one or the other, not both)
847 void print_reachable(const char* str,
848 VerifyOption vo, bool all) PRODUCT_RETURN;
850 // Clear the next marking bitmap (will be called concurrently).
851 void clearNextBitmap();
853 // These two do the work that needs to be done before and after the
854 // initial root checkpoint. Since this checkpoint can be done at two
855 // different points (i.e. an explicit pause or piggy-backed on a
856 // young collection), then it's nice to be able to easily share the
857 // pre/post code. It might be the case that we can put everything in
858 // the post method. TP
859 void checkpointRootsInitialPre();
860 void checkpointRootsInitialPost();
862 // Scan all the root regions and mark everything reachable from
863 // them.
864 void scanRootRegions();
866 // Scan a single root region and mark everything reachable from it.
867 void scanRootRegion(HeapRegion* hr, uint worker_id);
869 // Do concurrent phase of marking, to a tentative transitive closure.
870 void markFromRoots();
872 // Process all unprocessed SATB buffers. It is called at the
873 // beginning of an evacuation pause.
874 void drainAllSATBBuffers();
876 void checkpointRootsFinal(bool clear_all_soft_refs);
877 void checkpointRootsFinalWork();
878 void cleanup();
879 void completeCleanup();
881 // Mark in the previous bitmap. NB: this is usually read-only, so use
882 // this carefully!
883 inline void markPrev(oop p);
885 // Clears marks for all objects in the given range, for the prev,
886 // next, or both bitmaps. NB: the previous bitmap is usually
887 // read-only, so use this carefully!
888 void clearRangePrevBitmap(MemRegion mr);
889 void clearRangeNextBitmap(MemRegion mr);
890 void clearRangeBothBitmaps(MemRegion mr);
892 // Notify data structures that a GC has started.
893 void note_start_of_gc() {
894 _markStack.note_start_of_gc();
895 }
897 // Notify data structures that a GC is finished.
898 void note_end_of_gc() {
899 _markStack.note_end_of_gc();
900 }
902 // Iterate over the oops in the mark stack and all local queues. It
903 // also calls invalidate_entries_into_cset() on the region stack.
904 void oops_do(OopClosure* f);
906 // Verify that there are no CSet oops on the stacks (taskqueues /
907 // global mark stack), enqueued SATB buffers, per-thread SATB
908 // buffers, and fingers (global / per-task). The boolean parameters
909 // decide which of the above data structures to verify. If marking
910 // is not in progress, it's a no-op.
911 void verify_no_cset_oops(bool verify_stacks,
912 bool verify_enqueued_buffers,
913 bool verify_thread_buffers,
914 bool verify_fingers) PRODUCT_RETURN;
916 // It is called at the end of an evacuation pause during marking so
917 // that CM is notified of where the new end of the heap is. It
918 // doesn't do anything if concurrent_marking_in_progress() is false,
919 // unless the force parameter is true.
920 void update_g1_committed(bool force = false);
922 void complete_marking_in_collection_set();
924 // It indicates that a new collection set is being chosen.
925 void newCSet();
927 // It registers a collection set heap region with CM. This is used
928 // to determine whether any heap regions are located above the finger.
929 void registerCSetRegion(HeapRegion* hr);
931 // Resets the region fields of any active CMTask whose region fields
932 // are in the collection set (i.e. the region currently claimed by
933 // the CMTask will be evacuated and may be used, subsequently, as
934 // an alloc region). When this happens the region fields in the CMTask
935 // are stale and, hence, should be cleared causing the worker thread
936 // to claim a new region.
937 void reset_active_task_region_fields_in_cset();
939 // Registers the maximum region-end associated with a set of
940 // regions with CM. Again this is used to determine whether any
941 // heap regions are located above the finger.
942 void register_collection_set_finger(HeapWord* max_finger) {
943 // max_finger is the highest heap region end of the regions currently
944 // contained in the collection set. If this value is larger than
945 // _min_finger then we need to gray objects.
946 // This routine is like registerCSetRegion but for an entire
947 // collection of regions.
948 if (max_finger > _min_finger) {
949 _should_gray_objects = true;
950 }
951 }
953 // Returns "true" if at least one mark has been completed.
954 bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
956 bool isMarked(oop p) const {
957 assert(p != NULL && p->is_oop(), "expected an oop");
958 HeapWord* addr = (HeapWord*)p;
959 assert(addr >= _nextMarkBitMap->startWord() ||
960 addr < _nextMarkBitMap->endWord(), "in a region");
962 return _nextMarkBitMap->isMarked(addr);
963 }
965 inline bool not_yet_marked(oop p) const;
967 // XXX Debug code
968 bool containing_card_is_marked(void* p);
969 bool containing_cards_are_marked(void* start, void* last);
971 bool isPrevMarked(oop p) const {
972 assert(p != NULL && p->is_oop(), "expected an oop");
973 HeapWord* addr = (HeapWord*)p;
974 assert(addr >= _prevMarkBitMap->startWord() ||
975 addr < _prevMarkBitMap->endWord(), "in a region");
977 return _prevMarkBitMap->isMarked(addr);
978 }
980 inline bool do_yield_check(uint worker_i = 0);
981 inline bool should_yield();
983 // Called to abort the marking cycle after a Full GC takes palce.
984 void abort();
986 // This prints the global/local fingers. It is used for debugging.
987 NOT_PRODUCT(void print_finger();)
989 void print_summary_info();
991 void print_worker_threads_on(outputStream* st) const;
993 // The following indicate whether a given verbose level has been
994 // set. Notice that anything above stats is conditional to
995 // _MARKING_VERBOSE_ having been set to 1
996 bool verbose_stats() {
997 return _verbose_level >= stats_verbose;
998 }
999 bool verbose_low() {
1000 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
1001 }
1002 bool verbose_medium() {
1003 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
1004 }
1005 bool verbose_high() {
1006 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
1007 }
1009 // Counting data structure accessors
1011 // Returns the card number of the bottom of the G1 heap.
1012 // Used in biasing indices into accounting card bitmaps.
1013 intptr_t heap_bottom_card_num() const {
1014 return _heap_bottom_card_num;
1015 }
1017 // Returns the card bitmap for a given task or worker id.
1018 BitMap* count_card_bitmap_for(uint worker_id) {
1019 assert(0 <= worker_id && worker_id < _max_task_num, "oob");
1020 assert(_count_card_bitmaps != NULL, "uninitialized");
1021 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
1022 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
1023 return task_card_bm;
1024 }
1026 // Returns the array containing the marked bytes for each region,
1027 // for the given worker or task id.
1028 size_t* count_marked_bytes_array_for(uint worker_id) {
1029 assert(0 <= worker_id && worker_id < _max_task_num, "oob");
1030 assert(_count_marked_bytes != NULL, "uninitialized");
1031 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
1032 assert(marked_bytes_array != NULL, "uninitialized");
1033 return marked_bytes_array;
1034 }
1036 // Returns the index in the liveness accounting card table bitmap
1037 // for the given address
1038 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
1040 // Counts the size of the given memory region in the the given
1041 // marked_bytes array slot for the given HeapRegion.
1042 // Sets the bits in the given card bitmap that are associated with the
1043 // cards that are spanned by the memory region.
1044 inline void count_region(MemRegion mr, HeapRegion* hr,
1045 size_t* marked_bytes_array,
1046 BitMap* task_card_bm);
1048 // Counts the given memory region in the task/worker counting
1049 // data structures for the given worker id.
1050 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
1052 // Counts the given memory region in the task/worker counting
1053 // data structures for the given worker id.
1054 inline void count_region(MemRegion mr, uint worker_id);
1056 // Counts the given object in the given task/worker counting
1057 // data structures.
1058 inline void count_object(oop obj, HeapRegion* hr,
1059 size_t* marked_bytes_array,
1060 BitMap* task_card_bm);
1062 // Counts the given object in the task/worker counting data
1063 // structures for the given worker id.
1064 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
1066 // Attempts to mark the given object and, if successful, counts
1067 // the object in the given task/worker counting structures.
1068 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
1069 size_t* marked_bytes_array,
1070 BitMap* task_card_bm);
1072 // Attempts to mark the given object and, if successful, counts
1073 // the object in the task/worker counting structures for the
1074 // given worker id.
1075 inline bool par_mark_and_count(oop obj, size_t word_size,
1076 HeapRegion* hr, uint worker_id);
1078 // Attempts to mark the given object and, if successful, counts
1079 // the object in the task/worker counting structures for the
1080 // given worker id.
1081 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
1083 // Similar to the above routine but we don't know the heap region that
1084 // contains the object to be marked/counted, which this routine looks up.
1085 inline bool par_mark_and_count(oop obj, uint worker_id);
1087 // Similar to the above routine but there are times when we cannot
1088 // safely calculate the size of obj due to races and we, therefore,
1089 // pass the size in as a parameter. It is the caller's reponsibility
1090 // to ensure that the size passed in for obj is valid.
1091 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
1093 // Unconditionally mark the given object, and unconditinally count
1094 // the object in the counting structures for worker id 0.
1095 // Should *not* be called from parallel code.
1096 inline bool mark_and_count(oop obj, HeapRegion* hr);
1098 // Similar to the above routine but we don't know the heap region that
1099 // contains the object to be marked/counted, which this routine looks up.
1100 // Should *not* be called from parallel code.
1101 inline bool mark_and_count(oop obj);
1103 protected:
1104 // Clear all the per-task bitmaps and arrays used to store the
1105 // counting data.
1106 void clear_all_count_data();
1108 // Aggregates the counting data for each worker/task
1109 // that was constructed while marking. Also sets
1110 // the amount of marked bytes for each region and
1111 // the top at concurrent mark count.
1112 void aggregate_count_data();
1114 // Verification routine
1115 void verify_count_data();
1116 };
1118 // A class representing a marking task.
1119 class CMTask : public TerminatorTerminator {
1120 private:
1121 enum PrivateConstants {
1122 // the regular clock call is called once the scanned words reaches
1123 // this limit
1124 words_scanned_period = 12*1024,
1125 // the regular clock call is called once the number of visited
1126 // references reaches this limit
1127 refs_reached_period = 384,
1128 // initial value for the hash seed, used in the work stealing code
1129 init_hash_seed = 17,
1130 // how many entries will be transferred between global stack and
1131 // local queues
1132 global_stack_transfer_size = 16
1133 };
1135 int _task_id;
1136 G1CollectedHeap* _g1h;
1137 ConcurrentMark* _cm;
1138 CMBitMap* _nextMarkBitMap;
1139 // the task queue of this task
1140 CMTaskQueue* _task_queue;
1141 private:
1142 // the task queue set---needed for stealing
1143 CMTaskQueueSet* _task_queues;
1144 // indicates whether the task has been claimed---this is only for
1145 // debugging purposes
1146 bool _claimed;
1148 // number of calls to this task
1149 int _calls;
1151 // when the virtual timer reaches this time, the marking step should
1152 // exit
1153 double _time_target_ms;
1154 // the start time of the current marking step
1155 double _start_time_ms;
1157 // the oop closure used for iterations over oops
1158 G1CMOopClosure* _cm_oop_closure;
1160 // the region this task is scanning, NULL if we're not scanning any
1161 HeapRegion* _curr_region;
1162 // the local finger of this task, NULL if we're not scanning a region
1163 HeapWord* _finger;
1164 // limit of the region this task is scanning, NULL if we're not scanning one
1165 HeapWord* _region_limit;
1167 // This is used only when we scan regions popped from the region
1168 // stack. It records what the last object on such a region we
1169 // scanned was. It is used to ensure that, if we abort region
1170 // iteration, we do not rescan the first part of the region. This
1171 // should be NULL when we're not scanning a region from the region
1172 // stack.
1173 HeapWord* _region_finger;
1175 // If we abort while scanning a region we record the remaining
1176 // unscanned portion and check this field when marking restarts.
1177 // This avoids having to push on the region stack while other
1178 // marking threads may still be popping regions.
1179 // If we were to push the unscanned portion directly to the
1180 // region stack then we would need to using locking versions
1181 // of the push and pop operations.
1182 MemRegion _aborted_region;
1184 // the number of words this task has scanned
1185 size_t _words_scanned;
1186 // When _words_scanned reaches this limit, the regular clock is
1187 // called. Notice that this might be decreased under certain
1188 // circumstances (i.e. when we believe that we did an expensive
1189 // operation).
1190 size_t _words_scanned_limit;
1191 // the initial value of _words_scanned_limit (i.e. what it was
1192 // before it was decreased).
1193 size_t _real_words_scanned_limit;
1195 // the number of references this task has visited
1196 size_t _refs_reached;
1197 // When _refs_reached reaches this limit, the regular clock is
1198 // called. Notice this this might be decreased under certain
1199 // circumstances (i.e. when we believe that we did an expensive
1200 // operation).
1201 size_t _refs_reached_limit;
1202 // the initial value of _refs_reached_limit (i.e. what it was before
1203 // it was decreased).
1204 size_t _real_refs_reached_limit;
1206 // used by the work stealing stuff
1207 int _hash_seed;
1208 // if this is true, then the task has aborted for some reason
1209 bool _has_aborted;
1210 // set when the task aborts because it has met its time quota
1211 bool _has_timed_out;
1212 // true when we're draining SATB buffers; this avoids the task
1213 // aborting due to SATB buffers being available (as we're already
1214 // dealing with them)
1215 bool _draining_satb_buffers;
1217 // number sequence of past step times
1218 NumberSeq _step_times_ms;
1219 // elapsed time of this task
1220 double _elapsed_time_ms;
1221 // termination time of this task
1222 double _termination_time_ms;
1223 // when this task got into the termination protocol
1224 double _termination_start_time_ms;
1226 // true when the task is during a concurrent phase, false when it is
1227 // in the remark phase (so, in the latter case, we do not have to
1228 // check all the things that we have to check during the concurrent
1229 // phase, i.e. SATB buffer availability...)
1230 bool _concurrent;
1232 TruncatedSeq _marking_step_diffs_ms;
1234 // Counting data structures. Embedding the task's marked_bytes_array
1235 // and card bitmap into the actual task saves having to go through
1236 // the ConcurrentMark object.
1237 size_t* _marked_bytes_array;
1238 BitMap* _card_bm;
1240 // LOTS of statistics related with this task
1241 #if _MARKING_STATS_
1242 NumberSeq _all_clock_intervals_ms;
1243 double _interval_start_time_ms;
1245 int _aborted;
1246 int _aborted_overflow;
1247 int _aborted_cm_aborted;
1248 int _aborted_yield;
1249 int _aborted_timed_out;
1250 int _aborted_satb;
1251 int _aborted_termination;
1253 int _steal_attempts;
1254 int _steals;
1256 int _clock_due_to_marking;
1257 int _clock_due_to_scanning;
1259 int _local_pushes;
1260 int _local_pops;
1261 int _local_max_size;
1262 int _objs_scanned;
1264 int _global_pushes;
1265 int _global_pops;
1266 int _global_max_size;
1268 int _global_transfers_to;
1269 int _global_transfers_from;
1271 int _region_stack_pops;
1273 int _regions_claimed;
1274 int _objs_found_on_bitmap;
1276 int _satb_buffers_processed;
1277 #endif // _MARKING_STATS_
1279 // it updates the local fields after this task has claimed
1280 // a new region to scan
1281 void setup_for_region(HeapRegion* hr);
1282 // it brings up-to-date the limit of the region
1283 void update_region_limit();
1285 // called when either the words scanned or the refs visited limit
1286 // has been reached
1287 void reached_limit();
1288 // recalculates the words scanned and refs visited limits
1289 void recalculate_limits();
1290 // decreases the words scanned and refs visited limits when we reach
1291 // an expensive operation
1292 void decrease_limits();
1293 // it checks whether the words scanned or refs visited reached their
1294 // respective limit and calls reached_limit() if they have
1295 void check_limits() {
1296 if (_words_scanned >= _words_scanned_limit ||
1297 _refs_reached >= _refs_reached_limit) {
1298 reached_limit();
1299 }
1300 }
1301 // this is supposed to be called regularly during a marking step as
1302 // it checks a bunch of conditions that might cause the marking step
1303 // to abort
1304 void regular_clock_call();
1305 bool concurrent() { return _concurrent; }
1307 public:
1308 // It resets the task; it should be called right at the beginning of
1309 // a marking phase.
1310 void reset(CMBitMap* _nextMarkBitMap);
1311 // it clears all the fields that correspond to a claimed region.
1312 void clear_region_fields();
1314 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1316 // The main method of this class which performs a marking step
1317 // trying not to exceed the given duration. However, it might exit
1318 // prematurely, according to some conditions (i.e. SATB buffers are
1319 // available for processing).
1320 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1322 // These two calls start and stop the timer
1323 void record_start_time() {
1324 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1325 }
1326 void record_end_time() {
1327 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1328 }
1330 // returns the task ID
1331 int task_id() { return _task_id; }
1333 // From TerminatorTerminator. It determines whether this task should
1334 // exit the termination protocol after it's entered it.
1335 virtual bool should_exit_termination();
1337 // Resets the local region fields after a task has finished scanning a
1338 // region; or when they have become stale as a result of the region
1339 // being evacuated.
1340 void giveup_current_region();
1342 HeapWord* finger() { return _finger; }
1344 bool has_aborted() { return _has_aborted; }
1345 void set_has_aborted() { _has_aborted = true; }
1346 void clear_has_aborted() { _has_aborted = false; }
1347 bool has_timed_out() { return _has_timed_out; }
1348 bool claimed() { return _claimed; }
1350 // Support routines for the partially scanned region that may be
1351 // recorded as a result of aborting while draining the CMRegionStack
1352 MemRegion aborted_region() { return _aborted_region; }
1353 void set_aborted_region(MemRegion mr)
1354 { _aborted_region = mr; }
1356 // Clears any recorded partially scanned region
1357 void clear_aborted_region() { set_aborted_region(MemRegion()); }
1359 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1361 // It grays the object by marking it and, if necessary, pushing it
1362 // on the local queue
1363 inline void deal_with_reference(oop obj);
1365 // It scans an object and visits its children.
1366 void scan_object(oop obj);
1368 // It pushes an object on the local queue.
1369 inline void push(oop obj);
1371 // These two move entries to/from the global stack.
1372 void move_entries_to_global_stack();
1373 void get_entries_from_global_stack();
1375 // It pops and scans objects from the local queue. If partially is
1376 // true, then it stops when the queue size is of a given limit. If
1377 // partially is false, then it stops when the queue is empty.
1378 void drain_local_queue(bool partially);
1379 // It moves entries from the global stack to the local queue and
1380 // drains the local queue. If partially is true, then it stops when
1381 // both the global stack and the local queue reach a given size. If
1382 // partially if false, it tries to empty them totally.
1383 void drain_global_stack(bool partially);
1384 // It keeps picking SATB buffers and processing them until no SATB
1385 // buffers are available.
1386 void drain_satb_buffers();
1388 // It keeps popping regions from the region stack and processing
1389 // them until the region stack is empty.
1390 void drain_region_stack(BitMapClosure* closure);
1392 // moves the local finger to a new location
1393 inline void move_finger_to(HeapWord* new_finger) {
1394 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1395 _finger = new_finger;
1396 }
1398 // moves the region finger to a new location
1399 inline void move_region_finger_to(HeapWord* new_finger) {
1400 assert(new_finger < _cm->finger(), "invariant");
1401 _region_finger = new_finger;
1402 }
1404 CMTask(int task_num, ConcurrentMark *cm,
1405 size_t* marked_bytes, BitMap* card_bm,
1406 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1408 // it prints statistics associated with this task
1409 void print_stats();
1411 #if _MARKING_STATS_
1412 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1413 #endif // _MARKING_STATS_
1414 };
1416 // Class that's used to to print out per-region liveness
1417 // information. It's currently used at the end of marking and also
1418 // after we sort the old regions at the end of the cleanup operation.
1419 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1420 private:
1421 outputStream* _out;
1423 // Accumulators for these values.
1424 size_t _total_used_bytes;
1425 size_t _total_capacity_bytes;
1426 size_t _total_prev_live_bytes;
1427 size_t _total_next_live_bytes;
1429 // These are set up when we come across a "stars humongous" region
1430 // (as this is where most of this information is stored, not in the
1431 // subsequent "continues humongous" regions). After that, for every
1432 // region in a given humongous region series we deduce the right
1433 // values for it by simply subtracting the appropriate amount from
1434 // these fields. All these values should reach 0 after we've visited
1435 // the last region in the series.
1436 size_t _hum_used_bytes;
1437 size_t _hum_capacity_bytes;
1438 size_t _hum_prev_live_bytes;
1439 size_t _hum_next_live_bytes;
1441 static double perc(size_t val, size_t total) {
1442 if (total == 0) {
1443 return 0.0;
1444 } else {
1445 return 100.0 * ((double) val / (double) total);
1446 }
1447 }
1449 static double bytes_to_mb(size_t val) {
1450 return (double) val / (double) M;
1451 }
1453 // See the .cpp file.
1454 size_t get_hum_bytes(size_t* hum_bytes);
1455 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1456 size_t* prev_live_bytes, size_t* next_live_bytes);
1458 public:
1459 // The header and footer are printed in the constructor and
1460 // destructor respectively.
1461 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1462 virtual bool doHeapRegion(HeapRegion* r);
1463 ~G1PrintRegionLivenessInfoClosure();
1464 };
1466 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP