Thu, 15 Aug 2013 10:52:18 +0200
7145569: G1: optimize nmethods scanning
Summary: Add a list of nmethods to the RSet for a region that contain references into the region. Skip scanning the code cache during root scanning and scan the nmethod lists during RSet scanning instead.
Reviewed-by: tschatzl, brutisso, mgerdin, twisti, kvn
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 bool do_object_b(oop obj);
48 };
50 // A generic CM bit map. This is essentially a wrapper around the BitMap
51 // class, with one bit per (1<<_shifter) HeapWords.
53 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
54 protected:
55 HeapWord* _bmStartWord; // base address of range covered by map
56 size_t _bmWordSize; // map size (in #HeapWords covered)
57 const int _shifter; // map to char or bit
58 VirtualSpace _virtual_space; // underlying the bit map
59 BitMap _bm; // the bit map itself
61 public:
62 // constructor
63 CMBitMapRO(int shifter);
65 enum { do_yield = true };
67 // inquiries
68 HeapWord* startWord() const { return _bmStartWord; }
69 size_t sizeInWords() const { return _bmWordSize; }
70 // the following is one past the last word in space
71 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
73 // read marks
75 bool isMarked(HeapWord* addr) const {
76 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
77 "outside underlying space?");
78 return _bm.at(heapWordToOffset(addr));
79 }
81 // iteration
82 inline bool iterate(BitMapClosure* cl, MemRegion mr);
83 inline bool iterate(BitMapClosure* cl);
85 // Return the address corresponding to the next marked bit at or after
86 // "addr", and before "limit", if "limit" is non-NULL. If there is no
87 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
88 HeapWord* getNextMarkedWordAddress(HeapWord* addr,
89 HeapWord* limit = NULL) const;
90 // Return the address corresponding to the next unmarked bit at or after
91 // "addr", and before "limit", if "limit" is non-NULL. If there is no
92 // such bit, returns "limit" if that is non-NULL, or else "endWord()".
93 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
94 HeapWord* limit = NULL) const;
96 // conversion utilities
97 HeapWord* offsetToHeapWord(size_t offset) const {
98 return _bmStartWord + (offset << _shifter);
99 }
100 size_t heapWordToOffset(HeapWord* addr) const {
101 return pointer_delta(addr, _bmStartWord) >> _shifter;
102 }
103 int heapWordDiffToOffsetDiff(size_t diff) const;
105 // The argument addr should be the start address of a valid object
106 HeapWord* nextObject(HeapWord* addr) {
107 oop obj = (oop) addr;
108 HeapWord* res = addr + obj->size();
109 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
110 return res;
111 }
113 void print_on_error(outputStream* st, const char* prefix) const;
115 // debugging
116 NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
117 };
119 class CMBitMap : public CMBitMapRO {
121 public:
122 // constructor
123 CMBitMap(int shifter) :
124 CMBitMapRO(shifter) {}
126 // Allocates the back store for the marking bitmap
127 bool allocate(ReservedSpace heap_rs);
129 // write marks
130 void mark(HeapWord* addr) {
131 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
132 "outside underlying space?");
133 _bm.set_bit(heapWordToOffset(addr));
134 }
135 void clear(HeapWord* addr) {
136 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
137 "outside underlying space?");
138 _bm.clear_bit(heapWordToOffset(addr));
139 }
140 bool parMark(HeapWord* addr) {
141 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
142 "outside underlying space?");
143 return _bm.par_set_bit(heapWordToOffset(addr));
144 }
145 bool parClear(HeapWord* addr) {
146 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
147 "outside underlying space?");
148 return _bm.par_clear_bit(heapWordToOffset(addr));
149 }
150 void markRange(MemRegion mr);
151 void clearAll();
152 void clearRange(MemRegion mr);
154 // Starting at the bit corresponding to "addr" (inclusive), find the next
155 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find
156 // the end of this run (stopping at "end_addr"). Return the MemRegion
157 // covering from the start of the region corresponding to the first bit
158 // of the run to the end of the region corresponding to the last bit of
159 // the run. If there is no "1" bit at or after "addr", return an empty
160 // MemRegion.
161 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
162 };
164 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
165 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
166 VirtualSpace _virtual_space; // Underlying backing store for actual stack
167 ConcurrentMark* _cm;
168 oop* _base; // bottom of stack
169 jint _index; // one more than last occupied index
170 jint _capacity; // max #elements
171 jint _saved_index; // value of _index saved at start of GC
172 NOT_PRODUCT(jint _max_depth;) // max depth plumbed during run
174 bool _overflow;
175 bool _should_expand;
176 DEBUG_ONLY(bool _drain_in_progress;)
177 DEBUG_ONLY(bool _drain_in_progress_yields;)
179 public:
180 CMMarkStack(ConcurrentMark* cm);
181 ~CMMarkStack();
183 #ifndef PRODUCT
184 jint max_depth() const {
185 return _max_depth;
186 }
187 #endif
189 bool allocate(size_t capacity);
191 oop pop() {
192 if (!isEmpty()) {
193 return _base[--_index] ;
194 }
195 return NULL;
196 }
198 // If overflow happens, don't do the push, and record the overflow.
199 // *Requires* that "ptr" is already marked.
200 void push(oop ptr) {
201 if (isFull()) {
202 // Record overflow.
203 _overflow = true;
204 return;
205 } else {
206 _base[_index++] = ptr;
207 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
208 }
209 }
210 // Non-block impl. Note: concurrency is allowed only with other
211 // "par_push" operations, not with "pop" or "drain". We would need
212 // parallel versions of them if such concurrency was desired.
213 void par_push(oop ptr);
215 // Pushes the first "n" elements of "ptr_arr" on the stack.
216 // Non-block impl. Note: concurrency is allowed only with other
217 // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
218 void par_adjoin_arr(oop* ptr_arr, int n);
220 // Pushes the first "n" elements of "ptr_arr" on the stack.
221 // Locking impl: concurrency is allowed only with
222 // "par_push_arr" and/or "par_pop_arr" operations, which use the same
223 // locking strategy.
224 void par_push_arr(oop* ptr_arr, int n);
226 // If returns false, the array was empty. Otherwise, removes up to "max"
227 // elements from the stack, and transfers them to "ptr_arr" in an
228 // unspecified order. The actual number transferred is given in "n" ("n
229 // == 0" is deliberately redundant with the return value.) Locking impl:
230 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
231 // operations, which use the same locking strategy.
232 bool par_pop_arr(oop* ptr_arr, int max, int* n);
234 // Drain the mark stack, applying the given closure to all fields of
235 // objects on the stack. (That is, continue until the stack is empty,
236 // even if closure applications add entries to the stack.) The "bm"
237 // argument, if non-null, may be used to verify that only marked objects
238 // are on the mark stack. If "yield_after" is "true", then the
239 // concurrent marker performing the drain offers to yield after
240 // processing each object. If a yield occurs, stops the drain operation
241 // and returns false. Otherwise, returns true.
242 template<class OopClosureClass>
243 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
245 bool isEmpty() { return _index == 0; }
246 bool isFull() { return _index == _capacity; }
247 int maxElems() { return _capacity; }
249 bool overflow() { return _overflow; }
250 void clear_overflow() { _overflow = false; }
252 bool should_expand() const { return _should_expand; }
253 void set_should_expand();
255 // Expand the stack, typically in response to an overflow condition
256 void expand();
258 int size() { return _index; }
260 void setEmpty() { _index = 0; clear_overflow(); }
262 // Record the current index.
263 void note_start_of_gc();
265 // Make sure that we have not added any entries to the stack during GC.
266 void note_end_of_gc();
268 // iterate over the oops in the mark stack, up to the bound recorded via
269 // the call above.
270 void oops_do(OopClosure* f);
271 };
273 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
274 private:
275 #ifndef PRODUCT
276 uintx _num_remaining;
277 bool _force;
278 #endif // !defined(PRODUCT)
280 public:
281 void init() PRODUCT_RETURN;
282 void update() PRODUCT_RETURN;
283 bool should_force() PRODUCT_RETURN_( return false; );
284 };
286 // this will enable a variety of different statistics per GC task
287 #define _MARKING_STATS_ 0
288 // this will enable the higher verbose levels
289 #define _MARKING_VERBOSE_ 0
291 #if _MARKING_STATS_
292 #define statsOnly(statement) \
293 do { \
294 statement ; \
295 } while (0)
296 #else // _MARKING_STATS_
297 #define statsOnly(statement) \
298 do { \
299 } while (0)
300 #endif // _MARKING_STATS_
302 typedef enum {
303 no_verbose = 0, // verbose turned off
304 stats_verbose, // only prints stats at the end of marking
305 low_verbose, // low verbose, mostly per region and per major event
306 medium_verbose, // a bit more detailed than low
307 high_verbose // per object verbose
308 } CMVerboseLevel;
310 class YoungList;
312 // Root Regions are regions that are not empty at the beginning of a
313 // marking cycle and which we might collect during an evacuation pause
314 // while the cycle is active. Given that, during evacuation pauses, we
315 // do not copy objects that are explicitly marked, what we have to do
316 // for the root regions is to scan them and mark all objects reachable
317 // from them. According to the SATB assumptions, we only need to visit
318 // each object once during marking. So, as long as we finish this scan
319 // before the next evacuation pause, we can copy the objects from the
320 // root regions without having to mark them or do anything else to them.
321 //
322 // Currently, we only support root region scanning once (at the start
323 // of the marking cycle) and the root regions are all the survivor
324 // regions populated during the initial-mark pause.
325 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
326 private:
327 YoungList* _young_list;
328 ConcurrentMark* _cm;
330 volatile bool _scan_in_progress;
331 volatile bool _should_abort;
332 HeapRegion* volatile _next_survivor;
334 public:
335 CMRootRegions();
336 // We actually do most of the initialization in this method.
337 void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
339 // Reset the claiming / scanning of the root regions.
340 void prepare_for_scan();
342 // Forces get_next() to return NULL so that the iteration aborts early.
343 void abort() { _should_abort = true; }
345 // Return true if the CM thread are actively scanning root regions,
346 // false otherwise.
347 bool scan_in_progress() { return _scan_in_progress; }
349 // Claim the next root region to scan atomically, or return NULL if
350 // all have been claimed.
351 HeapRegion* claim_next();
353 // Flag that we're done with root region scanning and notify anyone
354 // who's waiting on it. If aborted is false, assume that all regions
355 // have been claimed.
356 void scan_finished();
358 // If CM threads are still scanning root regions, wait until they
359 // are done. Return true if we had to wait, false otherwise.
360 bool wait_until_scan_finished();
361 };
363 class ConcurrentMarkThread;
365 class ConcurrentMark: public CHeapObj<mtGC> {
366 friend class CMMarkStack;
367 friend class ConcurrentMarkThread;
368 friend class CMTask;
369 friend class CMBitMapClosure;
370 friend class CMGlobalObjectClosure;
371 friend class CMRemarkTask;
372 friend class CMConcurrentMarkingTask;
373 friend class G1ParNoteEndTask;
374 friend class CalcLiveObjectsClosure;
375 friend class G1CMRefProcTaskProxy;
376 friend class G1CMRefProcTaskExecutor;
377 friend class G1CMKeepAliveAndDrainClosure;
378 friend class G1CMDrainMarkingStackClosure;
380 protected:
381 ConcurrentMarkThread* _cmThread; // the thread doing the work
382 G1CollectedHeap* _g1h; // the heap.
383 uint _parallel_marking_threads; // the number of marking
384 // threads we're use
385 uint _max_parallel_marking_threads; // max number of marking
386 // threads we'll ever use
387 double _sleep_factor; // how much we have to sleep, with
388 // respect to the work we just did, to
389 // meet the marking overhead goal
390 double _marking_task_overhead; // marking target overhead for
391 // a single task
393 // same as the two above, but for the cleanup task
394 double _cleanup_sleep_factor;
395 double _cleanup_task_overhead;
397 FreeRegionList _cleanup_list;
399 // Concurrent marking support structures
400 CMBitMap _markBitMap1;
401 CMBitMap _markBitMap2;
402 CMBitMapRO* _prevMarkBitMap; // completed mark bitmap
403 CMBitMap* _nextMarkBitMap; // under-construction mark bitmap
405 BitMap _region_bm;
406 BitMap _card_bm;
408 // Heap bounds
409 HeapWord* _heap_start;
410 HeapWord* _heap_end;
412 // Root region tracking and claiming.
413 CMRootRegions _root_regions;
415 // For gray objects
416 CMMarkStack _markStack; // Grey objects behind global finger.
417 HeapWord* volatile _finger; // the global finger, region aligned,
418 // always points to the end of the
419 // last claimed region
421 // marking tasks
422 uint _max_worker_id;// maximum worker id
423 uint _active_tasks; // task num currently active
424 CMTask** _tasks; // task queue array (max_worker_id len)
425 CMTaskQueueSet* _task_queues; // task queue set
426 ParallelTaskTerminator _terminator; // for termination
428 // Two sync barriers that are used to synchronise tasks when an
429 // overflow occurs. The algorithm is the following. All tasks enter
430 // the first one to ensure that they have all stopped manipulating
431 // the global data structures. After they exit it, they re-initialise
432 // their data structures and task 0 re-initialises the global data
433 // structures. Then, they enter the second sync barrier. This
434 // ensure, that no task starts doing work before all data
435 // structures (local and global) have been re-initialised. When they
436 // exit it, they are free to start working again.
437 WorkGangBarrierSync _first_overflow_barrier_sync;
438 WorkGangBarrierSync _second_overflow_barrier_sync;
440 // this is set by any task, when an overflow on the global data
441 // structures is detected.
442 volatile bool _has_overflown;
443 // true: marking is concurrent, false: we're in remark
444 volatile bool _concurrent;
445 // set at the end of a Full GC so that marking aborts
446 volatile bool _has_aborted;
448 // used when remark aborts due to an overflow to indicate that
449 // another concurrent marking phase should start
450 volatile bool _restart_for_overflow;
452 // This is true from the very start of concurrent marking until the
453 // point when all the tasks complete their work. It is really used
454 // to determine the points between the end of concurrent marking and
455 // time of remark.
456 volatile bool _concurrent_marking_in_progress;
458 // verbose level
459 CMVerboseLevel _verbose_level;
461 // All of these times are in ms.
462 NumberSeq _init_times;
463 NumberSeq _remark_times;
464 NumberSeq _remark_mark_times;
465 NumberSeq _remark_weak_ref_times;
466 NumberSeq _cleanup_times;
467 double _total_counting_time;
468 double _total_rs_scrub_time;
470 double* _accum_task_vtime; // accumulated task vtime
472 FlexibleWorkGang* _parallel_workers;
474 ForceOverflowSettings _force_overflow_conc;
475 ForceOverflowSettings _force_overflow_stw;
477 void weakRefsWork(bool clear_all_soft_refs);
479 void swapMarkBitMaps();
481 // It resets the global marking data structures, as well as the
482 // task local ones; should be called during initial mark.
483 void reset();
485 // Resets all the marking data structures. Called when we have to restart
486 // marking or when marking completes (via set_non_marking_state below).
487 void reset_marking_state(bool clear_overflow = true);
489 // We do this after we're done with marking so that the marking data
490 // structures are initialised to a sensible and predictable state.
491 void set_non_marking_state();
493 // Called to indicate how many threads are currently active.
494 void set_concurrency(uint active_tasks);
496 // It should be called to indicate which phase we're in (concurrent
497 // mark or remark) and how many threads are currently active.
498 void set_concurrency_and_phase(uint active_tasks, bool concurrent);
500 // prints all gathered CM-related statistics
501 void print_stats();
503 bool cleanup_list_is_empty() {
504 return _cleanup_list.is_empty();
505 }
507 // accessor methods
508 uint parallel_marking_threads() const { return _parallel_marking_threads; }
509 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
510 double sleep_factor() { return _sleep_factor; }
511 double marking_task_overhead() { return _marking_task_overhead;}
512 double cleanup_sleep_factor() { return _cleanup_sleep_factor; }
513 double cleanup_task_overhead() { return _cleanup_task_overhead;}
515 bool use_parallel_marking_threads() const {
516 assert(parallel_marking_threads() <=
517 max_parallel_marking_threads(), "sanity");
518 assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
519 parallel_marking_threads() > 0,
520 "parallel workers not set up correctly");
521 return _parallel_workers != NULL;
522 }
524 HeapWord* finger() { return _finger; }
525 bool concurrent() { return _concurrent; }
526 uint active_tasks() { return _active_tasks; }
527 ParallelTaskTerminator* terminator() { return &_terminator; }
529 // It claims the next available region to be scanned by a marking
530 // task/thread. It might return NULL if the next region is empty or
531 // we have run out of regions. In the latter case, out_of_regions()
532 // determines whether we've really run out of regions or the task
533 // should call claim_region() again. This might seem a bit
534 // awkward. Originally, the code was written so that claim_region()
535 // either successfully returned with a non-empty region or there
536 // were no more regions to be claimed. The problem with this was
537 // that, in certain circumstances, it iterated over large chunks of
538 // the heap finding only empty regions and, while it was working, it
539 // was preventing the calling task to call its regular clock
540 // method. So, this way, each task will spend very little time in
541 // claim_region() and is allowed to call the regular clock method
542 // frequently.
543 HeapRegion* claim_region(uint worker_id);
545 // It determines whether we've run out of regions to scan.
546 bool out_of_regions() { return _finger == _heap_end; }
548 // Returns the task with the given id
549 CMTask* task(int id) {
550 assert(0 <= id && id < (int) _active_tasks,
551 "task id not within active bounds");
552 return _tasks[id];
553 }
555 // Returns the task queue with the given id
556 CMTaskQueue* task_queue(int id) {
557 assert(0 <= id && id < (int) _active_tasks,
558 "task queue id not within active bounds");
559 return (CMTaskQueue*) _task_queues->queue(id);
560 }
562 // Returns the task queue set
563 CMTaskQueueSet* task_queues() { return _task_queues; }
565 // Access / manipulation of the overflow flag which is set to
566 // indicate that the global stack has overflown
567 bool has_overflown() { return _has_overflown; }
568 void set_has_overflown() { _has_overflown = true; }
569 void clear_has_overflown() { _has_overflown = false; }
570 bool restart_for_overflow() { return _restart_for_overflow; }
572 // Methods to enter the two overflow sync barriers
573 void enter_first_sync_barrier(uint worker_id);
574 void enter_second_sync_barrier(uint worker_id);
576 ForceOverflowSettings* force_overflow_conc() {
577 return &_force_overflow_conc;
578 }
580 ForceOverflowSettings* force_overflow_stw() {
581 return &_force_overflow_stw;
582 }
584 ForceOverflowSettings* force_overflow() {
585 if (concurrent()) {
586 return force_overflow_conc();
587 } else {
588 return force_overflow_stw();
589 }
590 }
592 // Live Data Counting data structures...
593 // These data structures are initialized at the start of
594 // marking. They are written to while marking is active.
595 // They are aggregated during remark; the aggregated values
596 // are then used to populate the _region_bm, _card_bm, and
597 // the total live bytes, which are then subsequently updated
598 // during cleanup.
600 // An array of bitmaps (one bit map per task). Each bitmap
601 // is used to record the cards spanned by the live objects
602 // marked by that task/worker.
603 BitMap* _count_card_bitmaps;
605 // Used to record the number of marked live bytes
606 // (for each region, by worker thread).
607 size_t** _count_marked_bytes;
609 // Card index of the bottom of the G1 heap. Used for biasing indices into
610 // the card bitmaps.
611 intptr_t _heap_bottom_card_num;
613 // Set to true when initialization is complete
614 bool _completed_initialization;
616 public:
617 // Manipulation of the global mark stack.
618 // Notice that the first mark_stack_push is CAS-based, whereas the
619 // two below are Mutex-based. This is OK since the first one is only
620 // called during evacuation pauses and doesn't compete with the
621 // other two (which are called by the marking tasks during
622 // concurrent marking or remark).
623 bool mark_stack_push(oop p) {
624 _markStack.par_push(p);
625 if (_markStack.overflow()) {
626 set_has_overflown();
627 return false;
628 }
629 return true;
630 }
631 bool mark_stack_push(oop* arr, int n) {
632 _markStack.par_push_arr(arr, n);
633 if (_markStack.overflow()) {
634 set_has_overflown();
635 return false;
636 }
637 return true;
638 }
639 void mark_stack_pop(oop* arr, int max, int* n) {
640 _markStack.par_pop_arr(arr, max, n);
641 }
642 size_t mark_stack_size() { return _markStack.size(); }
643 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
644 bool mark_stack_overflow() { return _markStack.overflow(); }
645 bool mark_stack_empty() { return _markStack.isEmpty(); }
647 CMRootRegions* root_regions() { return &_root_regions; }
649 bool concurrent_marking_in_progress() {
650 return _concurrent_marking_in_progress;
651 }
652 void set_concurrent_marking_in_progress() {
653 _concurrent_marking_in_progress = true;
654 }
655 void clear_concurrent_marking_in_progress() {
656 _concurrent_marking_in_progress = false;
657 }
659 void update_accum_task_vtime(int i, double vtime) {
660 _accum_task_vtime[i] += vtime;
661 }
663 double all_task_accum_vtime() {
664 double ret = 0.0;
665 for (uint i = 0; i < _max_worker_id; ++i)
666 ret += _accum_task_vtime[i];
667 return ret;
668 }
670 // Attempts to steal an object from the task queues of other tasks
671 bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
672 return _task_queues->steal(worker_id, hash_seed, obj);
673 }
675 ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
676 ~ConcurrentMark();
678 ConcurrentMarkThread* cmThread() { return _cmThread; }
680 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
681 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; }
683 // Returns the number of GC threads to be used in a concurrent
684 // phase based on the number of GC threads being used in a STW
685 // phase.
686 uint scale_parallel_threads(uint n_par_threads);
688 // Calculates the number of GC threads to be used in a concurrent phase.
689 uint calc_parallel_marking_threads();
691 // The following three are interaction between CM and
692 // G1CollectedHeap
694 // This notifies CM that a root during initial-mark needs to be
695 // grayed. It is MT-safe. word_size is the size of the object in
696 // words. It is passed explicitly as sometimes we cannot calculate
697 // it from the given object because it might be in an inconsistent
698 // state (e.g., in to-space and being copied). So the caller is
699 // responsible for dealing with this issue (e.g., get the size from
700 // the from-space image when the to-space image might be
701 // inconsistent) and always passing the size. hr is the region that
702 // contains the object and it's passed optionally from callers who
703 // might already have it (no point in recalculating it).
704 inline void grayRoot(oop obj, size_t word_size,
705 uint worker_id, HeapRegion* hr = NULL);
707 // It iterates over the heap and for each object it comes across it
708 // will dump the contents of its reference fields, as well as
709 // liveness information for the object and its referents. The dump
710 // will be written to a file with the following name:
711 // G1PrintReachableBaseFile + "." + str.
712 // vo decides whether the prev (vo == UsePrevMarking), the next
713 // (vo == UseNextMarking) marking information, or the mark word
714 // (vo == UseMarkWord) will be used to determine the liveness of
715 // each object / referent.
716 // If all is true, all objects in the heap will be dumped, otherwise
717 // only the live ones. In the dump the following symbols / breviations
718 // are used:
719 // M : an explicitly live object (its bitmap bit is set)
720 // > : an implicitly live object (over tams)
721 // O : an object outside the G1 heap (typically: in the perm gen)
722 // NOT : a reference field whose referent is not live
723 // AND MARKED : indicates that an object is both explicitly and
724 // implicitly live (it should be one or the other, not both)
725 void print_reachable(const char* str,
726 VerifyOption vo, bool all) PRODUCT_RETURN;
728 // Clear the next marking bitmap (will be called concurrently).
729 void clearNextBitmap();
731 // These two do the work that needs to be done before and after the
732 // initial root checkpoint. Since this checkpoint can be done at two
733 // different points (i.e. an explicit pause or piggy-backed on a
734 // young collection), then it's nice to be able to easily share the
735 // pre/post code. It might be the case that we can put everything in
736 // the post method. TP
737 void checkpointRootsInitialPre();
738 void checkpointRootsInitialPost();
740 // Scan all the root regions and mark everything reachable from
741 // them.
742 void scanRootRegions();
744 // Scan a single root region and mark everything reachable from it.
745 void scanRootRegion(HeapRegion* hr, uint worker_id);
747 // Do concurrent phase of marking, to a tentative transitive closure.
748 void markFromRoots();
750 void checkpointRootsFinal(bool clear_all_soft_refs);
751 void checkpointRootsFinalWork();
752 void cleanup();
753 void completeCleanup();
755 // Mark in the previous bitmap. NB: this is usually read-only, so use
756 // this carefully!
757 inline void markPrev(oop p);
759 // Clears marks for all objects in the given range, for the prev,
760 // next, or both bitmaps. NB: the previous bitmap is usually
761 // read-only, so use this carefully!
762 void clearRangePrevBitmap(MemRegion mr);
763 void clearRangeNextBitmap(MemRegion mr);
764 void clearRangeBothBitmaps(MemRegion mr);
766 // Notify data structures that a GC has started.
767 void note_start_of_gc() {
768 _markStack.note_start_of_gc();
769 }
771 // Notify data structures that a GC is finished.
772 void note_end_of_gc() {
773 _markStack.note_end_of_gc();
774 }
776 // Verify that there are no CSet oops on the stacks (taskqueues /
777 // global mark stack), enqueued SATB buffers, per-thread SATB
778 // buffers, and fingers (global / per-task). The boolean parameters
779 // decide which of the above data structures to verify. If marking
780 // is not in progress, it's a no-op.
781 void verify_no_cset_oops(bool verify_stacks,
782 bool verify_enqueued_buffers,
783 bool verify_thread_buffers,
784 bool verify_fingers) PRODUCT_RETURN;
786 // It is called at the end of an evacuation pause during marking so
787 // that CM is notified of where the new end of the heap is. It
788 // doesn't do anything if concurrent_marking_in_progress() is false,
789 // unless the force parameter is true.
790 void update_g1_committed(bool force = false);
792 bool isMarked(oop p) const {
793 assert(p != NULL && p->is_oop(), "expected an oop");
794 HeapWord* addr = (HeapWord*)p;
795 assert(addr >= _nextMarkBitMap->startWord() ||
796 addr < _nextMarkBitMap->endWord(), "in a region");
798 return _nextMarkBitMap->isMarked(addr);
799 }
801 inline bool not_yet_marked(oop p) const;
803 // XXX Debug code
804 bool containing_card_is_marked(void* p);
805 bool containing_cards_are_marked(void* start, void* last);
807 bool isPrevMarked(oop p) const {
808 assert(p != NULL && p->is_oop(), "expected an oop");
809 HeapWord* addr = (HeapWord*)p;
810 assert(addr >= _prevMarkBitMap->startWord() ||
811 addr < _prevMarkBitMap->endWord(), "in a region");
813 return _prevMarkBitMap->isMarked(addr);
814 }
816 inline bool do_yield_check(uint worker_i = 0);
817 inline bool should_yield();
819 // Called to abort the marking cycle after a Full GC takes palce.
820 void abort();
822 bool has_aborted() { return _has_aborted; }
824 // This prints the global/local fingers. It is used for debugging.
825 NOT_PRODUCT(void print_finger();)
827 void print_summary_info();
829 void print_worker_threads_on(outputStream* st) const;
831 void print_on_error(outputStream* st) const;
833 // The following indicate whether a given verbose level has been
834 // set. Notice that anything above stats is conditional to
835 // _MARKING_VERBOSE_ having been set to 1
836 bool verbose_stats() {
837 return _verbose_level >= stats_verbose;
838 }
839 bool verbose_low() {
840 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
841 }
842 bool verbose_medium() {
843 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
844 }
845 bool verbose_high() {
846 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
847 }
849 // Liveness counting
851 // Utility routine to set an exclusive range of cards on the given
852 // card liveness bitmap
853 inline void set_card_bitmap_range(BitMap* card_bm,
854 BitMap::idx_t start_idx,
855 BitMap::idx_t end_idx,
856 bool is_par);
858 // Returns the card number of the bottom of the G1 heap.
859 // Used in biasing indices into accounting card bitmaps.
860 intptr_t heap_bottom_card_num() const {
861 return _heap_bottom_card_num;
862 }
864 // Returns the card bitmap for a given task or worker id.
865 BitMap* count_card_bitmap_for(uint worker_id) {
866 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
867 assert(_count_card_bitmaps != NULL, "uninitialized");
868 BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
869 assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
870 return task_card_bm;
871 }
873 // Returns the array containing the marked bytes for each region,
874 // for the given worker or task id.
875 size_t* count_marked_bytes_array_for(uint worker_id) {
876 assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
877 assert(_count_marked_bytes != NULL, "uninitialized");
878 size_t* marked_bytes_array = _count_marked_bytes[worker_id];
879 assert(marked_bytes_array != NULL, "uninitialized");
880 return marked_bytes_array;
881 }
883 // Returns the index in the liveness accounting card table bitmap
884 // for the given address
885 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
887 // Counts the size of the given memory region in the the given
888 // marked_bytes array slot for the given HeapRegion.
889 // Sets the bits in the given card bitmap that are associated with the
890 // cards that are spanned by the memory region.
891 inline void count_region(MemRegion mr, HeapRegion* hr,
892 size_t* marked_bytes_array,
893 BitMap* task_card_bm);
895 // Counts the given memory region in the task/worker counting
896 // data structures for the given worker id.
897 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
899 // Counts the given memory region in the task/worker counting
900 // data structures for the given worker id.
901 inline void count_region(MemRegion mr, uint worker_id);
903 // Counts the given object in the given task/worker counting
904 // data structures.
905 inline void count_object(oop obj, HeapRegion* hr,
906 size_t* marked_bytes_array,
907 BitMap* task_card_bm);
909 // Counts the given object in the task/worker counting data
910 // structures for the given worker id.
911 inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
913 // Attempts to mark the given object and, if successful, counts
914 // the object in the given task/worker counting structures.
915 inline bool par_mark_and_count(oop obj, HeapRegion* hr,
916 size_t* marked_bytes_array,
917 BitMap* task_card_bm);
919 // Attempts to mark the given object and, if successful, counts
920 // the object in the task/worker counting structures for the
921 // given worker id.
922 inline bool par_mark_and_count(oop obj, size_t word_size,
923 HeapRegion* hr, uint worker_id);
925 // Attempts to mark the given object and, if successful, counts
926 // the object in the task/worker counting structures for the
927 // given worker id.
928 inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
930 // Similar to the above routine but we don't know the heap region that
931 // contains the object to be marked/counted, which this routine looks up.
932 inline bool par_mark_and_count(oop obj, uint worker_id);
934 // Similar to the above routine but there are times when we cannot
935 // safely calculate the size of obj due to races and we, therefore,
936 // pass the size in as a parameter. It is the caller's reponsibility
937 // to ensure that the size passed in for obj is valid.
938 inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
940 // Unconditionally mark the given object, and unconditinally count
941 // the object in the counting structures for worker id 0.
942 // Should *not* be called from parallel code.
943 inline bool mark_and_count(oop obj, HeapRegion* hr);
945 // Similar to the above routine but we don't know the heap region that
946 // contains the object to be marked/counted, which this routine looks up.
947 // Should *not* be called from parallel code.
948 inline bool mark_and_count(oop obj);
950 // Returns true if initialization was successfully completed.
951 bool completed_initialization() const {
952 return _completed_initialization;
953 }
955 protected:
956 // Clear all the per-task bitmaps and arrays used to store the
957 // counting data.
958 void clear_all_count_data();
960 // Aggregates the counting data for each worker/task
961 // that was constructed while marking. Also sets
962 // the amount of marked bytes for each region and
963 // the top at concurrent mark count.
964 void aggregate_count_data();
966 // Verification routine
967 void verify_count_data();
968 };
970 // A class representing a marking task.
971 class CMTask : public TerminatorTerminator {
972 private:
973 enum PrivateConstants {
974 // the regular clock call is called once the scanned words reaches
975 // this limit
976 words_scanned_period = 12*1024,
977 // the regular clock call is called once the number of visited
978 // references reaches this limit
979 refs_reached_period = 384,
980 // initial value for the hash seed, used in the work stealing code
981 init_hash_seed = 17,
982 // how many entries will be transferred between global stack and
983 // local queues
984 global_stack_transfer_size = 16
985 };
987 uint _worker_id;
988 G1CollectedHeap* _g1h;
989 ConcurrentMark* _cm;
990 CMBitMap* _nextMarkBitMap;
991 // the task queue of this task
992 CMTaskQueue* _task_queue;
993 private:
994 // the task queue set---needed for stealing
995 CMTaskQueueSet* _task_queues;
996 // indicates whether the task has been claimed---this is only for
997 // debugging purposes
998 bool _claimed;
1000 // number of calls to this task
1001 int _calls;
1003 // when the virtual timer reaches this time, the marking step should
1004 // exit
1005 double _time_target_ms;
1006 // the start time of the current marking step
1007 double _start_time_ms;
1009 // the oop closure used for iterations over oops
1010 G1CMOopClosure* _cm_oop_closure;
1012 // the region this task is scanning, NULL if we're not scanning any
1013 HeapRegion* _curr_region;
1014 // the local finger of this task, NULL if we're not scanning a region
1015 HeapWord* _finger;
1016 // limit of the region this task is scanning, NULL if we're not scanning one
1017 HeapWord* _region_limit;
1019 // the number of words this task has scanned
1020 size_t _words_scanned;
1021 // When _words_scanned reaches this limit, the regular clock is
1022 // called. Notice that this might be decreased under certain
1023 // circumstances (i.e. when we believe that we did an expensive
1024 // operation).
1025 size_t _words_scanned_limit;
1026 // the initial value of _words_scanned_limit (i.e. what it was
1027 // before it was decreased).
1028 size_t _real_words_scanned_limit;
1030 // the number of references this task has visited
1031 size_t _refs_reached;
1032 // When _refs_reached reaches this limit, the regular clock is
1033 // called. Notice this this might be decreased under certain
1034 // circumstances (i.e. when we believe that we did an expensive
1035 // operation).
1036 size_t _refs_reached_limit;
1037 // the initial value of _refs_reached_limit (i.e. what it was before
1038 // it was decreased).
1039 size_t _real_refs_reached_limit;
1041 // used by the work stealing stuff
1042 int _hash_seed;
1043 // if this is true, then the task has aborted for some reason
1044 bool _has_aborted;
1045 // set when the task aborts because it has met its time quota
1046 bool _has_timed_out;
1047 // true when we're draining SATB buffers; this avoids the task
1048 // aborting due to SATB buffers being available (as we're already
1049 // dealing with them)
1050 bool _draining_satb_buffers;
1052 // number sequence of past step times
1053 NumberSeq _step_times_ms;
1054 // elapsed time of this task
1055 double _elapsed_time_ms;
1056 // termination time of this task
1057 double _termination_time_ms;
1058 // when this task got into the termination protocol
1059 double _termination_start_time_ms;
1061 // true when the task is during a concurrent phase, false when it is
1062 // in the remark phase (so, in the latter case, we do not have to
1063 // check all the things that we have to check during the concurrent
1064 // phase, i.e. SATB buffer availability...)
1065 bool _concurrent;
1067 TruncatedSeq _marking_step_diffs_ms;
1069 // Counting data structures. Embedding the task's marked_bytes_array
1070 // and card bitmap into the actual task saves having to go through
1071 // the ConcurrentMark object.
1072 size_t* _marked_bytes_array;
1073 BitMap* _card_bm;
1075 // LOTS of statistics related with this task
1076 #if _MARKING_STATS_
1077 NumberSeq _all_clock_intervals_ms;
1078 double _interval_start_time_ms;
1080 int _aborted;
1081 int _aborted_overflow;
1082 int _aborted_cm_aborted;
1083 int _aborted_yield;
1084 int _aborted_timed_out;
1085 int _aborted_satb;
1086 int _aborted_termination;
1088 int _steal_attempts;
1089 int _steals;
1091 int _clock_due_to_marking;
1092 int _clock_due_to_scanning;
1094 int _local_pushes;
1095 int _local_pops;
1096 int _local_max_size;
1097 int _objs_scanned;
1099 int _global_pushes;
1100 int _global_pops;
1101 int _global_max_size;
1103 int _global_transfers_to;
1104 int _global_transfers_from;
1106 int _regions_claimed;
1107 int _objs_found_on_bitmap;
1109 int _satb_buffers_processed;
1110 #endif // _MARKING_STATS_
1112 // it updates the local fields after this task has claimed
1113 // a new region to scan
1114 void setup_for_region(HeapRegion* hr);
1115 // it brings up-to-date the limit of the region
1116 void update_region_limit();
1118 // called when either the words scanned or the refs visited limit
1119 // has been reached
1120 void reached_limit();
1121 // recalculates the words scanned and refs visited limits
1122 void recalculate_limits();
1123 // decreases the words scanned and refs visited limits when we reach
1124 // an expensive operation
1125 void decrease_limits();
1126 // it checks whether the words scanned or refs visited reached their
1127 // respective limit and calls reached_limit() if they have
1128 void check_limits() {
1129 if (_words_scanned >= _words_scanned_limit ||
1130 _refs_reached >= _refs_reached_limit) {
1131 reached_limit();
1132 }
1133 }
1134 // this is supposed to be called regularly during a marking step as
1135 // it checks a bunch of conditions that might cause the marking step
1136 // to abort
1137 void regular_clock_call();
1138 bool concurrent() { return _concurrent; }
1140 public:
1141 // It resets the task; it should be called right at the beginning of
1142 // a marking phase.
1143 void reset(CMBitMap* _nextMarkBitMap);
1144 // it clears all the fields that correspond to a claimed region.
1145 void clear_region_fields();
1147 void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1149 // The main method of this class which performs a marking step
1150 // trying not to exceed the given duration. However, it might exit
1151 // prematurely, according to some conditions (i.e. SATB buffers are
1152 // available for processing).
1153 void do_marking_step(double target_ms,
1154 bool do_termination,
1155 bool is_serial);
1157 // These two calls start and stop the timer
1158 void record_start_time() {
1159 _elapsed_time_ms = os::elapsedTime() * 1000.0;
1160 }
1161 void record_end_time() {
1162 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1163 }
1165 // returns the worker ID associated with this task.
1166 uint worker_id() { return _worker_id; }
1168 // From TerminatorTerminator. It determines whether this task should
1169 // exit the termination protocol after it's entered it.
1170 virtual bool should_exit_termination();
1172 // Resets the local region fields after a task has finished scanning a
1173 // region; or when they have become stale as a result of the region
1174 // being evacuated.
1175 void giveup_current_region();
1177 HeapWord* finger() { return _finger; }
1179 bool has_aborted() { return _has_aborted; }
1180 void set_has_aborted() { _has_aborted = true; }
1181 void clear_has_aborted() { _has_aborted = false; }
1182 bool has_timed_out() { return _has_timed_out; }
1183 bool claimed() { return _claimed; }
1185 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1187 // It grays the object by marking it and, if necessary, pushing it
1188 // on the local queue
1189 inline void deal_with_reference(oop obj);
1191 // It scans an object and visits its children.
1192 void scan_object(oop obj);
1194 // It pushes an object on the local queue.
1195 inline void push(oop obj);
1197 // These two move entries to/from the global stack.
1198 void move_entries_to_global_stack();
1199 void get_entries_from_global_stack();
1201 // It pops and scans objects from the local queue. If partially is
1202 // true, then it stops when the queue size is of a given limit. If
1203 // partially is false, then it stops when the queue is empty.
1204 void drain_local_queue(bool partially);
1205 // It moves entries from the global stack to the local queue and
1206 // drains the local queue. If partially is true, then it stops when
1207 // both the global stack and the local queue reach a given size. If
1208 // partially if false, it tries to empty them totally.
1209 void drain_global_stack(bool partially);
1210 // It keeps picking SATB buffers and processing them until no SATB
1211 // buffers are available.
1212 void drain_satb_buffers();
1214 // moves the local finger to a new location
1215 inline void move_finger_to(HeapWord* new_finger) {
1216 assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1217 _finger = new_finger;
1218 }
1220 CMTask(uint worker_id, ConcurrentMark *cm,
1221 size_t* marked_bytes, BitMap* card_bm,
1222 CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1224 // it prints statistics associated with this task
1225 void print_stats();
1227 #if _MARKING_STATS_
1228 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1229 #endif // _MARKING_STATS_
1230 };
1232 // Class that's used to to print out per-region liveness
1233 // information. It's currently used at the end of marking and also
1234 // after we sort the old regions at the end of the cleanup operation.
1235 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1236 private:
1237 outputStream* _out;
1239 // Accumulators for these values.
1240 size_t _total_used_bytes;
1241 size_t _total_capacity_bytes;
1242 size_t _total_prev_live_bytes;
1243 size_t _total_next_live_bytes;
1245 // These are set up when we come across a "stars humongous" region
1246 // (as this is where most of this information is stored, not in the
1247 // subsequent "continues humongous" regions). After that, for every
1248 // region in a given humongous region series we deduce the right
1249 // values for it by simply subtracting the appropriate amount from
1250 // these fields. All these values should reach 0 after we've visited
1251 // the last region in the series.
1252 size_t _hum_used_bytes;
1253 size_t _hum_capacity_bytes;
1254 size_t _hum_prev_live_bytes;
1255 size_t _hum_next_live_bytes;
1257 // Accumulator for the remembered set size
1258 size_t _total_remset_bytes;
1260 // Accumulator for strong code roots memory size
1261 size_t _total_strong_code_roots_bytes;
1263 static double perc(size_t val, size_t total) {
1264 if (total == 0) {
1265 return 0.0;
1266 } else {
1267 return 100.0 * ((double) val / (double) total);
1268 }
1269 }
1271 static double bytes_to_mb(size_t val) {
1272 return (double) val / (double) M;
1273 }
1275 // See the .cpp file.
1276 size_t get_hum_bytes(size_t* hum_bytes);
1277 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1278 size_t* prev_live_bytes, size_t* next_live_bytes);
1280 public:
1281 // The header and footer are printed in the constructor and
1282 // destructor respectively.
1283 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1284 virtual bool doHeapRegion(HeapRegion* r);
1285 ~G1PrintRegionLivenessInfoClosure();
1286 };
1288 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP