Tue, 14 Jun 2011 10:33:43 -0400
7046558: G1: concurrent marking optimizations
Summary: Some optimizations to improve the concurrent marking phase: specialize the main oop closure, make sure a few methods in the fast path are properly inlined, a few more bits and pieces, and some cosmetic fixes.
Reviewed-by: stefank, johnc
1 /*
2 * Copyright (c) 2001, 2011, 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 bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }
88 bool iterate(BitMapClosure* cl, MemRegion mr);
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 _oops_do_bound; // Number of elements to include in next iteration.
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 size; a subsequent "oops_do" will iterate only over
251 // indices valid at the time of this call.
252 void set_oops_do_bound(jint bound = -1) {
253 if (bound == -1) {
254 _oops_do_bound = _index;
255 } else {
256 _oops_do_bound = bound;
257 }
258 }
259 jint oops_do_bound() { return _oops_do_bound; }
260 // iterate over the oops in the mark stack, up to the bound recorded via
261 // the call above.
262 void oops_do(OopClosure* f);
263 };
265 class CMRegionStack VALUE_OBJ_CLASS_SPEC {
266 MemRegion* _base;
267 jint _capacity;
268 jint _index;
269 jint _oops_do_bound;
270 bool _overflow;
271 public:
272 CMRegionStack();
273 ~CMRegionStack();
274 void allocate(size_t size);
276 // This is lock-free; assumes that it will only be called in parallel
277 // with other "push" operations (no pops).
278 void push_lock_free(MemRegion mr);
280 // Lock-free; assumes that it will only be called in parallel
281 // with other "pop" operations (no pushes).
282 MemRegion pop_lock_free();
284 #if 0
285 // The routines that manipulate the region stack with a lock are
286 // not currently used. They should be retained, however, as a
287 // diagnostic aid.
289 // These two are the implementations that use a lock. They can be
290 // called concurrently with each other but they should not be called
291 // concurrently with the lock-free versions (push() / pop()).
292 void push_with_lock(MemRegion mr);
293 MemRegion pop_with_lock();
294 #endif
296 bool isEmpty() { return _index == 0; }
297 bool isFull() { return _index == _capacity; }
299 bool overflow() { return _overflow; }
300 void clear_overflow() { _overflow = false; }
302 int size() { return _index; }
304 // It iterates over the entries in the region stack and it
305 // invalidates (i.e. assigns MemRegion()) the ones that point to
306 // regions in the collection set.
307 bool invalidate_entries_into_cset();
309 // This gives an upper bound up to which the iteration in
310 // invalidate_entries_into_cset() will reach. This prevents
311 // newly-added entries to be unnecessarily scanned.
312 void set_oops_do_bound() {
313 _oops_do_bound = _index;
314 }
316 void setEmpty() { _index = 0; clear_overflow(); }
317 };
319 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
320 private:
321 #ifndef PRODUCT
322 uintx _num_remaining;
323 bool _force;
324 #endif // !defined(PRODUCT)
326 public:
327 void init() PRODUCT_RETURN;
328 void update() PRODUCT_RETURN;
329 bool should_force() PRODUCT_RETURN_( return false; );
330 };
332 // this will enable a variety of different statistics per GC task
333 #define _MARKING_STATS_ 0
334 // this will enable the higher verbose levels
335 #define _MARKING_VERBOSE_ 0
337 #if _MARKING_STATS_
338 #define statsOnly(statement) \
339 do { \
340 statement ; \
341 } while (0)
342 #else // _MARKING_STATS_
343 #define statsOnly(statement) \
344 do { \
345 } while (0)
346 #endif // _MARKING_STATS_
348 typedef enum {
349 no_verbose = 0, // verbose turned off
350 stats_verbose, // only prints stats at the end of marking
351 low_verbose, // low verbose, mostly per region and per major event
352 medium_verbose, // a bit more detailed than low
353 high_verbose // per object verbose
354 } CMVerboseLevel;
357 class ConcurrentMarkThread;
359 class ConcurrentMark: public CHeapObj {
360 friend class ConcurrentMarkThread;
361 friend class CMTask;
362 friend class CMBitMapClosure;
363 friend class CSMarkOopClosure;
364 friend class CMGlobalObjectClosure;
365 friend class CMRemarkTask;
366 friend class CMConcurrentMarkingTask;
367 friend class G1ParNoteEndTask;
368 friend class CalcLiveObjectsClosure;
369 friend class G1RefProcTaskProxy;
370 friend class G1RefProcTaskExecutor;
371 friend class G1CMParKeepAliveAndDrainClosure;
372 friend class G1CMParDrainMarkingStackClosure;
374 protected:
375 ConcurrentMarkThread* _cmThread; // the thread doing the work
376 G1CollectedHeap* _g1h; // the heap.
377 size_t _parallel_marking_threads; // the number of marking
378 // threads we'll use
379 double _sleep_factor; // how much we have to sleep, with
380 // respect to the work we just did, to
381 // meet the marking overhead goal
382 double _marking_task_overhead; // marking target overhead for
383 // a single task
385 // same as the two above, but for the cleanup task
386 double _cleanup_sleep_factor;
387 double _cleanup_task_overhead;
389 FreeRegionList _cleanup_list;
391 // CMS marking support structures
392 CMBitMap _markBitMap1;
393 CMBitMap _markBitMap2;
394 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
395 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
396 bool _at_least_one_mark_complete;
398 BitMap _region_bm;
399 BitMap _card_bm;
401 // Heap bounds
402 HeapWord* _heap_start;
403 HeapWord* _heap_end;
405 // For gray objects
406 CMMarkStack _markStack; // Grey objects behind global finger.
407 CMRegionStack _regionStack; // Grey regions behind global finger.
408 HeapWord* volatile _finger; // the global finger, region aligned,
409 // always points to the end of the
410 // last claimed region
412 // marking tasks
413 size_t _max_task_num; // maximum task number
414 size_t _active_tasks; // task num currently active
415 CMTask** _tasks; // task queue array (max_task_num len)
416 CMTaskQueueSet* _task_queues; // task queue set
417 ParallelTaskTerminator _terminator; // for termination
419 // Two sync barriers that are used to synchronise tasks when an
420 // overflow occurs. The algorithm is the following. All tasks enter
421 // the first one to ensure that they have all stopped manipulating
422 // the global data structures. After they exit it, they re-initialise
423 // their data structures and task 0 re-initialises the global data
424 // structures. Then, they enter the second sync barrier. This
425 // ensure, that no task starts doing work before all data
426 // structures (local and global) have been re-initialised. When they
427 // exit it, they are free to start working again.
428 WorkGangBarrierSync _first_overflow_barrier_sync;
429 WorkGangBarrierSync _second_overflow_barrier_sync;
432 // this is set by any task, when an overflow on the global data
433 // structures is detected.
434 volatile bool _has_overflown;
435 // true: marking is concurrent, false: we're in remark
436 volatile bool _concurrent;
437 // set at the end of a Full GC so that marking aborts
438 volatile bool _has_aborted;
440 // used when remark aborts due to an overflow to indicate that
441 // another concurrent marking phase should start
442 volatile bool _restart_for_overflow;
444 // This is true from the very start of concurrent marking until the
445 // point when all the tasks complete their work. It is really used
446 // to determine the points between the end of concurrent marking and
447 // time of remark.
448 volatile bool _concurrent_marking_in_progress;
450 // verbose level
451 CMVerboseLevel _verbose_level;
453 // These two fields are used to implement the optimisation that
454 // avoids pushing objects on the global/region stack if there are
455 // no collection set regions above the lowest finger.
457 // This is the lowest finger (among the global and local fingers),
458 // which is calculated before a new collection set is chosen.
459 HeapWord* _min_finger;
460 // If this flag is true, objects/regions that are marked below the
461 // finger should be pushed on the stack(s). If this is flag is
462 // false, it is safe not to push them on the stack(s).
463 bool _should_gray_objects;
465 // All of these times are in ms.
466 NumberSeq _init_times;
467 NumberSeq _remark_times;
468 NumberSeq _remark_mark_times;
469 NumberSeq _remark_weak_ref_times;
470 NumberSeq _cleanup_times;
471 double _total_counting_time;
472 double _total_rs_scrub_time;
474 double* _accum_task_vtime; // accumulated task vtime
476 WorkGang* _parallel_workers;
478 ForceOverflowSettings _force_overflow_conc;
479 ForceOverflowSettings _force_overflow_stw;
481 void weakRefsWork(bool clear_all_soft_refs);
483 void swapMarkBitMaps();
485 // It resets the global marking data structures, as well as the
486 // task local ones; should be called during initial mark.
487 void reset();
488 // It resets all the marking data structures.
489 void clear_marking_state(bool clear_overflow = true);
491 // It should be called to indicate which phase we're in (concurrent
492 // mark or remark) and how many threads are currently active.
493 void set_phase(size_t active_tasks, bool concurrent);
494 // We do this after we're done with marking so that the marking data
495 // structures are initialised to a sensible and predictable state.
496 void set_non_marking_state();
498 // prints all gathered CM-related statistics
499 void print_stats();
501 bool cleanup_list_is_empty() {
502 return _cleanup_list.is_empty();
503 }
505 // accessor methods
506 size_t parallel_marking_threads() { return _parallel_marking_threads; }
507 double sleep_factor() { return _sleep_factor; }
508 double marking_task_overhead() { return _marking_task_overhead;}
509 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
510 double cleanup_task_overhead() { return _cleanup_task_overhead;}
512 HeapWord* finger() { return _finger; }
513 bool concurrent() { return _concurrent; }
514 size_t active_tasks() { return _active_tasks; }
515 ParallelTaskTerminator* terminator() { return &_terminator; }
517 // It claims the next available region to be scanned by a marking
518 // task. It might return NULL if the next region is empty or we have
519 // run out of regions. In the latter case, out_of_regions()
520 // determines whether we've really run out of regions or the task
521 // should call claim_region() again. This might seem a bit
522 // awkward. Originally, the code was written so that claim_region()
523 // either successfully returned with a non-empty region or there
524 // were no more regions to be claimed. The problem with this was
525 // that, in certain circumstances, it iterated over large chunks of
526 // the heap finding only empty regions and, while it was working, it
527 // was preventing the calling task to call its regular clock
528 // method. So, this way, each task will spend very little time in
529 // claim_region() and is allowed to call the regular clock method
530 // frequently.
531 HeapRegion* claim_region(int task);
533 // It determines whether we've run out of regions to scan.
534 bool out_of_regions() { return _finger == _heap_end; }
536 // Returns the task with the given id
537 CMTask* task(int id) {
538 assert(0 <= id && id < (int) _active_tasks,
539 "task id not within active bounds");
540 return _tasks[id];
541 }
543 // Returns the task queue with the given id
544 CMTaskQueue* task_queue(int id) {
545 assert(0 <= id && id < (int) _active_tasks,
546 "task queue id not within active bounds");
547 return (CMTaskQueue*) _task_queues->queue(id);
548 }
550 // Returns the task queue set
551 CMTaskQueueSet* task_queues() { return _task_queues; }
553 // Access / manipulation of the overflow flag which is set to
554 // indicate that the global stack or region stack has overflown
555 bool has_overflown() { return _has_overflown; }
556 void set_has_overflown() { _has_overflown = true; }
557 void clear_has_overflown() { _has_overflown = false; }
559 bool has_aborted() { return _has_aborted; }
560 bool restart_for_overflow() { return _restart_for_overflow; }
562 // Methods to enter the two overflow sync barriers
563 void enter_first_sync_barrier(int task_num);
564 void enter_second_sync_barrier(int task_num);
566 ForceOverflowSettings* force_overflow_conc() {
567 return &_force_overflow_conc;
568 }
570 ForceOverflowSettings* force_overflow_stw() {
571 return &_force_overflow_stw;
572 }
574 ForceOverflowSettings* force_overflow() {
575 if (concurrent()) {
576 return force_overflow_conc();
577 } else {
578 return force_overflow_stw();
579 }
580 }
582 public:
583 // Manipulation of the global mark stack.
584 // Notice that the first mark_stack_push is CAS-based, whereas the
585 // two below are Mutex-based. This is OK since the first one is only
586 // called during evacuation pauses and doesn't compete with the
587 // other two (which are called by the marking tasks during
588 // concurrent marking or remark).
589 bool mark_stack_push(oop p) {
590 _markStack.par_push(p);
591 if (_markStack.overflow()) {
592 set_has_overflown();
593 return false;
594 }
595 return true;
596 }
597 bool mark_stack_push(oop* arr, int n) {
598 _markStack.par_push_arr(arr, n);
599 if (_markStack.overflow()) {
600 set_has_overflown();
601 return false;
602 }
603 return true;
604 }
605 void mark_stack_pop(oop* arr, int max, int* n) {
606 _markStack.par_pop_arr(arr, max, n);
607 }
608 size_t mark_stack_size() { return _markStack.size(); }
609 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
610 bool mark_stack_overflow() { return _markStack.overflow(); }
611 bool mark_stack_empty() { return _markStack.isEmpty(); }
613 // (Lock-free) Manipulation of the region stack
614 bool region_stack_push_lock_free(MemRegion mr) {
615 // Currently we only call the lock-free version during evacuation
616 // pauses.
617 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
619 _regionStack.push_lock_free(mr);
620 if (_regionStack.overflow()) {
621 set_has_overflown();
622 return false;
623 }
624 return true;
625 }
627 // Lock-free version of region-stack pop. Should only be
628 // called in tandem with other lock-free pops.
629 MemRegion region_stack_pop_lock_free() {
630 return _regionStack.pop_lock_free();
631 }
633 #if 0
634 // The routines that manipulate the region stack with a lock are
635 // not currently used. They should be retained, however, as a
636 // diagnostic aid.
638 bool region_stack_push_with_lock(MemRegion mr) {
639 // Currently we only call the lock-based version during either
640 // concurrent marking or remark.
641 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
642 "if we are at a safepoint it should be the remark safepoint");
644 _regionStack.push_with_lock(mr);
645 if (_regionStack.overflow()) {
646 set_has_overflown();
647 return false;
648 }
649 return true;
650 }
652 MemRegion region_stack_pop_with_lock() {
653 // Currently we only call the lock-based version during either
654 // concurrent marking or remark.
655 assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
656 "if we are at a safepoint it should be the remark safepoint");
658 return _regionStack.pop_with_lock();
659 }
660 #endif
662 int region_stack_size() { return _regionStack.size(); }
663 bool region_stack_overflow() { return _regionStack.overflow(); }
664 bool region_stack_empty() { return _regionStack.isEmpty(); }
666 // Iterate over any regions that were aborted while draining the
667 // region stack (any such regions are saved in the corresponding
668 // CMTask) and invalidate (i.e. assign to the empty MemRegion())
669 // any regions that point into the collection set.
670 bool invalidate_aborted_regions_in_cset();
672 // Returns true if there are any aborted memory regions.
673 bool has_aborted_regions();
675 bool concurrent_marking_in_progress() {
676 return _concurrent_marking_in_progress;
677 }
678 void set_concurrent_marking_in_progress() {
679 _concurrent_marking_in_progress = true;
680 }
681 void clear_concurrent_marking_in_progress() {
682 _concurrent_marking_in_progress = false;
683 }
685 void update_accum_task_vtime(int i, double vtime) {
686 _accum_task_vtime[i] += vtime;
687 }
689 double all_task_accum_vtime() {
690 double ret = 0.0;
691 for (int i = 0; i < (int)_max_task_num; ++i)
692 ret += _accum_task_vtime[i];
693 return ret;
694 }
696 // Attempts to steal an object from the task queues of other tasks
697 bool try_stealing(int task_num, int* hash_seed, oop& obj) {
698 return _task_queues->steal(task_num, hash_seed, obj);
699 }
701 // It grays an object by first marking it. Then, if it's behind the
702 // global finger, it also pushes it on the global stack.
703 void deal_with_reference(oop obj);
705 ConcurrentMark(ReservedSpace rs, int max_regions);
706 ~ConcurrentMark();
707 ConcurrentMarkThread* cmThread() { return _cmThread; }
709 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
710 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
712 // The following three are interaction between CM and
713 // G1CollectedHeap
715 // This notifies CM that a root during initial-mark needs to be
716 // grayed and it's MT-safe. Currently, we just mark it. But, in the
717 // future, we can experiment with pushing it on the stack and we can
718 // do this without changing G1CollectedHeap.
719 void grayRoot(oop p);
720 // It's used during evacuation pauses to gray a region, if
721 // necessary, and it's MT-safe. It assumes that the caller has
722 // marked any objects on that region. If _should_gray_objects is
723 // true and we're still doing concurrent marking, the region is
724 // pushed on the region stack, if it is located below the global
725 // finger, otherwise we do nothing.
726 void grayRegionIfNecessary(MemRegion mr);
727 // It's used during evacuation pauses to mark and, if necessary,
728 // gray a single object and it's MT-safe. It assumes the caller did
729 // not mark the object. If _should_gray_objects is true and we're
730 // still doing concurrent marking, the objects is pushed on the
731 // global stack, if it is located below the global finger, otherwise
732 // we do nothing.
733 void markAndGrayObjectIfNecessary(oop p);
735 // It iterates over the heap and for each object it comes across it
736 // will dump the contents of its reference fields, as well as
737 // liveness information for the object and its referents. The dump
738 // will be written to a file with the following name:
739 // G1PrintReachableBaseFile + "." + str. use_prev_marking decides
740 // whether the prev (use_prev_marking == true) or next
741 // (use_prev_marking == false) marking information will be used to
742 // determine the liveness of each object / referent. If all is true,
743 // all objects in the heap will be dumped, otherwise only the live
744 // ones. In the dump the following symbols / abbreviations are used:
745 // M : an explicitly live object (its bitmap bit is set)
746 // > : an implicitly live object (over tams)
747 // O : an object outside the G1 heap (typically: in the perm gen)
748 // NOT : a reference field whose referent is not live
749 // AND MARKED : indicates that an object is both explicitly and
750 // implicitly live (it should be one or the other, not both)
751 void print_reachable(const char* str,
752 bool use_prev_marking, bool all) PRODUCT_RETURN;
754 // Clear the next marking bitmap (will be called concurrently).
755 void clearNextBitmap();
757 // main CMS steps and related support
758 void checkpointRootsInitial();
760 // These two do the work that needs to be done before and after the
761 // initial root checkpoint. Since this checkpoint can be done at two
762 // different points (i.e. an explicit pause or piggy-backed on a
763 // young collection), then it's nice to be able to easily share the
764 // pre/post code. It might be the case that we can put everything in
765 // the post method. TP
766 void checkpointRootsInitialPre();
767 void checkpointRootsInitialPost();
769 // Do concurrent phase of marking, to a tentative transitive closure.
770 void markFromRoots();
772 // Process all unprocessed SATB buffers. It is called at the
773 // beginning of an evacuation pause.
774 void drainAllSATBBuffers();
776 void checkpointRootsFinal(bool clear_all_soft_refs);
777 void checkpointRootsFinalWork();
778 void calcDesiredRegions();
779 void cleanup();
780 void completeCleanup();
782 // Mark in the previous bitmap. NB: this is usually read-only, so use
783 // this carefully!
784 void markPrev(oop p);
785 void clear(oop p);
786 // Clears marks for all objects in the given range, for both prev and
787 // next bitmaps. NB: the previous bitmap is usually read-only, so use
788 // this carefully!
789 void clearRangeBothMaps(MemRegion mr);
791 // Record the current top of the mark and region stacks; a
792 // subsequent oops_do() on the mark stack and
793 // invalidate_entries_into_cset() on the region stack will iterate
794 // only over indices valid at the time of this call.
795 void set_oops_do_bound() {
796 _markStack.set_oops_do_bound();
797 _regionStack.set_oops_do_bound();
798 }
799 // Iterate over the oops in the mark stack and all local queues. It
800 // also calls invalidate_entries_into_cset() on the region stack.
801 void oops_do(OopClosure* f);
802 // It is called at the end of an evacuation pause during marking so
803 // that CM is notified of where the new end of the heap is. It
804 // doesn't do anything if concurrent_marking_in_progress() is false,
805 // unless the force parameter is true.
806 void update_g1_committed(bool force = false);
808 void complete_marking_in_collection_set();
810 // It indicates that a new collection set is being chosen.
811 void newCSet();
813 // It registers a collection set heap region with CM. This is used
814 // to determine whether any heap regions are located above the finger.
815 void registerCSetRegion(HeapRegion* hr);
817 // Resets the region fields of any active CMTask whose region fields
818 // are in the collection set (i.e. the region currently claimed by
819 // the CMTask will be evacuated and may be used, subsequently, as
820 // an alloc region). When this happens the region fields in the CMTask
821 // are stale and, hence, should be cleared causing the worker thread
822 // to claim a new region.
823 void reset_active_task_region_fields_in_cset();
825 // Registers the maximum region-end associated with a set of
826 // regions with CM. Again this is used to determine whether any
827 // heap regions are located above the finger.
828 void register_collection_set_finger(HeapWord* max_finger) {
829 // max_finger is the highest heap region end of the regions currently
830 // contained in the collection set. If this value is larger than
831 // _min_finger then we need to gray objects.
832 // This routine is like registerCSetRegion but for an entire
833 // collection of regions.
834 if (max_finger > _min_finger)
835 _should_gray_objects = true;
836 }
838 // Returns "true" if at least one mark has been completed.
839 bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
841 bool isMarked(oop p) const {
842 assert(p != NULL && p->is_oop(), "expected an oop");
843 HeapWord* addr = (HeapWord*)p;
844 assert(addr >= _nextMarkBitMap->startWord() ||
845 addr < _nextMarkBitMap->endWord(), "in a region");
847 return _nextMarkBitMap->isMarked(addr);
848 }
850 inline bool not_yet_marked(oop p) const;
852 // XXX Debug code
853 bool containing_card_is_marked(void* p);
854 bool containing_cards_are_marked(void* start, void* last);
856 bool isPrevMarked(oop p) const {
857 assert(p != NULL && p->is_oop(), "expected an oop");
858 HeapWord* addr = (HeapWord*)p;
859 assert(addr >= _prevMarkBitMap->startWord() ||
860 addr < _prevMarkBitMap->endWord(), "in a region");
862 return _prevMarkBitMap->isMarked(addr);
863 }
865 inline bool do_yield_check(int worker_i = 0);
866 inline bool should_yield();
868 // Called to abort the marking cycle after a Full GC takes palce.
869 void abort();
871 // This prints the global/local fingers. It is used for debugging.
872 NOT_PRODUCT(void print_finger();)
874 void print_summary_info();
876 void print_worker_threads_on(outputStream* st) const;
878 // The following indicate whether a given verbose level has been
879 // set. Notice that anything above stats is conditional to
880 // _MARKING_VERBOSE_ having been set to 1
881 bool verbose_stats()
882 { return _verbose_level >= stats_verbose; }
883 bool verbose_low()
884 { return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; }
885 bool verbose_medium()
886 { return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; }
887 bool verbose_high()
888 { return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; }
889 };
891 // A class representing a marking task.
892 class CMTask : public TerminatorTerminator {
893 private:
894 enum PrivateConstants {
895 // the regular clock call is called once the scanned words reaches
896 // this limit
897 words_scanned_period = 12*1024,
898 // the regular clock call is called once the number of visited
899 // references reaches this limit
900 refs_reached_period = 384,
901 // initial value for the hash seed, used in the work stealing code
902 init_hash_seed = 17,
903 // how many entries will be transferred between global stack and
904 // local queues
905 global_stack_transfer_size = 16
906 };
908 int _task_id;
909 G1CollectedHeap* _g1h;
910 ConcurrentMark* _cm;
911 CMBitMap* _nextMarkBitMap;
912 // the task queue of this task
913 CMTaskQueue* _task_queue;
914 private:
915 // the task queue set---needed for stealing
916 CMTaskQueueSet* _task_queues;
917 // indicates whether the task has been claimed---this is only for
918 // debugging purposes
919 bool _claimed;
921 // number of calls to this task
922 int _calls;
924 // when the virtual timer reaches this time, the marking step should
925 // exit
926 double _time_target_ms;
927 // the start time of the current marking step
928 double _start_time_ms;
930 // the oop closure used for iterations over oops
931 G1CMOopClosure* _cm_oop_closure;
933 // the region this task is scanning, NULL if we're not scanning any
934 HeapRegion* _curr_region;
935 // the local finger of this task, NULL if we're not scanning a region
936 HeapWord* _finger;
937 // limit of the region this task is scanning, NULL if we're not scanning one
938 HeapWord* _region_limit;
940 // This is used only when we scan regions popped from the region
941 // stack. It records what the last object on such a region we
942 // scanned was. It is used to ensure that, if we abort region
943 // iteration, we do not rescan the first part of the region. This
944 // should be NULL when we're not scanning a region from the region
945 // stack.
946 HeapWord* _region_finger;
948 // If we abort while scanning a region we record the remaining
949 // unscanned portion and check this field when marking restarts.
950 // This avoids having to push on the region stack while other
951 // marking threads may still be popping regions.
952 // If we were to push the unscanned portion directly to the
953 // region stack then we would need to using locking versions
954 // of the push and pop operations.
955 MemRegion _aborted_region;
957 // the number of words this task has scanned
958 size_t _words_scanned;
959 // When _words_scanned reaches this limit, the regular clock is
960 // called. Notice that this might be decreased under certain
961 // circumstances (i.e. when we believe that we did an expensive
962 // operation).
963 size_t _words_scanned_limit;
964 // the initial value of _words_scanned_limit (i.e. what it was
965 // before it was decreased).
966 size_t _real_words_scanned_limit;
968 // the number of references this task has visited
969 size_t _refs_reached;
970 // When _refs_reached reaches this limit, the regular clock is
971 // called. Notice this this might be decreased under certain
972 // circumstances (i.e. when we believe that we did an expensive
973 // operation).
974 size_t _refs_reached_limit;
975 // the initial value of _refs_reached_limit (i.e. what it was before
976 // it was decreased).
977 size_t _real_refs_reached_limit;
979 // used by the work stealing stuff
980 int _hash_seed;
981 // if this is true, then the task has aborted for some reason
982 bool _has_aborted;
983 // set when the task aborts because it has met its time quota
984 bool _has_timed_out;
985 // true when we're draining SATB buffers; this avoids the task
986 // aborting due to SATB buffers being available (as we're already
987 // dealing with them)
988 bool _draining_satb_buffers;
990 // number sequence of past step times
991 NumberSeq _step_times_ms;
992 // elapsed time of this task
993 double _elapsed_time_ms;
994 // termination time of this task
995 double _termination_time_ms;
996 // when this task got into the termination protocol
997 double _termination_start_time_ms;
999 // true when the task is during a concurrent phase, false when it is
1000 // in the remark phase (so, in the latter case, we do not have to
1001 // check all the things that we have to check during the concurrent
1002 // phase, i.e. SATB buffer availability...)
1003 bool _concurrent;
1005 TruncatedSeq _marking_step_diffs_ms;
1007 // LOTS of statistics related with this task
1008 #if _MARKING_STATS_
1009 NumberSeq _all_clock_intervals_ms;
1010 double _interval_start_time_ms;
1012 int _aborted;
1013 int _aborted_overflow;
1014 int _aborted_cm_aborted;
1015 int _aborted_yield;
1016 int _aborted_timed_out;
1017 int _aborted_satb;
1018 int _aborted_termination;
1020 int _steal_attempts;
1021 int _steals;
1023 int _clock_due_to_marking;
1024 int _clock_due_to_scanning;
1026 int _local_pushes;
1027 int _local_pops;
1028 int _local_max_size;
1029 int _objs_scanned;
1031 int _global_pushes;
1032 int _global_pops;
1033 int _global_max_size;
1035 int _global_transfers_to;
1036 int _global_transfers_from;
1038 int _region_stack_pops;
1040 int _regions_claimed;
1041 int _objs_found_on_bitmap;
1043 int _satb_buffers_processed;
1044 #endif // _MARKING_STATS_
1046 // it updates the local fields after this task has claimed
1047 // a new region to scan
1048 void setup_for_region(HeapRegion* hr);
1049 // it brings up-to-date the limit of the region
1050 void update_region_limit();
1052 // called when either the words scanned or the refs visited limit
1053 // has been reached
1054 void reached_limit();
1055 // recalculates the words scanned and refs visited limits
1056 void recalculate_limits();
1057 // decreases the words scanned and refs visited limits when we reach
1058 // an expensive operation
1059 void decrease_limits();
1060 // it checks whether the words scanned or refs visited reached their
1061 // respective limit and calls reached_limit() if they have
1062 void check_limits() {
1063 if (_words_scanned >= _words_scanned_limit ||
1064 _refs_reached >= _refs_reached_limit)
1065 reached_limit();
1066 }
1067 // this is supposed to be called regularly during a marking step as
1068 // it checks a bunch of conditions that might cause the marking step
1069 // to abort
1070 void regular_clock_call();
1071 bool concurrent() { return _concurrent; }
1073 public:
1074 // It resets the task; it should be called right at the beginning of
1075 // a marking phase.
1076 void reset(CMBitMap* _nextMarkBitMap);
1077 // it clears all the fields that correspond to a claimed region.
1078 void clear_region_fields();
1080 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1082 // The main method of this class which performs a marking step
1083 // trying not to exceed the given duration. However, it might exit
1084 // prematurely, according to some conditions (i.e. SATB buffers are
1085 // available for processing).
1086 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1088 // These two calls start and stop the timer
1089 void record_start_time() {
1090 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1091 }
1092 void record_end_time() {
1093 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1094 }
1096 // returns the task ID
1097 int task_id() { return _task_id; }
1099 // From TerminatorTerminator. It determines whether this task should
1100 // exit the termination protocol after it's entered it.
1101 virtual bool should_exit_termination();
1103 // Resets the local region fields after a task has finished scanning a
1104 // region; or when they have become stale as a result of the region
1105 // being evacuated.
1106 void giveup_current_region();
1108 HeapWord* finger() { return _finger; }
1110 bool has_aborted() { return _has_aborted; }
1111 void set_has_aborted() { _has_aborted = true; }
1112 void clear_has_aborted() { _has_aborted = false; }
1113 bool has_timed_out() { return _has_timed_out; }
1114 bool claimed() { return _claimed; }
1116 // Support routines for the partially scanned region that may be
1117 // recorded as a result of aborting while draining the CMRegionStack
1118 MemRegion aborted_region() { return _aborted_region; }
1119 void set_aborted_region(MemRegion mr)
1120 { _aborted_region = mr; }
1122 // Clears any recorded partially scanned region
1123 void clear_aborted_region() { set_aborted_region(MemRegion()); }
1125 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1127 // It grays the object by marking it and, if necessary, pushing it
1128 // on the local queue
1129 inline void deal_with_reference(oop obj);
1131 // It scans an object and visits its children.
1132 void scan_object(oop obj);
1134 // It pushes an object on the local queue.
1135 inline void push(oop obj);
1137 // These two move entries to/from the global stack.
1138 void move_entries_to_global_stack();
1139 void get_entries_from_global_stack();
1141 // It pops and scans objects from the local queue. If partially is
1142 // true, then it stops when the queue size is of a given limit. If
1143 // partially is false, then it stops when the queue is empty.
1144 void drain_local_queue(bool partially);
1145 // It moves entries from the global stack to the local queue and
1146 // drains the local queue. If partially is true, then it stops when
1147 // both the global stack and the local queue reach a given size. If
1148 // partially if false, it tries to empty them totally.
1149 void drain_global_stack(bool partially);
1150 // It keeps picking SATB buffers and processing them until no SATB
1151 // buffers are available.
1152 void drain_satb_buffers();
1153 // It keeps popping regions from the region stack and processing
1154 // them until the region stack is empty.
1155 void drain_region_stack(BitMapClosure* closure);
1157 // moves the local finger to a new location
1158 inline void move_finger_to(HeapWord* new_finger) {
1159 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1160 _finger = new_finger;
1161 }
1163 // moves the region finger to a new location
1164 inline void move_region_finger_to(HeapWord* new_finger) {
1165 assert(new_finger < _cm->finger(), "invariant");
1166 _region_finger = new_finger;
1167 }
1169 CMTask(int task_num, ConcurrentMark *cm,
1170 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1172 // it prints statistics associated with this task
1173 void print_stats();
1175 #if _MARKING_STATS_
1176 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1177 #endif // _MARKING_STATS_
1178 };
1180 // Class that's used to to print out per-region liveness
1181 // information. It's currently used at the end of marking and also
1182 // after we sort the old regions at the end of the cleanup operation.
1183 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1184 private:
1185 outputStream* _out;
1187 // Accumulators for these values.
1188 size_t _total_used_bytes;
1189 size_t _total_capacity_bytes;
1190 size_t _total_prev_live_bytes;
1191 size_t _total_next_live_bytes;
1193 // These are set up when we come across a "stars humongous" region
1194 // (as this is where most of this information is stored, not in the
1195 // subsequent "continues humongous" regions). After that, for every
1196 // region in a given humongous region series we deduce the right
1197 // values for it by simply subtracting the appropriate amount from
1198 // these fields. All these values should reach 0 after we've visited
1199 // the last region in the series.
1200 size_t _hum_used_bytes;
1201 size_t _hum_capacity_bytes;
1202 size_t _hum_prev_live_bytes;
1203 size_t _hum_next_live_bytes;
1205 static double perc(size_t val, size_t total) {
1206 if (total == 0) {
1207 return 0.0;
1208 } else {
1209 return 100.0 * ((double) val / (double) total);
1210 }
1211 }
1213 static double bytes_to_mb(size_t val) {
1214 return (double) val / (double) M;
1215 }
1217 // See the .cpp file.
1218 size_t get_hum_bytes(size_t* hum_bytes);
1219 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1220 size_t* prev_live_bytes, size_t* next_live_bytes);
1222 public:
1223 // The header and footer are printed in the constructor and
1224 // destructor respectively.
1225 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1226 virtual bool doHeapRegion(HeapRegion* r);
1227 ~G1PrintRegionLivenessInfoClosure();
1228 };
1230 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP