Tue, 19 Mar 2013 00:57:39 -0700
8009940: G1: assert(_finger == _heap_end) failed, concurrentMark.cpp:809
Summary: Skip reference processing if the global marking stack overflows during remark. Refactor and rename set_phase(); move code that sets the concurrency level into its own routine. Do not call set_phase() from within parallel reference processing; use the concurrency level routine instead. The marking state should only set reset by CMTask[0] during the concurrent phase of the marking cycle; if an overflow occurs at any stage during the remark, the marking state will be reset after reference processing.
Reviewed-by: brutisso, jmasa
1 /*
2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
28 #include "gc_implementation/g1/heapRegionSets.hpp"
29 #include "utilities/taskqueue.hpp"
31 class G1CollectedHeap;
32 class CMTask;
33 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue;
34 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
36 // Closure used by CM during concurrent reference discovery
37 // and reference processing (during remarking) to determine
38 // if a particular object is alive. It is primarily used
39 // to determine if referents of discovered reference objects
40 // are alive. An instance is also embedded into the
41 // reference processor as the _is_alive_non_header field
42 class G1CMIsAliveClosure: public BoolObjectClosure {
43 G1CollectedHeap* _g1;
44 public:
45 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
47 void do_object(oop obj) {
48 ShouldNotCallThis();
49 }
50 bool do_object_b(oop obj);
51 };
53 // A generic CM bit map. This is essentially a wrapper around the BitMap
54 // class, with one bit per (1<<_shifter) HeapWords.
56 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
57 protected:
58 HeapWord* _bmStartWord; // base address of range covered by map
59 size_t _bmWordSize; // map size (in #HeapWords covered)
60 const int _shifter; // map to char or bit
61 VirtualSpace _virtual_space; // underlying the bit map
62 BitMap _bm; // the bit map itself
64 public:
65 // constructor
66 CMBitMapRO(int shifter);
68 enum { do_yield = true };
70 // inquiries
71 HeapWord* startWord() const { return _bmStartWord; }
72 size_t sizeInWords() const { return _bmWordSize; }
73 // the following is one past the last word in space
74 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
76 // read marks
78 bool isMarked(HeapWord* addr) const {
79 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
80 "outside underlying space?");
81 return _bm.at(heapWordToOffset(addr));
82 }
84 // iteration
85 inline bool iterate(BitMapClosure* cl, MemRegion mr);
86 inline bool iterate(BitMapClosure* cl);
88 // Return the address corresponding to the next marked bit at or after
89 // "addr", and before "limit", if "limit" is non-NULL. If there is no
90 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
91 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
92 HeapWord* limit = NULL) const;
93 // Return the address corresponding to the next unmarked bit at or after
94 // "addr", and before "limit", if "limit" is non-NULL. If there is no
95 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
96 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
97 HeapWord* limit = NULL) const;
99 // conversion utilities
100 HeapWord* offsetToHeapWord(size_t offset) const {
101 return _bmStartWord + (offset << _shifter);
102 }
103 size_t heapWordToOffset(HeapWord* addr) const {
104 return pointer_delta(addr, _bmStartWord) >> _shifter;
105 }
106 int heapWordDiffToOffsetDiff(size_t diff) const;
108 // The argument addr should be the start address of a valid object
109 HeapWord* nextObject(HeapWord* addr) {
110 oop obj = (oop) addr;
111 HeapWord* res = addr + obj->size();
112 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
113 return res;
114 }
116 // debugging
117 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
118 };
120 class CMBitMap : public CMBitMapRO {
122 public:
123 // constructor
124 CMBitMap(int shifter) :
125 CMBitMapRO(shifter) {}
127 // Allocates the back store for the marking bitmap
128 bool allocate(ReservedSpace heap_rs);
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 ConcurrentMarking in the G1 collector.
166 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
167 VirtualSpace _virtual_space; // Underlying backing store for actual stack
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 bool _should_expand;
177 DEBUG_ONLY(bool _drain_in_progress;)
178 DEBUG_ONLY(bool _drain_in_progress_yields;)
180 public:
181 CMMarkStack(ConcurrentMark* cm);
182 ~CMMarkStack();
184 #ifndef PRODUCT
185 jint max_depth() const {
186 return _max_depth;
187 }
188 #endif
190 bool allocate(size_t capacity);
192 oop pop() {
193 if (!isEmpty()) {
194 return _base[--_index] ;
195 }
196 return NULL;
197 }
199 // If overflow happens, don't do the push, and record the overflow.
200 // *Requires* that "ptr" is already marked.
201 void push(oop ptr) {
202 if (isFull()) {
203 // Record overflow.
204 _overflow = true;
205 return;
206 } else {
207 _base[_index++] = ptr;
208 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
209 }
210 }
211 // Non-block impl. Note: concurrency is allowed only with other
212 // "par_push" operations, not with "pop" or "drain". We would need
213 // parallel versions of them if such concurrency was desired.
214 void par_push(oop ptr);
216 // Pushes the first "n" elements of "ptr_arr" on the stack.
217 // Non-block impl. Note: concurrency is allowed only with other
218 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
219 void par_adjoin_arr(oop* ptr_arr, int n);
221 // Pushes the first "n" elements of "ptr_arr" on the stack.
222 // Locking impl: concurrency is allowed only with
223 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
224 // locking strategy.
225 void par_push_arr(oop* ptr_arr, int n);
227 // If returns false, the array was empty. Otherwise, removes up to "max"
228 // elements from the stack, and transfers them to "ptr_arr" in an
229 // unspecified order. The actual number transferred is given in "n" ("n
230 // == 0" is deliberately redundant with the return value.) Locking impl:
231 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
232 // operations, which use the same locking strategy.
233 bool par_pop_arr(oop* ptr_arr, int max, int* n);
235 // Drain the mark stack, applying the given closure to all fields of
236 // objects on the stack. (That is, continue until the stack is empty,
237 // even if closure applications add entries to the stack.) The "bm"
238 // argument, if non-null, may be used to verify that only marked objects
239 // are on the mark stack. If "yield_after" is "true", then the
240 // concurrent marker performing the drain offers to yield after
241 // processing each object. If a yield occurs, stops the drain operation
242 // and returns false. Otherwise, returns true.
243 template<class OopClosureClass>
244 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
246 bool isEmpty() { return _index == 0; }
247 bool isFull() { return _index == _capacity; }
248 int maxElems() { return _capacity; }
250 bool overflow() { return _overflow; }
251 void clear_overflow() { _overflow = false; }
253 bool should_expand() const { return _should_expand; }
254 void set_should_expand();
256 // Expand the stack, typically in response to an overflow condition
257 void expand();
259 int size() { return _index; }
261 void setEmpty() { _index = 0; clear_overflow(); }
263 // Record the current index.
264 void note_start_of_gc();
266 // Make sure that we have not added any entries to the stack during GC.
267 void note_end_of_gc();
269 // iterate over the oops in the mark stack, up to the bound recorded via
270 // the call above.
271 void oops_do(OopClosure* f);
272 };
274 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
275 private:
276 #ifndef PRODUCT
277 uintx _num_remaining;
278 bool _force;
279 #endif // !defined(PRODUCT)
281 public:
282 void init() PRODUCT_RETURN;
283 void update() PRODUCT_RETURN;
284 bool should_force() PRODUCT_RETURN_( return false; );
285 };
287 // this will enable a variety of different statistics per GC task
288 #define _MARKING_STATS_ 0
289 // this will enable the higher verbose levels
290 #define _MARKING_VERBOSE_ 0
292 #if _MARKING_STATS_
293 #define statsOnly(statement) \
294 do { \
295 statement ; \
296 } while (0)
297 #else // _MARKING_STATS_
298 #define statsOnly(statement) \
299 do { \
300 } while (0)
301 #endif // _MARKING_STATS_
303 typedef enum {
304 no_verbose = 0, // verbose turned off
305 stats_verbose, // only prints stats at the end of marking
306 low_verbose, // low verbose, mostly per region and per major event
307 medium_verbose, // a bit more detailed than low
308 high_verbose // per object verbose
309 } CMVerboseLevel;
311 class YoungList;
313 // Root Regions are regions that are not empty at the beginning of a
314 // marking cycle and which we might collect during an evacuation pause
315 // while the cycle is active. Given that, during evacuation pauses, we
316 // do not copy objects that are explicitly marked, what we have to do
317 // for the root regions is to scan them and mark all objects reachable
318 // from them. According to the SATB assumptions, we only need to visit
319 // each object once during marking. So, as long as we finish this scan
320 // before the next evacuation pause, we can copy the objects from the
321 // root regions without having to mark them or do anything else to them.
322 //
323 // Currently, we only support root region scanning once (at the start
324 // of the marking cycle) and the root regions are all the survivor
325 // regions populated during the initial-mark pause.
326 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
327 private:
328 YoungList* _young_list;
329 ConcurrentMark* _cm;
331 volatile bool _scan_in_progress;
332 volatile bool _should_abort;
333 HeapRegion* volatile _next_survivor;
335 public:
336 CMRootRegions();
337 // We actually do most of the initialization in this method.
338 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
340 // Reset the claiming / scanning of the root regions.
341 void prepare_for_scan();
343 // Forces get_next() to return NULL so that the iteration aborts early.
344 void abort() { _should_abort = true; }
346 // Return true if the CM thread are actively scanning root regions,
347 // false otherwise.
348 bool scan_in_progress() { return _scan_in_progress; }
350 // Claim the next root region to scan atomically, or return NULL if
351 // all have been claimed.
352 HeapRegion* claim_next();
354 // Flag that we're done with root region scanning and notify anyone
355 // who's waiting on it. If aborted is false, assume that all regions
356 // have been claimed.
357 void scan_finished();
359 // If CM threads are still scanning root regions, wait until they
360 // are done. Return true if we had to wait, false otherwise.
361 bool wait_until_scan_finished();
362 };
364 class ConcurrentMarkThread;
366 class ConcurrentMark: public CHeapObj<mtGC> {
367 friend class CMMarkStack;
368 friend class ConcurrentMarkThread;
369 friend class CMTask;
370 friend class CMBitMapClosure;
371 friend class CMGlobalObjectClosure;
372 friend class CMRemarkTask;
373 friend class CMConcurrentMarkingTask;
374 friend class G1ParNoteEndTask;
375 friend class CalcLiveObjectsClosure;
376 friend class G1CMRefProcTaskProxy;
377 friend class G1CMRefProcTaskExecutor;
378 friend class G1CMKeepAliveAndDrainClosure;
379 friend class G1CMDrainMarkingStackClosure;
381 protected:
382 ConcurrentMarkThread* _cmThread; // the thread doing the work
383 G1CollectedHeap* _g1h; // the heap.
384 uint _parallel_marking_threads; // the number of marking
385 // threads we're use
386 uint _max_parallel_marking_threads; // max number of marking
387 // threads we'll ever use
388 double _sleep_factor; // how much we have to sleep, with
389 // respect to the work we just did, to
390 // meet the marking overhead goal
391 double _marking_task_overhead; // marking target overhead for
392 // a single task
394 // same as the two above, but for the cleanup task
395 double _cleanup_sleep_factor;
396 double _cleanup_task_overhead;
398 FreeRegionList _cleanup_list;
400 // Concurrent marking support structures
401 CMBitMap _markBitMap1;
402 CMBitMap _markBitMap2;
403 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
404 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
406 BitMap _region_bm;
407 BitMap _card_bm;
409 // Heap bounds
410 HeapWord* _heap_start;
411 HeapWord* _heap_end;
413 // Root region tracking and claiming.
414 CMRootRegions _root_regions;
416 // For gray objects
417 CMMarkStack _markStack; // Grey objects behind global finger.
418 HeapWord* volatile _finger; // the global finger, region aligned,
419 // always points to the end of the
420 // last claimed region
422 // marking tasks
423 uint _max_worker_id;// maximum worker id
424 uint _active_tasks; // task num currently active
425 CMTask** _tasks; // task queue array (max_worker_id len)
426 CMTaskQueueSet* _task_queues; // task queue set
427 ParallelTaskTerminator _terminator; // for termination
429 // Two sync barriers that are used to synchronise tasks when an
430 // overflow occurs. The algorithm is the following. All tasks enter
431 // the first one to ensure that they have all stopped manipulating
432 // the global data structures. After they exit it, they re-initialise
433 // their data structures and task 0 re-initialises the global data
434 // structures. Then, they enter the second sync barrier. This
435 // ensure, that no task starts doing work before all data
436 // structures (local and global) have been re-initialised. When they
437 // exit it, they are free to start working again.
438 WorkGangBarrierSync _first_overflow_barrier_sync;
439 WorkGangBarrierSync _second_overflow_barrier_sync;
441 // this is set by any task, when an overflow on the global data
442 // structures is detected.
443 volatile bool _has_overflown;
444 // true: marking is concurrent, false: we're in remark
445 volatile bool _concurrent;
446 // set at the end of a Full GC so that marking aborts
447 volatile bool _has_aborted;
449 // used when remark aborts due to an overflow to indicate that
450 // another concurrent marking phase should start
451 volatile bool _restart_for_overflow;
453 // This is true from the very start of concurrent marking until the
454 // point when all the tasks complete their work. It is really used
455 // to determine the points between the end of concurrent marking and
456 // time of remark.
457 volatile bool _concurrent_marking_in_progress;
459 // verbose level
460 CMVerboseLevel _verbose_level;
462 // All of these times are in ms.
463 NumberSeq _init_times;
464 NumberSeq _remark_times;
465 NumberSeq _remark_mark_times;
466 NumberSeq _remark_weak_ref_times;
467 NumberSeq _cleanup_times;
468 double _total_counting_time;
469 double _total_rs_scrub_time;
471 double* _accum_task_vtime; // accumulated task vtime
473 FlexibleWorkGang* _parallel_workers;
475 ForceOverflowSettings _force_overflow_conc;
476 ForceOverflowSettings _force_overflow_stw;
478 void weakRefsWork(bool clear_all_soft_refs);
480 void swapMarkBitMaps();
482 // It resets the global marking data structures, as well as the
483 // task local ones; should be called during initial mark.
484 void reset();
486 // Resets all the marking data structures. Called when we have to restart
487 // marking or when marking completes (via set_non_marking_state below).
488 void reset_marking_state(bool clear_overflow = true);
490 // We do this after we're done with marking so that the marking data
491 // structures are initialised to a sensible and predictable state.
492 void set_non_marking_state();
494 // Called to indicate how many threads are currently active.
495 void set_concurrency(uint active_tasks);
497 // It should be called to indicate which phase we're in (concurrent
498 // mark or remark) and how many threads are currently active.
499 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
501 // prints all gathered CM-related statistics
502 void print_stats();
504 bool cleanup_list_is_empty() {
505 return _cleanup_list.is_empty();
506 }
508 // accessor methods
509 uint parallel_marking_threads() const { return _parallel_marking_threads; }
510 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
511 double sleep_factor() { return _sleep_factor; }
512 double marking_task_overhead() { return _marking_task_overhead;}
513 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
514 double cleanup_task_overhead() { return _cleanup_task_overhead;}
516 bool use_parallel_marking_threads() const {
517 assert(parallel_marking_threads() <=
518 max_parallel_marking_threads(), "sanity");
519 assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
520 parallel_marking_threads() > 0,
521 "parallel workers not set up correctly");
522 return _parallel_workers != NULL;
523 }
525 HeapWord* finger() { return _finger; }
526 bool concurrent() { return _concurrent; }
527 uint active_tasks() { return _active_tasks; }
528 ParallelTaskTerminator* terminator() { return &_terminator; }
530 // It claims the next available region to be scanned by a marking
531 // task/thread. It might return NULL if the next region is empty or
532 // we have run out of regions. In the latter case, out_of_regions()
533 // determines whether we've really run out of regions or the task
534 // should call claim_region() again. This might seem a bit
535 // awkward. Originally, the code was written so that claim_region()
536 // either successfully returned with a non-empty region or there
537 // were no more regions to be claimed. The problem with this was
538 // that, in certain circumstances, it iterated over large chunks of
539 // the heap finding only empty regions and, while it was working, it
540 // was preventing the calling task to call its regular clock
541 // method. So, this way, each task will spend very little time in
542 // claim_region() and is allowed to call the regular clock method
543 // frequently.
544 HeapRegion* claim_region(uint worker_id);
546 // It determines whether we've run out of regions to scan.
547 bool out_of_regions() { return _finger == _heap_end; }
549 // Returns the task with the given id
550 CMTask* task(int id) {
551 assert(0 <= id && id < (int) _active_tasks,
552 "task id not within active bounds");
553 return _tasks[id];
554 }
556 // Returns the task queue with the given id
557 CMTaskQueue* task_queue(int id) {
558 assert(0 <= id && id < (int) _active_tasks,
559 "task queue id not within active bounds");
560 return (CMTaskQueue*) _task_queues->queue(id);
561 }
563 // Returns the task queue set
564 CMTaskQueueSet* task_queues() { return _task_queues; }
566 // Access / manipulation of the overflow flag which is set to
567 // indicate that the global stack has overflown
568 bool has_overflown() { return _has_overflown; }
569 void set_has_overflown() { _has_overflown = true; }
570 void clear_has_overflown() { _has_overflown = false; }
571 bool restart_for_overflow() { return _restart_for_overflow; }
573 bool has_aborted() { return _has_aborted; }
575 // Methods to enter the two overflow sync barriers
576 void enter_first_sync_barrier(uint worker_id);
577 void enter_second_sync_barrier(uint worker_id);
579 ForceOverflowSettings* force_overflow_conc() {
580 return &_force_overflow_conc;
581 }
583 ForceOverflowSettings* force_overflow_stw() {
584 return &_force_overflow_stw;
585 }
587 ForceOverflowSettings* force_overflow() {
588 if (concurrent()) {
589 return force_overflow_conc();
590 } else {
591 return force_overflow_stw();
592 }
593 }
595 // Live Data Counting data structures...
596 // These data structures are initialized at the start of
597 // marking. They are written to while marking is active.
598 // They are aggregated during remark; the aggregated values
599 // are then used to populate the _region_bm, _card_bm, and
600 // the total live bytes, which are then subsequently updated
601 // during cleanup.
603 // An array of bitmaps (one bit map per task). Each bitmap
604 // is used to record the cards spanned by the live objects
605 // marked by that task/worker.
606 BitMap* _count_card_bitmaps;
608 // Used to record the number of marked live bytes
609 // (for each region, by worker thread).
610 size_t** _count_marked_bytes;
612 // Card index of the bottom of the G1 heap. Used for biasing indices into
613 // the card bitmaps.
614 intptr_t _heap_bottom_card_num;
616 // Set to true when initialization is complete
617 bool _completed_initialization;
619 public:
620 // Manipulation of the global mark stack.
621 // Notice that the first mark_stack_push is CAS-based, whereas the
622 // two below are Mutex-based. This is OK since the first one is only
623 // called during evacuation pauses and doesn't compete with the
624 // other two (which are called by the marking tasks during
625 // concurrent marking or remark).
626 bool mark_stack_push(oop p) {
627 _markStack.par_push(p);
628 if (_markStack.overflow()) {
629 set_has_overflown();
630 return false;
631 }
632 return true;
633 }
634 bool mark_stack_push(oop* arr, int n) {
635 _markStack.par_push_arr(arr, n);
636 if (_markStack.overflow()) {
637 set_has_overflown();
638 return false;
639 }
640 return true;
641 }
642 void mark_stack_pop(oop* arr, int max, int* n) {
643 _markStack.par_pop_arr(arr, max, n);
644 }
645 size_t mark_stack_size() { return _markStack.size(); }
646 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
647 bool mark_stack_overflow() { return _markStack.overflow(); }
648 bool mark_stack_empty() { return _markStack.isEmpty(); }
650 CMRootRegions* root_regions() { return &_root_regions; }
652 bool concurrent_marking_in_progress() {
653 return _concurrent_marking_in_progress;
654 }
655 void set_concurrent_marking_in_progress() {
656 _concurrent_marking_in_progress = true;
657 }
658 void clear_concurrent_marking_in_progress() {
659 _concurrent_marking_in_progress = false;
660 }
662 void update_accum_task_vtime(int i, double vtime) {
663 _accum_task_vtime[i] += vtime;
664 }
666 double all_task_accum_vtime() {
667 double ret = 0.0;
668 for (uint i = 0; i < _max_worker_id; ++i)
669 ret += _accum_task_vtime[i];
670 return ret;
671 }
673 // Attempts to steal an object from the task queues of other tasks
674 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
675 return _task_queues->steal(worker_id, hash_seed, obj);
676 }
678 ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
679 ~ConcurrentMark();
681 ConcurrentMarkThread* cmThread() { return _cmThread; }
683 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
684 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
686 // Returns the number of GC threads to be used in a concurrent
687 // phase based on the number of GC threads being used in a STW
688 // phase.
689 uint scale_parallel_threads(uint n_par_threads);
691 // Calculates the number of GC threads to be used in a concurrent phase.
692 uint calc_parallel_marking_threads();
694 // The following three are interaction between CM and
695 // G1CollectedHeap
697 // This notifies CM that a root during initial-mark needs to be
698 // grayed. It is MT-safe. word_size is the size of the object in
699 // words. It is passed explicitly as sometimes we cannot calculate
700 // it from the given object because it might be in an inconsistent
701 // state (e.g., in to-space and being copied). So the caller is
702 // responsible for dealing with this issue (e.g., get the size from
703 // the from-space image when the to-space image might be
704 // inconsistent) and always passing the size. hr is the region that
705 // contains the object and it's passed optionally from callers who
706 // might already have it (no point in recalculating it).
707 inline void grayRoot(oop obj, size_t word_size,
708 uint worker_id, HeapRegion* hr = NULL);
710 // It iterates over the heap and for each object it comes across it
711 // will dump the contents of its reference fields, as well as
712 // liveness information for the object and its referents. The dump
713 // will be written to a file with the following name:
714 // G1PrintReachableBaseFile + "." + str.
715 // vo decides whether the prev (vo == UsePrevMarking), the next
716 // (vo == UseNextMarking) marking information, or the mark word
717 // (vo == UseMarkWord) will be used to determine the liveness of
718 // each object / referent.
719 // If all is true, all objects in the heap will be dumped, otherwise
720 // only the live ones. In the dump the following symbols / breviations
721 // are used:
722 // M : an explicitly live object (its bitmap bit is set)
723 // > : an implicitly live object (over tams)
724 // O : an object outside the G1 heap (typically: in the perm gen)
725 // NOT : a reference field whose referent is not live
726 // AND MARKED : indicates that an object is both explicitly and
727 // implicitly live (it should be one or the other, not both)
728 void print_reachable(const char* str,
729 VerifyOption vo, bool all) PRODUCT_RETURN;
731 // Clear the next marking bitmap (will be called concurrently).
732 void clearNextBitmap();
734 // These two do the work that needs to be done before and after the
735 // initial root checkpoint. Since this checkpoint can be done at two
736 // different points (i.e. an explicit pause or piggy-backed on a
737 // young collection), then it's nice to be able to easily share the
738 // pre/post code. It might be the case that we can put everything in
739 // the post method. TP
740 void checkpointRootsInitialPre();
741 void checkpointRootsInitialPost();
743 // Scan all the root regions and mark everything reachable from
744 // them.
745 void scanRootRegions();
747 // Scan a single root region and mark everything reachable from it.
748 void scanRootRegion(HeapRegion* hr, uint worker_id);
750 // Do concurrent phase of marking, to a tentative transitive closure.
751 void markFromRoots();
753 void checkpointRootsFinal(bool clear_all_soft_refs);
754 void checkpointRootsFinalWork();
755 void cleanup();
756 void completeCleanup();
758 // Mark in the previous bitmap. NB: this is usually read-only, so use
759 // this carefully!
760 inline void markPrev(oop p);
762 // Clears marks for all objects in the given range, for the prev,
763 // next, or both bitmaps. NB: the previous bitmap is usually
764 // read-only, so use this carefully!
765 void clearRangePrevBitmap(MemRegion mr);
766 void clearRangeNextBitmap(MemRegion mr);
767 void clearRangeBothBitmaps(MemRegion mr);
769 // Notify data structures that a GC has started.
770 void note_start_of_gc() {
771 _markStack.note_start_of_gc();
772 }
774 // Notify data structures that a GC is finished.
775 void note_end_of_gc() {
776 _markStack.note_end_of_gc();
777 }
779 // Verify that there are no CSet oops on the stacks (taskqueues /
780 // global mark stack), enqueued SATB buffers, per-thread SATB
781 // buffers, and fingers (global / per-task). The boolean parameters
782 // decide which of the above data structures to verify. If marking
783 // is not in progress, it's a no-op.
784 void verify_no_cset_oops(bool verify_stacks,
785 bool verify_enqueued_buffers,
786 bool verify_thread_buffers,
787 bool verify_fingers) PRODUCT_RETURN;
789 // It is called at the end of an evacuation pause during marking so
790 // that CM is notified of where the new end of the heap is. It
791 // doesn't do anything if concurrent_marking_in_progress() is false,
792 // unless the force parameter is true.
793 void update_g1_committed(bool force = false);
795 bool isMarked(oop p) const {
796 assert(p != NULL && p->is_oop(), "expected an oop");
797 HeapWord* addr = (HeapWord*)p;
798 assert(addr >= _nextMarkBitMap->startWord() ||
799 addr < _nextMarkBitMap->endWord(), "in a region");
801 return _nextMarkBitMap->isMarked(addr);
802 }
804 inline bool not_yet_marked(oop p) const;
806 // XXX Debug code
807 bool containing_card_is_marked(void* p);
808 bool containing_cards_are_marked(void* start, void* last);
810 bool isPrevMarked(oop p) const {
811 assert(p != NULL && p->is_oop(), "expected an oop");
812 HeapWord* addr = (HeapWord*)p;
813 assert(addr >= _prevMarkBitMap->startWord() ||
814 addr < _prevMarkBitMap->endWord(), "in a region");
816 return _prevMarkBitMap->isMarked(addr);
817 }
819 inline bool do_yield_check(uint worker_i = 0);
820 inline bool should_yield();
822 // Called to abort the marking cycle after a Full GC takes palce.
823 void abort();
825 // This prints the global/local fingers. It is used for debugging.
826 NOT_PRODUCT(void print_finger();)
828 void print_summary_info();
830 void print_worker_threads_on(outputStream* st) const;
832 // The following indicate whether a given verbose level has been
833 // set. Notice that anything above stats is conditional to
834 // _MARKING_VERBOSE_ having been set to 1
835 bool verbose_stats() {
836 return _verbose_level >= stats_verbose;
837 }
838 bool verbose_low() {
839 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
840 }
841 bool verbose_medium() {
842 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
843 }
844 bool verbose_high() {
845 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
846 }
848 // Liveness counting
850 // Utility routine to set an exclusive range of cards on the given
851 // card liveness bitmap
852 inline void set_card_bitmap_range(BitMap* card_bm,
853 BitMap::idx_t start_idx,
854 BitMap::idx_t end_idx,
855 bool is_par);
857 // Returns the card number of the bottom of the G1 heap.
858 // Used in biasing indices into accounting card bitmaps.
859 intptr_t heap_bottom_card_num() const {
860 return _heap_bottom_card_num;
861 }
863 // Returns the card bitmap for a given task or worker id.
864 BitMap* count_card_bitmap_for(uint worker_id) {
865 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
866 assert(_count_card_bitmaps != NULL, "uninitialized");
867 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
868 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
869 return task_card_bm;
870 }
872 // Returns the array containing the marked bytes for each region,
873 // for the given worker or task id.
874 size_t* count_marked_bytes_array_for(uint worker_id) {
875 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
876 assert(_count_marked_bytes != NULL, "uninitialized");
877 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
878 assert(marked_bytes_array != NULL, "uninitialized");
879 return marked_bytes_array;
880 }
882 // Returns the index in the liveness accounting card table bitmap
883 // for the given address
884 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
886 // Counts the size of the given memory region in the the given
887 // marked_bytes array slot for the given HeapRegion.
888 // Sets the bits in the given card bitmap that are associated with the
889 // cards that are spanned by the memory region.
890 inline void count_region(MemRegion mr, HeapRegion* hr,
891 size_t* marked_bytes_array,
892 BitMap* task_card_bm);
894 // Counts the given memory region in the task/worker counting
895 // data structures for the given worker id.
896 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
898 // Counts the given memory region in the task/worker counting
899 // data structures for the given worker id.
900 inline void count_region(MemRegion mr, uint worker_id);
902 // Counts the given object in the given task/worker counting
903 // data structures.
904 inline void count_object(oop obj, HeapRegion* hr,
905 size_t* marked_bytes_array,
906 BitMap* task_card_bm);
908 // Counts the given object in the task/worker counting data
909 // structures for the given worker id.
910 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
912 // Attempts to mark the given object and, if successful, counts
913 // the object in the given task/worker counting structures.
914 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
915 size_t* marked_bytes_array,
916 BitMap* task_card_bm);
918 // Attempts to mark the given object and, if successful, counts
919 // the object in the task/worker counting structures for the
920 // given worker id.
921 inline bool par_mark_and_count(oop obj, size_t word_size,
922 HeapRegion* hr, uint worker_id);
924 // Attempts to mark the given object and, if successful, counts
925 // the object in the task/worker counting structures for the
926 // given worker id.
927 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
929 // Similar to the above routine but we don't know the heap region that
930 // contains the object to be marked/counted, which this routine looks up.
931 inline bool par_mark_and_count(oop obj, uint worker_id);
933 // Similar to the above routine but there are times when we cannot
934 // safely calculate the size of obj due to races and we, therefore,
935 // pass the size in as a parameter. It is the caller's reponsibility
936 // to ensure that the size passed in for obj is valid.
937 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
939 // Unconditionally mark the given object, and unconditinally count
940 // the object in the counting structures for worker id 0.
941 // Should *not* be called from parallel code.
942 inline bool mark_and_count(oop obj, HeapRegion* hr);
944 // Similar to the above routine but we don't know the heap region that
945 // contains the object to be marked/counted, which this routine looks up.
946 // Should *not* be called from parallel code.
947 inline bool mark_and_count(oop obj);
949 // Returns true if initialization was successfully completed.
950 bool completed_initialization() const {
951 return _completed_initialization;
952 }
954 protected:
955 // Clear all the per-task bitmaps and arrays used to store the
956 // counting data.
957 void clear_all_count_data();
959 // Aggregates the counting data for each worker/task
960 // that was constructed while marking. Also sets
961 // the amount of marked bytes for each region and
962 // the top at concurrent mark count.
963 void aggregate_count_data();
965 // Verification routine
966 void verify_count_data();
967 };
969 // A class representing a marking task.
970 class CMTask : public TerminatorTerminator {
971 private:
972 enum PrivateConstants {
973 // the regular clock call is called once the scanned words reaches
974 // this limit
975 words_scanned_period = 12*1024,
976 // the regular clock call is called once the number of visited
977 // references reaches this limit
978 refs_reached_period = 384,
979 // initial value for the hash seed, used in the work stealing code
980 init_hash_seed = 17,
981 // how many entries will be transferred between global stack and
982 // local queues
983 global_stack_transfer_size = 16
984 };
986 uint _worker_id;
987 G1CollectedHeap* _g1h;
988 ConcurrentMark* _cm;
989 CMBitMap* _nextMarkBitMap;
990 // the task queue of this task
991 CMTaskQueue* _task_queue;
992 private:
993 // the task queue set---needed for stealing
994 CMTaskQueueSet* _task_queues;
995 // indicates whether the task has been claimed---this is only for
996 // debugging purposes
997 bool _claimed;
999 // number of calls to this task
1000 int _calls;
1002 // when the virtual timer reaches this time, the marking step should
1003 // exit
1004 double _time_target_ms;
1005 // the start time of the current marking step
1006 double _start_time_ms;
1008 // the oop closure used for iterations over oops
1009 G1CMOopClosure* _cm_oop_closure;
1011 // the region this task is scanning, NULL if we're not scanning any
1012 HeapRegion* _curr_region;
1013 // the local finger of this task, NULL if we're not scanning a region
1014 HeapWord* _finger;
1015 // limit of the region this task is scanning, NULL if we're not scanning one
1016 HeapWord* _region_limit;
1018 // the number of words this task has scanned
1019 size_t _words_scanned;
1020 // When _words_scanned reaches this limit, the regular clock is
1021 // called. Notice that this might be decreased under certain
1022 // circumstances (i.e. when we believe that we did an expensive
1023 // operation).
1024 size_t _words_scanned_limit;
1025 // the initial value of _words_scanned_limit (i.e. what it was
1026 // before it was decreased).
1027 size_t _real_words_scanned_limit;
1029 // the number of references this task has visited
1030 size_t _refs_reached;
1031 // When _refs_reached reaches this limit, the regular clock is
1032 // called. Notice this this might be decreased under certain
1033 // circumstances (i.e. when we believe that we did an expensive
1034 // operation).
1035 size_t _refs_reached_limit;
1036 // the initial value of _refs_reached_limit (i.e. what it was before
1037 // it was decreased).
1038 size_t _real_refs_reached_limit;
1040 // used by the work stealing stuff
1041 int _hash_seed;
1042 // if this is true, then the task has aborted for some reason
1043 bool _has_aborted;
1044 // set when the task aborts because it has met its time quota
1045 bool _has_timed_out;
1046 // true when we're draining SATB buffers; this avoids the task
1047 // aborting due to SATB buffers being available (as we're already
1048 // dealing with them)
1049 bool _draining_satb_buffers;
1051 // number sequence of past step times
1052 NumberSeq _step_times_ms;
1053 // elapsed time of this task
1054 double _elapsed_time_ms;
1055 // termination time of this task
1056 double _termination_time_ms;
1057 // when this task got into the termination protocol
1058 double _termination_start_time_ms;
1060 // true when the task is during a concurrent phase, false when it is
1061 // in the remark phase (so, in the latter case, we do not have to
1062 // check all the things that we have to check during the concurrent
1063 // phase, i.e. SATB buffer availability...)
1064 bool _concurrent;
1066 TruncatedSeq _marking_step_diffs_ms;
1068 // Counting data structures. Embedding the task's marked_bytes_array
1069 // and card bitmap into the actual task saves having to go through
1070 // the ConcurrentMark object.
1071 size_t* _marked_bytes_array;
1072 BitMap* _card_bm;
1074 // LOTS of statistics related with this task
1075 #if _MARKING_STATS_
1076 NumberSeq _all_clock_intervals_ms;
1077 double _interval_start_time_ms;
1079 int _aborted;
1080 int _aborted_overflow;
1081 int _aborted_cm_aborted;
1082 int _aborted_yield;
1083 int _aborted_timed_out;
1084 int _aborted_satb;
1085 int _aborted_termination;
1087 int _steal_attempts;
1088 int _steals;
1090 int _clock_due_to_marking;
1091 int _clock_due_to_scanning;
1093 int _local_pushes;
1094 int _local_pops;
1095 int _local_max_size;
1096 int _objs_scanned;
1098 int _global_pushes;
1099 int _global_pops;
1100 int _global_max_size;
1102 int _global_transfers_to;
1103 int _global_transfers_from;
1105 int _regions_claimed;
1106 int _objs_found_on_bitmap;
1108 int _satb_buffers_processed;
1109 #endif // _MARKING_STATS_
1111 // it updates the local fields after this task has claimed
1112 // a new region to scan
1113 void setup_for_region(HeapRegion* hr);
1114 // it brings up-to-date the limit of the region
1115 void update_region_limit();
1117 // called when either the words scanned or the refs visited limit
1118 // has been reached
1119 void reached_limit();
1120 // recalculates the words scanned and refs visited limits
1121 void recalculate_limits();
1122 // decreases the words scanned and refs visited limits when we reach
1123 // an expensive operation
1124 void decrease_limits();
1125 // it checks whether the words scanned or refs visited reached their
1126 // respective limit and calls reached_limit() if they have
1127 void check_limits() {
1128 if (_words_scanned >= _words_scanned_limit ||
1129 _refs_reached >= _refs_reached_limit) {
1130 reached_limit();
1131 }
1132 }
1133 // this is supposed to be called regularly during a marking step as
1134 // it checks a bunch of conditions that might cause the marking step
1135 // to abort
1136 void regular_clock_call();
1137 bool concurrent() { return _concurrent; }
1139 public:
1140 // It resets the task; it should be called right at the beginning of
1141 // a marking phase.
1142 void reset(CMBitMap* _nextMarkBitMap);
1143 // it clears all the fields that correspond to a claimed region.
1144 void clear_region_fields();
1146 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1148 // The main method of this class which performs a marking step
1149 // trying not to exceed the given duration. However, it might exit
1150 // prematurely, according to some conditions (i.e. SATB buffers are
1151 // available for processing).
1152 void do_marking_step(double target_ms,
1153 bool do_termination,
1154 bool is_serial);
1156 // These two calls start and stop the timer
1157 void record_start_time() {
1158 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1159 }
1160 void record_end_time() {
1161 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1162 }
1164 // returns the worker ID associated with this task.
1165 uint worker_id() { return _worker_id; }
1167 // From TerminatorTerminator. It determines whether this task should
1168 // exit the termination protocol after it's entered it.
1169 virtual bool should_exit_termination();
1171 // Resets the local region fields after a task has finished scanning a
1172 // region; or when they have become stale as a result of the region
1173 // being evacuated.
1174 void giveup_current_region();
1176 HeapWord* finger() { return _finger; }
1178 bool has_aborted() { return _has_aborted; }
1179 void set_has_aborted() { _has_aborted = true; }
1180 void clear_has_aborted() { _has_aborted = false; }
1181 bool has_timed_out() { return _has_timed_out; }
1182 bool claimed() { return _claimed; }
1184 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1186 // It grays the object by marking it and, if necessary, pushing it
1187 // on the local queue
1188 inline void deal_with_reference(oop obj);
1190 // It scans an object and visits its children.
1191 void scan_object(oop obj);
1193 // It pushes an object on the local queue.
1194 inline void push(oop obj);
1196 // These two move entries to/from the global stack.
1197 void move_entries_to_global_stack();
1198 void get_entries_from_global_stack();
1200 // It pops and scans objects from the local queue. If partially is
1201 // true, then it stops when the queue size is of a given limit. If
1202 // partially is false, then it stops when the queue is empty.
1203 void drain_local_queue(bool partially);
1204 // It moves entries from the global stack to the local queue and
1205 // drains the local queue. If partially is true, then it stops when
1206 // both the global stack and the local queue reach a given size. If
1207 // partially if false, it tries to empty them totally.
1208 void drain_global_stack(bool partially);
1209 // It keeps picking SATB buffers and processing them until no SATB
1210 // buffers are available.
1211 void drain_satb_buffers();
1213 // moves the local finger to a new location
1214 inline void move_finger_to(HeapWord* new_finger) {
1215 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1216 _finger = new_finger;
1217 }
1219 CMTask(uint worker_id, ConcurrentMark *cm,
1220 size_t* marked_bytes, BitMap* card_bm,
1221 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1223 // it prints statistics associated with this task
1224 void print_stats();
1226 #if _MARKING_STATS_
1227 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1228 #endif // _MARKING_STATS_
1229 };
1231 // Class that's used to to print out per-region liveness
1232 // information. It's currently used at the end of marking and also
1233 // after we sort the old regions at the end of the cleanup operation.
1234 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1235 private:
1236 outputStream* _out;
1238 // Accumulators for these values.
1239 size_t _total_used_bytes;
1240 size_t _total_capacity_bytes;
1241 size_t _total_prev_live_bytes;
1242 size_t _total_next_live_bytes;
1244 // These are set up when we come across a "stars humongous" region
1245 // (as this is where most of this information is stored, not in the
1246 // subsequent "continues humongous" regions). After that, for every
1247 // region in a given humongous region series we deduce the right
1248 // values for it by simply subtracting the appropriate amount from
1249 // these fields. All these values should reach 0 after we've visited
1250 // the last region in the series.
1251 size_t _hum_used_bytes;
1252 size_t _hum_capacity_bytes;
1253 size_t _hum_prev_live_bytes;
1254 size_t _hum_next_live_bytes;
1256 static double perc(size_t val, size_t total) {
1257 if (total == 0) {
1258 return 0.0;
1259 } else {
1260 return 100.0 * ((double) val / (double) total);
1261 }
1262 }
1264 static double bytes_to_mb(size_t val) {
1265 return (double) val / (double) M;
1266 }
1268 // See the .cpp file.
1269 size_t get_hum_bytes(size_t* hum_bytes);
1270 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1271 size_t* prev_live_bytes, size_t* next_live_bytes);
1273 public:
1274 // The header and footer are printed in the constructor and
1275 // destructor respectively.
1276 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1277 virtual bool doHeapRegion(HeapRegion* r);
1278 ~G1PrintRegionLivenessInfoClosure();
1279 };
1281 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP