Mon, 01 Oct 2012 09:28:13 -0700
8000244: G1: Ergonomically set MarkStackSize and use virtual space for global marking stack
Summary: Set the value of MarkStackSize to a value based on the number of parallel marking threads with a reasonable minimum. Expand the marking stack if we have to restart marking due to an overflow up to a reasonable maximum. Allocate the underlying space for the marking stack from virtual memory.
Reviewed-by: jmasa, brutisso
1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
28 #include "gc_implementation/g1/heapRegionSets.hpp"
29 #include "utilities/taskqueue.hpp"
31 class G1CollectedHeap;
32 class CMTask;
33 typedef GenericTaskQueue<oop, 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 // XXX Fix these so that offsets are size_t's...
101 HeapWord* offsetToHeapWord(size_t offset) const {
102 return _bmStartWord + (offset << _shifter);
103 }
104 size_t heapWordToOffset(HeapWord* addr) const {
105 return pointer_delta(addr, _bmStartWord) >> _shifter;
106 }
107 int heapWordDiffToOffsetDiff(size_t diff) const;
108 HeapWord* nextWord(HeapWord* addr) {
109 return offsetToHeapWord(heapWordToOffset(addr) + 1);
110 }
112 // debugging
113 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
114 };
116 class CMBitMap : public CMBitMapRO {
118 public:
119 // constructor
120 CMBitMap(int shifter) :
121 CMBitMapRO(shifter) {}
123 // Allocates the back store for the marking bitmap
124 bool allocate(ReservedSpace heap_rs);
126 // write marks
127 void mark(HeapWord* addr) {
128 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
129 "outside underlying space?");
130 _bm.set_bit(heapWordToOffset(addr));
131 }
132 void clear(HeapWord* addr) {
133 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
134 "outside underlying space?");
135 _bm.clear_bit(heapWordToOffset(addr));
136 }
137 bool parMark(HeapWord* addr) {
138 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
139 "outside underlying space?");
140 return _bm.par_set_bit(heapWordToOffset(addr));
141 }
142 bool parClear(HeapWord* addr) {
143 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
144 "outside underlying space?");
145 return _bm.par_clear_bit(heapWordToOffset(addr));
146 }
147 void markRange(MemRegion mr);
148 void clearAll();
149 void clearRange(MemRegion mr);
151 // Starting at the bit corresponding to "addr" (inclusive), find the next
152 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
153 // the end of this run (stopping at "end_addr"). Return the MemRegion
154 // covering from the start of the region corresponding to the first bit
155 // of the run to the end of the region corresponding to the last bit of
156 // the run. If there is no "1" bit at or after "addr", return an empty
157 // MemRegion.
158 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
159 };
161 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
162 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
163 VirtualSpace _virtual_space; // Underlying backing store for actual stack
164 ConcurrentMark* _cm;
165 oop* _base; // bottom of stack
166 jint _index; // one more than last occupied index
167 jint _capacity; // max #elements
168 jint _saved_index; // value of _index saved at start of GC
169 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
171 bool _overflow;
172 bool _should_expand;
173 DEBUG_ONLY(bool _drain_in_progress;)
174 DEBUG_ONLY(bool _drain_in_progress_yields;)
176 public:
177 CMMarkStack(ConcurrentMark* cm);
178 ~CMMarkStack();
180 #ifndef PRODUCT
181 jint max_depth() const {
182 return _max_depth;
183 }
184 #endif
186 bool allocate(size_t capacity);
188 oop pop() {
189 if (!isEmpty()) {
190 return _base[--_index] ;
191 }
192 return NULL;
193 }
195 // If overflow happens, don't do the push, and record the overflow.
196 // *Requires* that "ptr" is already marked.
197 void push(oop ptr) {
198 if (isFull()) {
199 // Record overflow.
200 _overflow = true;
201 return;
202 } else {
203 _base[_index++] = ptr;
204 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
205 }
206 }
207 // Non-block impl. Note: concurrency is allowed only with other
208 // "par_push" operations, not with "pop" or "drain". We would need
209 // parallel versions of them if such concurrency was desired.
210 void par_push(oop ptr);
212 // Pushes the first "n" elements of "ptr_arr" on the stack.
213 // Non-block impl. Note: concurrency is allowed only with other
214 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
215 void par_adjoin_arr(oop* ptr_arr, int n);
217 // Pushes the first "n" elements of "ptr_arr" on the stack.
218 // Locking impl: concurrency is allowed only with
219 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
220 // locking strategy.
221 void par_push_arr(oop* ptr_arr, int n);
223 // If returns false, the array was empty. Otherwise, removes up to "max"
224 // elements from the stack, and transfers them to "ptr_arr" in an
225 // unspecified order. The actual number transferred is given in "n" ("n
226 // == 0" is deliberately redundant with the return value.) Locking impl:
227 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
228 // operations, which use the same locking strategy.
229 bool par_pop_arr(oop* ptr_arr, int max, int* n);
231 // Drain the mark stack, applying the given closure to all fields of
232 // objects on the stack. (That is, continue until the stack is empty,
233 // even if closure applications add entries to the stack.) The "bm"
234 // argument, if non-null, may be used to verify that only marked objects
235 // are on the mark stack. If "yield_after" is "true", then the
236 // concurrent marker performing the drain offers to yield after
237 // processing each object. If a yield occurs, stops the drain operation
238 // and returns false. Otherwise, returns true.
239 template<class OopClosureClass>
240 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
242 bool isEmpty() { return _index == 0; }
243 bool isFull() { return _index == _capacity; }
244 int maxElems() { return _capacity; }
246 bool overflow() { return _overflow; }
247 void clear_overflow() { _overflow = false; }
249 bool should_expand() const { return _should_expand; }
250 void set_should_expand();
252 // Expand the stack, typically in response to an overflow condition
253 void expand();
255 int size() { return _index; }
257 void setEmpty() { _index = 0; clear_overflow(); }
259 // Record the current index.
260 void note_start_of_gc();
262 // Make sure that we have not added any entries to the stack during GC.
263 void note_end_of_gc();
265 // iterate over the oops in the mark stack, up to the bound recorded via
266 // the call above.
267 void oops_do(OopClosure* f);
268 };
270 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
271 private:
272 #ifndef PRODUCT
273 uintx _num_remaining;
274 bool _force;
275 #endif // !defined(PRODUCT)
277 public:
278 void init() PRODUCT_RETURN;
279 void update() PRODUCT_RETURN;
280 bool should_force() PRODUCT_RETURN_( return false; );
281 };
283 // this will enable a variety of different statistics per GC task
284 #define _MARKING_STATS_ 0
285 // this will enable the higher verbose levels
286 #define _MARKING_VERBOSE_ 0
288 #if _MARKING_STATS_
289 #define statsOnly(statement) \
290 do { \
291 statement ; \
292 } while (0)
293 #else // _MARKING_STATS_
294 #define statsOnly(statement) \
295 do { \
296 } while (0)
297 #endif // _MARKING_STATS_
299 typedef enum {
300 no_verbose = 0, // verbose turned off
301 stats_verbose, // only prints stats at the end of marking
302 low_verbose, // low verbose, mostly per region and per major event
303 medium_verbose, // a bit more detailed than low
304 high_verbose // per object verbose
305 } CMVerboseLevel;
307 class YoungList;
309 // Root Regions are regions that are not empty at the beginning of a
310 // marking cycle and which we might collect during an evacuation pause
311 // while the cycle is active. Given that, during evacuation pauses, we
312 // do not copy objects that are explicitly marked, what we have to do
313 // for the root regions is to scan them and mark all objects reachable
314 // from them. According to the SATB assumptions, we only need to visit
315 // each object once during marking. So, as long as we finish this scan
316 // before the next evacuation pause, we can copy the objects from the
317 // root regions without having to mark them or do anything else to them.
318 //
319 // Currently, we only support root region scanning once (at the start
320 // of the marking cycle) and the root regions are all the survivor
321 // regions populated during the initial-mark pause.
322 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
323 private:
324 YoungList* _young_list;
325 ConcurrentMark* _cm;
327 volatile bool _scan_in_progress;
328 volatile bool _should_abort;
329 HeapRegion* volatile _next_survivor;
331 public:
332 CMRootRegions();
333 // We actually do most of the initialization in this method.
334 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
336 // Reset the claiming / scanning of the root regions.
337 void prepare_for_scan();
339 // Forces get_next() to return NULL so that the iteration aborts early.
340 void abort() { _should_abort = true; }
342 // Return true if the CM thread are actively scanning root regions,
343 // false otherwise.
344 bool scan_in_progress() { return _scan_in_progress; }
346 // Claim the next root region to scan atomically, or return NULL if
347 // all have been claimed.
348 HeapRegion* claim_next();
350 // Flag that we're done with root region scanning and notify anyone
351 // who's waiting on it. If aborted is false, assume that all regions
352 // have been claimed.
353 void scan_finished();
355 // If CM threads are still scanning root regions, wait until they
356 // are done. Return true if we had to wait, false otherwise.
357 bool wait_until_scan_finished();
358 };
360 class ConcurrentMarkThread;
362 class ConcurrentMark: public CHeapObj<mtGC> {
363 friend class CMMarkStack;
364 friend class ConcurrentMarkThread;
365 friend class CMTask;
366 friend class CMBitMapClosure;
367 friend class CMGlobalObjectClosure;
368 friend class CMRemarkTask;
369 friend class CMConcurrentMarkingTask;
370 friend class G1ParNoteEndTask;
371 friend class CalcLiveObjectsClosure;
372 friend class G1CMRefProcTaskProxy;
373 friend class G1CMRefProcTaskExecutor;
374 friend class G1CMParKeepAliveAndDrainClosure;
375 friend class G1CMParDrainMarkingStackClosure;
377 protected:
378 ConcurrentMarkThread* _cmThread; // the thread doing the work
379 G1CollectedHeap* _g1h; // the heap.
380 uint _parallel_marking_threads; // the number of marking
381 // threads we're use
382 uint _max_parallel_marking_threads; // max number of marking
383 // threads we'll ever use
384 double _sleep_factor; // how much we have to sleep, with
385 // respect to the work we just did, to
386 // meet the marking overhead goal
387 double _marking_task_overhead; // marking target overhead for
388 // a single task
390 // same as the two above, but for the cleanup task
391 double _cleanup_sleep_factor;
392 double _cleanup_task_overhead;
394 FreeRegionList _cleanup_list;
396 // Concurrent marking support structures
397 CMBitMap _markBitMap1;
398 CMBitMap _markBitMap2;
399 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
400 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
402 BitMap _region_bm;
403 BitMap _card_bm;
405 // Heap bounds
406 HeapWord* _heap_start;
407 HeapWord* _heap_end;
409 // Root region tracking and claiming.
410 CMRootRegions _root_regions;
412 // For gray objects
413 CMMarkStack _markStack; // Grey objects behind global finger.
414 HeapWord* volatile _finger; // the global finger, region aligned,
415 // always points to the end of the
416 // last claimed region
418 // marking tasks
419 uint _max_worker_id;// maximum worker id
420 uint _active_tasks; // task num currently active
421 CMTask** _tasks; // task queue array (max_worker_id len)
422 CMTaskQueueSet* _task_queues; // task queue set
423 ParallelTaskTerminator _terminator; // for termination
425 // Two sync barriers that are used to synchronise tasks when an
426 // overflow occurs. The algorithm is the following. All tasks enter
427 // the first one to ensure that they have all stopped manipulating
428 // the global data structures. After they exit it, they re-initialise
429 // their data structures and task 0 re-initialises the global data
430 // structures. Then, they enter the second sync barrier. This
431 // ensure, that no task starts doing work before all data
432 // structures (local and global) have been re-initialised. When they
433 // exit it, they are free to start working again.
434 WorkGangBarrierSync _first_overflow_barrier_sync;
435 WorkGangBarrierSync _second_overflow_barrier_sync;
437 // this is set by any task, when an overflow on the global data
438 // structures is detected.
439 volatile bool _has_overflown;
440 // true: marking is concurrent, false: we're in remark
441 volatile bool _concurrent;
442 // set at the end of a Full GC so that marking aborts
443 volatile bool _has_aborted;
445 // used when remark aborts due to an overflow to indicate that
446 // another concurrent marking phase should start
447 volatile bool _restart_for_overflow;
449 // This is true from the very start of concurrent marking until the
450 // point when all the tasks complete their work. It is really used
451 // to determine the points between the end of concurrent marking and
452 // time of remark.
453 volatile bool _concurrent_marking_in_progress;
455 // verbose level
456 CMVerboseLevel _verbose_level;
458 // All of these times are in ms.
459 NumberSeq _init_times;
460 NumberSeq _remark_times;
461 NumberSeq _remark_mark_times;
462 NumberSeq _remark_weak_ref_times;
463 NumberSeq _cleanup_times;
464 double _total_counting_time;
465 double _total_rs_scrub_time;
467 double* _accum_task_vtime; // accumulated task vtime
469 FlexibleWorkGang* _parallel_workers;
471 ForceOverflowSettings _force_overflow_conc;
472 ForceOverflowSettings _force_overflow_stw;
474 void weakRefsWork(bool clear_all_soft_refs);
476 void swapMarkBitMaps();
478 // It resets the global marking data structures, as well as the
479 // task local ones; should be called during initial mark.
480 void reset();
481 // It resets all the marking data structures.
482 void clear_marking_state(bool clear_overflow = true);
484 // It should be called to indicate which phase we're in (concurrent
485 // mark or remark) and how many threads are currently active.
486 void set_phase(uint active_tasks, bool concurrent);
487 // We do this after we're done with marking so that the marking data
488 // structures are initialised to a sensible and predictable state.
489 void set_non_marking_state();
491 // prints all gathered CM-related statistics
492 void print_stats();
494 bool cleanup_list_is_empty() {
495 return _cleanup_list.is_empty();
496 }
498 // accessor methods
499 uint parallel_marking_threads() { return _parallel_marking_threads; }
500 uint max_parallel_marking_threads() { return _max_parallel_marking_threads;}
501 double sleep_factor() { return _sleep_factor; }
502 double marking_task_overhead() { return _marking_task_overhead;}
503 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
504 double cleanup_task_overhead() { return _cleanup_task_overhead;}
506 HeapWord* finger() { return _finger; }
507 bool concurrent() { return _concurrent; }
508 uint active_tasks() { return _active_tasks; }
509 ParallelTaskTerminator* terminator() { return &_terminator; }
511 // It claims the next available region to be scanned by a marking
512 // task/thread. It might return NULL if the next region is empty or
513 // we have run out of regions. In the latter case, out_of_regions()
514 // determines whether we've really run out of regions or the task
515 // should call claim_region() again. This might seem a bit
516 // awkward. Originally, the code was written so that claim_region()
517 // either successfully returned with a non-empty region or there
518 // were no more regions to be claimed. The problem with this was
519 // that, in certain circumstances, it iterated over large chunks of
520 // the heap finding only empty regions and, while it was working, it
521 // was preventing the calling task to call its regular clock
522 // method. So, this way, each task will spend very little time in
523 // claim_region() and is allowed to call the regular clock method
524 // frequently.
525 HeapRegion* claim_region(uint worker_id);
527 // It determines whether we've run out of regions to scan.
528 bool out_of_regions() { return _finger == _heap_end; }
530 // Returns the task with the given id
531 CMTask* task(int id) {
532 assert(0 <= id && id < (int) _active_tasks,
533 "task id not within active bounds");
534 return _tasks[id];
535 }
537 // Returns the task queue with the given id
538 CMTaskQueue* task_queue(int id) {
539 assert(0 <= id && id < (int) _active_tasks,
540 "task queue id not within active bounds");
541 return (CMTaskQueue*) _task_queues->queue(id);
542 }
544 // Returns the task queue set
545 CMTaskQueueSet* task_queues() { return _task_queues; }
547 // Access / manipulation of the overflow flag which is set to
548 // indicate that the global stack has overflown
549 bool has_overflown() { return _has_overflown; }
550 void set_has_overflown() { _has_overflown = true; }
551 void clear_has_overflown() { _has_overflown = false; }
552 bool restart_for_overflow() { return _restart_for_overflow; }
554 bool has_aborted() { return _has_aborted; }
556 // Methods to enter the two overflow sync barriers
557 void enter_first_sync_barrier(uint worker_id);
558 void enter_second_sync_barrier(uint worker_id);
560 ForceOverflowSettings* force_overflow_conc() {
561 return &_force_overflow_conc;
562 }
564 ForceOverflowSettings* force_overflow_stw() {
565 return &_force_overflow_stw;
566 }
568 ForceOverflowSettings* force_overflow() {
569 if (concurrent()) {
570 return force_overflow_conc();
571 } else {
572 return force_overflow_stw();
573 }
574 }
576 // Live Data Counting data structures...
577 // These data structures are initialized at the start of
578 // marking. They are written to while marking is active.
579 // They are aggregated during remark; the aggregated values
580 // are then used to populate the _region_bm, _card_bm, and
581 // the total live bytes, which are then subsequently updated
582 // during cleanup.
584 // An array of bitmaps (one bit map per task). Each bitmap
585 // is used to record the cards spanned by the live objects
586 // marked by that task/worker.
587 BitMap* _count_card_bitmaps;
589 // Used to record the number of marked live bytes
590 // (for each region, by worker thread).
591 size_t** _count_marked_bytes;
593 // Card index of the bottom of the G1 heap. Used for biasing indices into
594 // the card bitmaps.
595 intptr_t _heap_bottom_card_num;
597 // Set to true when initialization is complete
598 bool _completed_initialization;
600 public:
601 // Manipulation of the global mark stack.
602 // Notice that the first mark_stack_push is CAS-based, whereas the
603 // two below are Mutex-based. This is OK since the first one is only
604 // called during evacuation pauses and doesn't compete with the
605 // other two (which are called by the marking tasks during
606 // concurrent marking or remark).
607 bool mark_stack_push(oop p) {
608 _markStack.par_push(p);
609 if (_markStack.overflow()) {
610 set_has_overflown();
611 return false;
612 }
613 return true;
614 }
615 bool mark_stack_push(oop* arr, int n) {
616 _markStack.par_push_arr(arr, n);
617 if (_markStack.overflow()) {
618 set_has_overflown();
619 return false;
620 }
621 return true;
622 }
623 void mark_stack_pop(oop* arr, int max, int* n) {
624 _markStack.par_pop_arr(arr, max, n);
625 }
626 size_t mark_stack_size() { return _markStack.size(); }
627 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
628 bool mark_stack_overflow() { return _markStack.overflow(); }
629 bool mark_stack_empty() { return _markStack.isEmpty(); }
631 CMRootRegions* root_regions() { return &_root_regions; }
633 bool concurrent_marking_in_progress() {
634 return _concurrent_marking_in_progress;
635 }
636 void set_concurrent_marking_in_progress() {
637 _concurrent_marking_in_progress = true;
638 }
639 void clear_concurrent_marking_in_progress() {
640 _concurrent_marking_in_progress = false;
641 }
643 void update_accum_task_vtime(int i, double vtime) {
644 _accum_task_vtime[i] += vtime;
645 }
647 double all_task_accum_vtime() {
648 double ret = 0.0;
649 for (uint i = 0; i < _max_worker_id; ++i)
650 ret += _accum_task_vtime[i];
651 return ret;
652 }
654 // Attempts to steal an object from the task queues of other tasks
655 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
656 return _task_queues->steal(worker_id, hash_seed, obj);
657 }
659 ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
660 ~ConcurrentMark();
662 ConcurrentMarkThread* cmThread() { return _cmThread; }
664 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
665 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
667 // Returns the number of GC threads to be used in a concurrent
668 // phase based on the number of GC threads being used in a STW
669 // phase.
670 uint scale_parallel_threads(uint n_par_threads);
672 // Calculates the number of GC threads to be used in a concurrent phase.
673 uint calc_parallel_marking_threads();
675 // The following three are interaction between CM and
676 // G1CollectedHeap
678 // This notifies CM that a root during initial-mark needs to be
679 // grayed. It is MT-safe. word_size is the size of the object in
680 // words. It is passed explicitly as sometimes we cannot calculate
681 // it from the given object because it might be in an inconsistent
682 // state (e.g., in to-space and being copied). So the caller is
683 // responsible for dealing with this issue (e.g., get the size from
684 // the from-space image when the to-space image might be
685 // inconsistent) and always passing the size. hr is the region that
686 // contains the object and it's passed optionally from callers who
687 // might already have it (no point in recalculating it).
688 inline void grayRoot(oop obj, size_t word_size,
689 uint worker_id, HeapRegion* hr = NULL);
691 // It iterates over the heap and for each object it comes across it
692 // will dump the contents of its reference fields, as well as
693 // liveness information for the object and its referents. The dump
694 // will be written to a file with the following name:
695 // G1PrintReachableBaseFile + "." + str.
696 // vo decides whether the prev (vo == UsePrevMarking), the next
697 // (vo == UseNextMarking) marking information, or the mark word
698 // (vo == UseMarkWord) will be used to determine the liveness of
699 // each object / referent.
700 // If all is true, all objects in the heap will be dumped, otherwise
701 // only the live ones. In the dump the following symbols / breviations
702 // are used:
703 // M : an explicitly live object (its bitmap bit is set)
704 // > : an implicitly live object (over tams)
705 // O : an object outside the G1 heap (typically: in the perm gen)
706 // NOT : a reference field whose referent is not live
707 // AND MARKED : indicates that an object is both explicitly and
708 // implicitly live (it should be one or the other, not both)
709 void print_reachable(const char* str,
710 VerifyOption vo, bool all) PRODUCT_RETURN;
712 // Clear the next marking bitmap (will be called concurrently).
713 void clearNextBitmap();
715 // These two do the work that needs to be done before and after the
716 // initial root checkpoint. Since this checkpoint can be done at two
717 // different points (i.e. an explicit pause or piggy-backed on a
718 // young collection), then it's nice to be able to easily share the
719 // pre/post code. It might be the case that we can put everything in
720 // the post method. TP
721 void checkpointRootsInitialPre();
722 void checkpointRootsInitialPost();
724 // Scan all the root regions and mark everything reachable from
725 // them.
726 void scanRootRegions();
728 // Scan a single root region and mark everything reachable from it.
729 void scanRootRegion(HeapRegion* hr, uint worker_id);
731 // Do concurrent phase of marking, to a tentative transitive closure.
732 void markFromRoots();
734 void checkpointRootsFinal(bool clear_all_soft_refs);
735 void checkpointRootsFinalWork();
736 void cleanup();
737 void completeCleanup();
739 // Mark in the previous bitmap. NB: this is usually read-only, so use
740 // this carefully!
741 inline void markPrev(oop p);
743 // Clears marks for all objects in the given range, for the prev,
744 // next, or both bitmaps. NB: the previous bitmap is usually
745 // read-only, so use this carefully!
746 void clearRangePrevBitmap(MemRegion mr);
747 void clearRangeNextBitmap(MemRegion mr);
748 void clearRangeBothBitmaps(MemRegion mr);
750 // Notify data structures that a GC has started.
751 void note_start_of_gc() {
752 _markStack.note_start_of_gc();
753 }
755 // Notify data structures that a GC is finished.
756 void note_end_of_gc() {
757 _markStack.note_end_of_gc();
758 }
760 // Verify that there are no CSet oops on the stacks (taskqueues /
761 // global mark stack), enqueued SATB buffers, per-thread SATB
762 // buffers, and fingers (global / per-task). The boolean parameters
763 // decide which of the above data structures to verify. If marking
764 // is not in progress, it's a no-op.
765 void verify_no_cset_oops(bool verify_stacks,
766 bool verify_enqueued_buffers,
767 bool verify_thread_buffers,
768 bool verify_fingers) PRODUCT_RETURN;
770 // It is called at the end of an evacuation pause during marking so
771 // that CM is notified of where the new end of the heap is. It
772 // doesn't do anything if concurrent_marking_in_progress() is false,
773 // unless the force parameter is true.
774 void update_g1_committed(bool force = false);
776 bool isMarked(oop p) const {
777 assert(p != NULL && p->is_oop(), "expected an oop");
778 HeapWord* addr = (HeapWord*)p;
779 assert(addr >= _nextMarkBitMap->startWord() ||
780 addr < _nextMarkBitMap->endWord(), "in a region");
782 return _nextMarkBitMap->isMarked(addr);
783 }
785 inline bool not_yet_marked(oop p) const;
787 // XXX Debug code
788 bool containing_card_is_marked(void* p);
789 bool containing_cards_are_marked(void* start, void* last);
791 bool isPrevMarked(oop p) const {
792 assert(p != NULL && p->is_oop(), "expected an oop");
793 HeapWord* addr = (HeapWord*)p;
794 assert(addr >= _prevMarkBitMap->startWord() ||
795 addr < _prevMarkBitMap->endWord(), "in a region");
797 return _prevMarkBitMap->isMarked(addr);
798 }
800 inline bool do_yield_check(uint worker_i = 0);
801 inline bool should_yield();
803 // Called to abort the marking cycle after a Full GC takes palce.
804 void abort();
806 // This prints the global/local fingers. It is used for debugging.
807 NOT_PRODUCT(void print_finger();)
809 void print_summary_info();
811 void print_worker_threads_on(outputStream* st) const;
813 // The following indicate whether a given verbose level has been
814 // set. Notice that anything above stats is conditional to
815 // _MARKING_VERBOSE_ having been set to 1
816 bool verbose_stats() {
817 return _verbose_level >= stats_verbose;
818 }
819 bool verbose_low() {
820 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
821 }
822 bool verbose_medium() {
823 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
824 }
825 bool verbose_high() {
826 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
827 }
829 // Liveness counting
831 // Utility routine to set an exclusive range of cards on the given
832 // card liveness bitmap
833 inline void set_card_bitmap_range(BitMap* card_bm,
834 BitMap::idx_t start_idx,
835 BitMap::idx_t end_idx,
836 bool is_par);
838 // Returns the card number of the bottom of the G1 heap.
839 // Used in biasing indices into accounting card bitmaps.
840 intptr_t heap_bottom_card_num() const {
841 return _heap_bottom_card_num;
842 }
844 // Returns the card bitmap for a given task or worker id.
845 BitMap* count_card_bitmap_for(uint worker_id) {
846 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
847 assert(_count_card_bitmaps != NULL, "uninitialized");
848 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
849 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
850 return task_card_bm;
851 }
853 // Returns the array containing the marked bytes for each region,
854 // for the given worker or task id.
855 size_t* count_marked_bytes_array_for(uint worker_id) {
856 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
857 assert(_count_marked_bytes != NULL, "uninitialized");
858 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
859 assert(marked_bytes_array != NULL, "uninitialized");
860 return marked_bytes_array;
861 }
863 // Returns the index in the liveness accounting card table bitmap
864 // for the given address
865 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
867 // Counts the size of the given memory region in the the given
868 // marked_bytes array slot for the given HeapRegion.
869 // Sets the bits in the given card bitmap that are associated with the
870 // cards that are spanned by the memory region.
871 inline void count_region(MemRegion mr, HeapRegion* hr,
872 size_t* marked_bytes_array,
873 BitMap* task_card_bm);
875 // Counts the given memory region in the task/worker counting
876 // data structures for the given worker id.
877 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
879 // Counts the given memory region in the task/worker counting
880 // data structures for the given worker id.
881 inline void count_region(MemRegion mr, uint worker_id);
883 // Counts the given object in the given task/worker counting
884 // data structures.
885 inline void count_object(oop obj, HeapRegion* hr,
886 size_t* marked_bytes_array,
887 BitMap* task_card_bm);
889 // Counts the given object in the task/worker counting data
890 // structures for the given worker id.
891 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
893 // Attempts to mark the given object and, if successful, counts
894 // the object in the given task/worker counting structures.
895 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
896 size_t* marked_bytes_array,
897 BitMap* task_card_bm);
899 // Attempts to mark the given object and, if successful, counts
900 // the object in the task/worker counting structures for the
901 // given worker id.
902 inline bool par_mark_and_count(oop obj, size_t word_size,
903 HeapRegion* hr, uint worker_id);
905 // Attempts to mark the given object and, if successful, counts
906 // the object in the task/worker counting structures for the
907 // given worker id.
908 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
910 // Similar to the above routine but we don't know the heap region that
911 // contains the object to be marked/counted, which this routine looks up.
912 inline bool par_mark_and_count(oop obj, uint worker_id);
914 // Similar to the above routine but there are times when we cannot
915 // safely calculate the size of obj due to races and we, therefore,
916 // pass the size in as a parameter. It is the caller's reponsibility
917 // to ensure that the size passed in for obj is valid.
918 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
920 // Unconditionally mark the given object, and unconditinally count
921 // the object in the counting structures for worker id 0.
922 // Should *not* be called from parallel code.
923 inline bool mark_and_count(oop obj, HeapRegion* hr);
925 // Similar to the above routine but we don't know the heap region that
926 // contains the object to be marked/counted, which this routine looks up.
927 // Should *not* be called from parallel code.
928 inline bool mark_and_count(oop obj);
930 // Returns true if initialization was successfully completed.
931 bool completed_initialization() const {
932 return _completed_initialization;
933 }
935 protected:
936 // Clear all the per-task bitmaps and arrays used to store the
937 // counting data.
938 void clear_all_count_data();
940 // Aggregates the counting data for each worker/task
941 // that was constructed while marking. Also sets
942 // the amount of marked bytes for each region and
943 // the top at concurrent mark count.
944 void aggregate_count_data();
946 // Verification routine
947 void verify_count_data();
948 };
950 // A class representing a marking task.
951 class CMTask : public TerminatorTerminator {
952 private:
953 enum PrivateConstants {
954 // the regular clock call is called once the scanned words reaches
955 // this limit
956 words_scanned_period = 12*1024,
957 // the regular clock call is called once the number of visited
958 // references reaches this limit
959 refs_reached_period = 384,
960 // initial value for the hash seed, used in the work stealing code
961 init_hash_seed = 17,
962 // how many entries will be transferred between global stack and
963 // local queues
964 global_stack_transfer_size = 16
965 };
967 uint _worker_id;
968 G1CollectedHeap* _g1h;
969 ConcurrentMark* _cm;
970 CMBitMap* _nextMarkBitMap;
971 // the task queue of this task
972 CMTaskQueue* _task_queue;
973 private:
974 // the task queue set---needed for stealing
975 CMTaskQueueSet* _task_queues;
976 // indicates whether the task has been claimed---this is only for
977 // debugging purposes
978 bool _claimed;
980 // number of calls to this task
981 int _calls;
983 // when the virtual timer reaches this time, the marking step should
984 // exit
985 double _time_target_ms;
986 // the start time of the current marking step
987 double _start_time_ms;
989 // the oop closure used for iterations over oops
990 G1CMOopClosure* _cm_oop_closure;
992 // the region this task is scanning, NULL if we're not scanning any
993 HeapRegion* _curr_region;
994 // the local finger of this task, NULL if we're not scanning a region
995 HeapWord* _finger;
996 // limit of the region this task is scanning, NULL if we're not scanning one
997 HeapWord* _region_limit;
999 // the number of words this task has scanned
1000 size_t _words_scanned;
1001 // When _words_scanned reaches this limit, the regular clock is
1002 // called. Notice that this might be decreased under certain
1003 // circumstances (i.e. when we believe that we did an expensive
1004 // operation).
1005 size_t _words_scanned_limit;
1006 // the initial value of _words_scanned_limit (i.e. what it was
1007 // before it was decreased).
1008 size_t _real_words_scanned_limit;
1010 // the number of references this task has visited
1011 size_t _refs_reached;
1012 // When _refs_reached reaches this limit, the regular clock is
1013 // called. Notice this this might be decreased under certain
1014 // circumstances (i.e. when we believe that we did an expensive
1015 // operation).
1016 size_t _refs_reached_limit;
1017 // the initial value of _refs_reached_limit (i.e. what it was before
1018 // it was decreased).
1019 size_t _real_refs_reached_limit;
1021 // used by the work stealing stuff
1022 int _hash_seed;
1023 // if this is true, then the task has aborted for some reason
1024 bool _has_aborted;
1025 // set when the task aborts because it has met its time quota
1026 bool _has_timed_out;
1027 // true when we're draining SATB buffers; this avoids the task
1028 // aborting due to SATB buffers being available (as we're already
1029 // dealing with them)
1030 bool _draining_satb_buffers;
1032 // number sequence of past step times
1033 NumberSeq _step_times_ms;
1034 // elapsed time of this task
1035 double _elapsed_time_ms;
1036 // termination time of this task
1037 double _termination_time_ms;
1038 // when this task got into the termination protocol
1039 double _termination_start_time_ms;
1041 // true when the task is during a concurrent phase, false when it is
1042 // in the remark phase (so, in the latter case, we do not have to
1043 // check all the things that we have to check during the concurrent
1044 // phase, i.e. SATB buffer availability...)
1045 bool _concurrent;
1047 TruncatedSeq _marking_step_diffs_ms;
1049 // Counting data structures. Embedding the task's marked_bytes_array
1050 // and card bitmap into the actual task saves having to go through
1051 // the ConcurrentMark object.
1052 size_t* _marked_bytes_array;
1053 BitMap* _card_bm;
1055 // LOTS of statistics related with this task
1056 #if _MARKING_STATS_
1057 NumberSeq _all_clock_intervals_ms;
1058 double _interval_start_time_ms;
1060 int _aborted;
1061 int _aborted_overflow;
1062 int _aborted_cm_aborted;
1063 int _aborted_yield;
1064 int _aborted_timed_out;
1065 int _aborted_satb;
1066 int _aborted_termination;
1068 int _steal_attempts;
1069 int _steals;
1071 int _clock_due_to_marking;
1072 int _clock_due_to_scanning;
1074 int _local_pushes;
1075 int _local_pops;
1076 int _local_max_size;
1077 int _objs_scanned;
1079 int _global_pushes;
1080 int _global_pops;
1081 int _global_max_size;
1083 int _global_transfers_to;
1084 int _global_transfers_from;
1086 int _regions_claimed;
1087 int _objs_found_on_bitmap;
1089 int _satb_buffers_processed;
1090 #endif // _MARKING_STATS_
1092 // it updates the local fields after this task has claimed
1093 // a new region to scan
1094 void setup_for_region(HeapRegion* hr);
1095 // it brings up-to-date the limit of the region
1096 void update_region_limit();
1098 // called when either the words scanned or the refs visited limit
1099 // has been reached
1100 void reached_limit();
1101 // recalculates the words scanned and refs visited limits
1102 void recalculate_limits();
1103 // decreases the words scanned and refs visited limits when we reach
1104 // an expensive operation
1105 void decrease_limits();
1106 // it checks whether the words scanned or refs visited reached their
1107 // respective limit and calls reached_limit() if they have
1108 void check_limits() {
1109 if (_words_scanned >= _words_scanned_limit ||
1110 _refs_reached >= _refs_reached_limit) {
1111 reached_limit();
1112 }
1113 }
1114 // this is supposed to be called regularly during a marking step as
1115 // it checks a bunch of conditions that might cause the marking step
1116 // to abort
1117 void regular_clock_call();
1118 bool concurrent() { return _concurrent; }
1120 public:
1121 // It resets the task; it should be called right at the beginning of
1122 // a marking phase.
1123 void reset(CMBitMap* _nextMarkBitMap);
1124 // it clears all the fields that correspond to a claimed region.
1125 void clear_region_fields();
1127 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1129 // The main method of this class which performs a marking step
1130 // trying not to exceed the given duration. However, it might exit
1131 // prematurely, according to some conditions (i.e. SATB buffers are
1132 // available for processing).
1133 void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1135 // These two calls start and stop the timer
1136 void record_start_time() {
1137 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1138 }
1139 void record_end_time() {
1140 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1141 }
1143 // returns the worker ID associated with this task.
1144 uint worker_id() { return _worker_id; }
1146 // From TerminatorTerminator. It determines whether this task should
1147 // exit the termination protocol after it's entered it.
1148 virtual bool should_exit_termination();
1150 // Resets the local region fields after a task has finished scanning a
1151 // region; or when they have become stale as a result of the region
1152 // being evacuated.
1153 void giveup_current_region();
1155 HeapWord* finger() { return _finger; }
1157 bool has_aborted() { return _has_aborted; }
1158 void set_has_aborted() { _has_aborted = true; }
1159 void clear_has_aborted() { _has_aborted = false; }
1160 bool has_timed_out() { return _has_timed_out; }
1161 bool claimed() { return _claimed; }
1163 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1165 // It grays the object by marking it and, if necessary, pushing it
1166 // on the local queue
1167 inline void deal_with_reference(oop obj);
1169 // It scans an object and visits its children.
1170 void scan_object(oop obj);
1172 // It pushes an object on the local queue.
1173 inline void push(oop obj);
1175 // These two move entries to/from the global stack.
1176 void move_entries_to_global_stack();
1177 void get_entries_from_global_stack();
1179 // It pops and scans objects from the local queue. If partially is
1180 // true, then it stops when the queue size is of a given limit. If
1181 // partially is false, then it stops when the queue is empty.
1182 void drain_local_queue(bool partially);
1183 // It moves entries from the global stack to the local queue and
1184 // drains the local queue. If partially is true, then it stops when
1185 // both the global stack and the local queue reach a given size. If
1186 // partially if false, it tries to empty them totally.
1187 void drain_global_stack(bool partially);
1188 // It keeps picking SATB buffers and processing them until no SATB
1189 // buffers are available.
1190 void drain_satb_buffers();
1192 // moves the local finger to a new location
1193 inline void move_finger_to(HeapWord* new_finger) {
1194 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1195 _finger = new_finger;
1196 }
1198 CMTask(uint worker_id, ConcurrentMark *cm,
1199 size_t* marked_bytes, BitMap* card_bm,
1200 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1202 // it prints statistics associated with this task
1203 void print_stats();
1205 #if _MARKING_STATS_
1206 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1207 #endif // _MARKING_STATS_
1208 };
1210 // Class that's used to to print out per-region liveness
1211 // information. It's currently used at the end of marking and also
1212 // after we sort the old regions at the end of the cleanup operation.
1213 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1214 private:
1215 outputStream* _out;
1217 // Accumulators for these values.
1218 size_t _total_used_bytes;
1219 size_t _total_capacity_bytes;
1220 size_t _total_prev_live_bytes;
1221 size_t _total_next_live_bytes;
1223 // These are set up when we come across a "stars humongous" region
1224 // (as this is where most of this information is stored, not in the
1225 // subsequent "continues humongous" regions). After that, for every
1226 // region in a given humongous region series we deduce the right
1227 // values for it by simply subtracting the appropriate amount from
1228 // these fields. All these values should reach 0 after we've visited
1229 // the last region in the series.
1230 size_t _hum_used_bytes;
1231 size_t _hum_capacity_bytes;
1232 size_t _hum_prev_live_bytes;
1233 size_t _hum_next_live_bytes;
1235 static double perc(size_t val, size_t total) {
1236 if (total == 0) {
1237 return 0.0;
1238 } else {
1239 return 100.0 * ((double) val / (double) total);
1240 }
1241 }
1243 static double bytes_to_mb(size_t val) {
1244 return (double) val / (double) M;
1245 }
1247 // See the .cpp file.
1248 size_t get_hum_bytes(size_t* hum_bytes);
1249 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1250 size_t* prev_live_bytes, size_t* next_live_bytes);
1252 public:
1253 // The header and footer are printed in the constructor and
1254 // destructor respectively.
1255 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1256 virtual bool doHeapRegion(HeapRegion* r);
1257 ~G1PrintRegionLivenessInfoClosure();
1258 };
1260 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP