Thu, 20 Nov 2008 16:56:09 -0800
6684579: SoftReference processing can be made more efficient
Summary: For current soft-ref clearing policies, we can decide at marking time if a soft-reference will definitely not be cleared, postponing the decision of whether it will definitely be cleared to the final reference processing phase. This can be especially beneficial in the case of concurrent collectors where the marking is usually concurrent but reference processing is usually not.
Reviewed-by: jmasa
1 /*
2 * Copyright 2001-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // ConcurrentMarkSweepGeneration is in support of a concurrent
26 // mark-sweep old generation in the Detlefs-Printezis--Boehm-Demers-Schenker
27 // style. We assume, for now, that this generation is always the
28 // seniormost generation (modulo the PermGeneration), and for simplicity
29 // in the first implementation, that this generation is a single compactible
30 // space. Neither of these restrictions appears essential, and will be
31 // relaxed in the future when more time is available to implement the
32 // greater generality (and there's a need for it).
33 //
34 // Concurrent mode failures are currently handled by
35 // means of a sliding mark-compact.
37 class CMSAdaptiveSizePolicy;
38 class CMSConcMarkingTask;
39 class CMSGCAdaptivePolicyCounters;
40 class ConcurrentMarkSweepGeneration;
41 class ConcurrentMarkSweepPolicy;
42 class ConcurrentMarkSweepThread;
43 class CompactibleFreeListSpace;
44 class FreeChunk;
45 class PromotionInfo;
46 class ScanMarkedObjectsAgainCarefullyClosure;
48 // A generic CMS bit map. It's the basis for both the CMS marking bit map
49 // as well as for the mod union table (in each case only a subset of the
50 // methods are used). This is essentially a wrapper around the BitMap class,
51 // with one bit per (1<<_shifter) HeapWords. (i.e. for the marking bit map,
52 // we have _shifter == 0. and for the mod union table we have
53 // shifter == CardTableModRefBS::card_shift - LogHeapWordSize.)
54 // XXX 64-bit issues in BitMap?
55 class CMSBitMap VALUE_OBJ_CLASS_SPEC {
56 friend class VMStructs;
58 HeapWord* _bmStartWord; // base address of range covered by map
59 size_t _bmWordSize; // map size (in #HeapWords covered)
60 const int _shifter; // shifts to convert HeapWord to bit position
61 VirtualSpace _virtual_space; // underlying the bit map
62 BitMap _bm; // the bit map itself
63 public:
64 Mutex* const _lock; // mutex protecting _bm;
66 public:
67 // constructor
68 CMSBitMap(int shifter, int mutex_rank, const char* mutex_name);
70 // allocates the actual storage for the map
71 bool allocate(MemRegion mr);
72 // field getter
73 Mutex* lock() const { return _lock; }
74 // locking verifier convenience function
75 void assert_locked() const PRODUCT_RETURN;
77 // inquiries
78 HeapWord* startWord() const { return _bmStartWord; }
79 size_t sizeInWords() const { return _bmWordSize; }
80 size_t sizeInBits() const { return _bm.size(); }
81 // the following is one past the last word in space
82 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; }
84 // reading marks
85 bool isMarked(HeapWord* addr) const;
86 bool par_isMarked(HeapWord* addr) const; // do not lock checks
87 bool isUnmarked(HeapWord* addr) const;
88 bool isAllClear() const;
90 // writing marks
91 void mark(HeapWord* addr);
92 // For marking by parallel GC threads;
93 // returns true if we did, false if another thread did
94 bool par_mark(HeapWord* addr);
96 void mark_range(MemRegion mr);
97 void par_mark_range(MemRegion mr);
98 void mark_large_range(MemRegion mr);
99 void par_mark_large_range(MemRegion mr);
100 void par_clear(HeapWord* addr); // For unmarking by parallel GC threads.
101 void clear_range(MemRegion mr);
102 void par_clear_range(MemRegion mr);
103 void clear_large_range(MemRegion mr);
104 void par_clear_large_range(MemRegion mr);
105 void clear_all();
106 void clear_all_incrementally(); // Not yet implemented!!
108 NOT_PRODUCT(
109 // checks the memory region for validity
110 void region_invariant(MemRegion mr);
111 )
113 // iteration
114 void iterate(BitMapClosure* cl) {
115 _bm.iterate(cl);
116 }
117 void iterate(BitMapClosure* cl, HeapWord* left, HeapWord* right);
118 void dirty_range_iterate_clear(MemRegionClosure* cl);
119 void dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl);
121 // auxiliary support for iteration
122 HeapWord* getNextMarkedWordAddress(HeapWord* addr) const;
123 HeapWord* getNextMarkedWordAddress(HeapWord* start_addr,
124 HeapWord* end_addr) const;
125 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr) const;
126 HeapWord* getNextUnmarkedWordAddress(HeapWord* start_addr,
127 HeapWord* end_addr) const;
128 MemRegion getAndClearMarkedRegion(HeapWord* addr);
129 MemRegion getAndClearMarkedRegion(HeapWord* start_addr,
130 HeapWord* end_addr);
132 // conversion utilities
133 HeapWord* offsetToHeapWord(size_t offset) const;
134 size_t heapWordToOffset(HeapWord* addr) const;
135 size_t heapWordDiffToOffsetDiff(size_t diff) const;
137 // debugging
138 // is this address range covered by the bit-map?
139 NOT_PRODUCT(
140 bool covers(MemRegion mr) const;
141 bool covers(HeapWord* start, size_t size = 0) const;
142 )
143 void verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) PRODUCT_RETURN;
144 };
146 // Represents a marking stack used by the CMS collector.
147 // Ideally this should be GrowableArray<> just like MSC's marking stack(s).
148 class CMSMarkStack: public CHeapObj {
149 //
150 friend class CMSCollector; // to get at expasion stats further below
151 //
153 VirtualSpace _virtual_space; // space for the stack
154 oop* _base; // bottom of stack
155 size_t _index; // one more than last occupied index
156 size_t _capacity; // max #elements
157 Mutex _par_lock; // an advisory lock used in case of parallel access
158 NOT_PRODUCT(size_t _max_depth;) // max depth plumbed during run
160 protected:
161 size_t _hit_limit; // we hit max stack size limit
162 size_t _failed_double; // we failed expansion before hitting limit
164 public:
165 CMSMarkStack():
166 _par_lock(Mutex::event, "CMSMarkStack._par_lock", true),
167 _hit_limit(0),
168 _failed_double(0) {}
170 bool allocate(size_t size);
172 size_t capacity() const { return _capacity; }
174 oop pop() {
175 if (!isEmpty()) {
176 return _base[--_index] ;
177 }
178 return NULL;
179 }
181 bool push(oop ptr) {
182 if (isFull()) {
183 return false;
184 } else {
185 _base[_index++] = ptr;
186 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
187 return true;
188 }
189 }
191 bool isEmpty() const { return _index == 0; }
192 bool isFull() const {
193 assert(_index <= _capacity, "buffer overflow");
194 return _index == _capacity;
195 }
197 size_t length() { return _index; }
199 // "Parallel versions" of some of the above
200 oop par_pop() {
201 // lock and pop
202 MutexLockerEx x(&_par_lock, Mutex::_no_safepoint_check_flag);
203 return pop();
204 }
206 bool par_push(oop ptr) {
207 // lock and push
208 MutexLockerEx x(&_par_lock, Mutex::_no_safepoint_check_flag);
209 return push(ptr);
210 }
212 // Forcibly reset the stack, losing all of its contents.
213 void reset() {
214 _index = 0;
215 }
217 // Expand the stack, typically in response to an overflow condition
218 void expand();
220 // Compute the least valued stack element.
221 oop least_value(HeapWord* low) {
222 oop least = (oop)low;
223 for (size_t i = 0; i < _index; i++) {
224 least = MIN2(least, _base[i]);
225 }
226 return least;
227 }
229 // Exposed here to allow stack expansion in || case
230 Mutex* par_lock() { return &_par_lock; }
231 };
233 class CardTableRS;
234 class CMSParGCThreadState;
236 class ModUnionClosure: public MemRegionClosure {
237 protected:
238 CMSBitMap* _t;
239 public:
240 ModUnionClosure(CMSBitMap* t): _t(t) { }
241 void do_MemRegion(MemRegion mr);
242 };
244 class ModUnionClosurePar: public ModUnionClosure {
245 public:
246 ModUnionClosurePar(CMSBitMap* t): ModUnionClosure(t) { }
247 void do_MemRegion(MemRegion mr);
248 };
250 // Survivor Chunk Array in support of parallelization of
251 // Survivor Space rescan.
252 class ChunkArray: public CHeapObj {
253 size_t _index;
254 size_t _capacity;
255 HeapWord** _array; // storage for array
257 public:
258 ChunkArray() : _index(0), _capacity(0), _array(NULL) {}
259 ChunkArray(HeapWord** a, size_t c):
260 _index(0), _capacity(c), _array(a) {}
262 HeapWord** array() { return _array; }
263 void set_array(HeapWord** a) { _array = a; }
265 size_t capacity() { return _capacity; }
266 void set_capacity(size_t c) { _capacity = c; }
268 size_t end() {
269 assert(_index < capacity(), "_index out of bounds");
270 return _index;
271 } // exclusive
273 HeapWord* nth(size_t n) {
274 assert(n < end(), "Out of bounds access");
275 return _array[n];
276 }
278 void reset() {
279 _index = 0;
280 }
282 void record_sample(HeapWord* p, size_t sz) {
283 // For now we do not do anything with the size
284 if (_index < _capacity) {
285 _array[_index++] = p;
286 }
287 }
288 };
290 //
291 // Timing, allocation and promotion statistics for gc scheduling and incremental
292 // mode pacing. Most statistics are exponential averages.
293 //
294 class CMSStats VALUE_OBJ_CLASS_SPEC {
295 private:
296 ConcurrentMarkSweepGeneration* const _cms_gen; // The cms (old) gen.
298 // The following are exponential averages with factor alpha:
299 // avg = (100 - alpha) * avg + alpha * cur_sample
300 //
301 // The durations measure: end_time[n] - start_time[n]
302 // The periods measure: start_time[n] - start_time[n-1]
303 //
304 // The cms period and duration include only concurrent collections; time spent
305 // in foreground cms collections due to System.gc() or because of a failure to
306 // keep up are not included.
307 //
308 // There are 3 alphas to "bootstrap" the statistics. The _saved_alpha is the
309 // real value, but is used only after the first period. A value of 100 is
310 // used for the first sample so it gets the entire weight.
311 unsigned int _saved_alpha; // 0-100
312 unsigned int _gc0_alpha;
313 unsigned int _cms_alpha;
315 double _gc0_duration;
316 double _gc0_period;
317 size_t _gc0_promoted; // bytes promoted per gc0
318 double _cms_duration;
319 double _cms_duration_pre_sweep; // time from initiation to start of sweep
320 double _cms_duration_per_mb;
321 double _cms_period;
322 size_t _cms_allocated; // bytes of direct allocation per gc0 period
324 // Timers.
325 elapsedTimer _cms_timer;
326 TimeStamp _gc0_begin_time;
327 TimeStamp _cms_begin_time;
328 TimeStamp _cms_end_time;
330 // Snapshots of the amount used in the CMS generation.
331 size_t _cms_used_at_gc0_begin;
332 size_t _cms_used_at_gc0_end;
333 size_t _cms_used_at_cms_begin;
335 // Used to prevent the duty cycle from being reduced in the middle of a cms
336 // cycle.
337 bool _allow_duty_cycle_reduction;
339 enum {
340 _GC0_VALID = 0x1,
341 _CMS_VALID = 0x2,
342 _ALL_VALID = _GC0_VALID | _CMS_VALID
343 };
345 unsigned int _valid_bits;
347 unsigned int _icms_duty_cycle; // icms duty cycle (0-100).
349 protected:
351 // Return a duty cycle that avoids wild oscillations, by limiting the amount
352 // of change between old_duty_cycle and new_duty_cycle (the latter is treated
353 // as a recommended value).
354 static unsigned int icms_damped_duty_cycle(unsigned int old_duty_cycle,
355 unsigned int new_duty_cycle);
356 unsigned int icms_update_duty_cycle_impl();
358 public:
359 CMSStats(ConcurrentMarkSweepGeneration* cms_gen,
360 unsigned int alpha = CMSExpAvgFactor);
362 // Whether or not the statistics contain valid data; higher level statistics
363 // cannot be called until this returns true (they require at least one young
364 // gen and one cms cycle to have completed).
365 bool valid() const;
367 // Record statistics.
368 void record_gc0_begin();
369 void record_gc0_end(size_t cms_gen_bytes_used);
370 void record_cms_begin();
371 void record_cms_end();
373 // Allow management of the cms timer, which must be stopped/started around
374 // yield points.
375 elapsedTimer& cms_timer() { return _cms_timer; }
376 void start_cms_timer() { _cms_timer.start(); }
377 void stop_cms_timer() { _cms_timer.stop(); }
379 // Basic statistics; units are seconds or bytes.
380 double gc0_period() const { return _gc0_period; }
381 double gc0_duration() const { return _gc0_duration; }
382 size_t gc0_promoted() const { return _gc0_promoted; }
383 double cms_period() const { return _cms_period; }
384 double cms_duration() const { return _cms_duration; }
385 double cms_duration_per_mb() const { return _cms_duration_per_mb; }
386 size_t cms_allocated() const { return _cms_allocated; }
388 size_t cms_used_at_gc0_end() const { return _cms_used_at_gc0_end;}
390 // Seconds since the last background cms cycle began or ended.
391 double cms_time_since_begin() const;
392 double cms_time_since_end() const;
394 // Higher level statistics--caller must check that valid() returns true before
395 // calling.
397 // Returns bytes promoted per second of wall clock time.
398 double promotion_rate() const;
400 // Returns bytes directly allocated per second of wall clock time.
401 double cms_allocation_rate() const;
403 // Rate at which space in the cms generation is being consumed (sum of the
404 // above two).
405 double cms_consumption_rate() const;
407 // Returns an estimate of the number of seconds until the cms generation will
408 // fill up, assuming no collection work is done.
409 double time_until_cms_gen_full() const;
411 // Returns an estimate of the number of seconds remaining until
412 // the cms generation collection should start.
413 double time_until_cms_start() const;
415 // End of higher level statistics.
417 // Returns the cms incremental mode duty cycle, as a percentage (0-100).
418 unsigned int icms_duty_cycle() const { return _icms_duty_cycle; }
420 // Update the duty cycle and return the new value.
421 unsigned int icms_update_duty_cycle();
423 // Debugging.
424 void print_on(outputStream* st) const PRODUCT_RETURN;
425 void print() const { print_on(gclog_or_tty); }
426 };
428 // A closure related to weak references processing which
429 // we embed in the CMSCollector, since we need to pass
430 // it to the reference processor for secondary filtering
431 // of references based on reachability of referent;
432 // see role of _is_alive_non_header closure in the
433 // ReferenceProcessor class.
434 // For objects in the CMS generation, this closure checks
435 // if the object is "live" (reachable). Used in weak
436 // reference processing.
437 class CMSIsAliveClosure: public BoolObjectClosure {
438 const MemRegion _span;
439 const CMSBitMap* _bit_map;
441 friend class CMSCollector;
442 public:
443 CMSIsAliveClosure(MemRegion span,
444 CMSBitMap* bit_map):
445 _span(span),
446 _bit_map(bit_map) {
447 assert(!span.is_empty(), "Empty span could spell trouble");
448 }
450 void do_object(oop obj) {
451 assert(false, "not to be invoked");
452 }
454 bool do_object_b(oop obj);
455 };
458 // Implements AbstractRefProcTaskExecutor for CMS.
459 class CMSRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
460 public:
462 CMSRefProcTaskExecutor(CMSCollector& collector)
463 : _collector(collector)
464 { }
466 // Executes a task using worker threads.
467 virtual void execute(ProcessTask& task);
468 virtual void execute(EnqueueTask& task);
469 private:
470 CMSCollector& _collector;
471 };
474 class CMSCollector: public CHeapObj {
475 friend class VMStructs;
476 friend class ConcurrentMarkSweepThread;
477 friend class ConcurrentMarkSweepGeneration;
478 friend class CompactibleFreeListSpace;
479 friend class CMSParRemarkTask;
480 friend class CMSConcMarkingTask;
481 friend class CMSRefProcTaskProxy;
482 friend class CMSRefProcTaskExecutor;
483 friend class ScanMarkedObjectsAgainCarefullyClosure; // for sampling eden
484 friend class SurvivorSpacePrecleanClosure; // --- ditto -------
485 friend class PushOrMarkClosure; // to access _restart_addr
486 friend class Par_PushOrMarkClosure; // to access _restart_addr
487 friend class MarkFromRootsClosure; // -- ditto --
488 // ... and for clearing cards
489 friend class Par_MarkFromRootsClosure; // to access _restart_addr
490 // ... and for clearing cards
491 friend class Par_ConcMarkingClosure; // to access _restart_addr etc.
492 friend class MarkFromRootsVerifyClosure; // to access _restart_addr
493 friend class PushAndMarkVerifyClosure; // -- ditto --
494 friend class MarkRefsIntoAndScanClosure; // to access _overflow_list
495 friend class PushAndMarkClosure; // -- ditto --
496 friend class Par_PushAndMarkClosure; // -- ditto --
497 friend class CMSKeepAliveClosure; // -- ditto --
498 friend class CMSDrainMarkingStackClosure; // -- ditto --
499 friend class CMSInnerParMarkAndPushClosure; // -- ditto --
500 NOT_PRODUCT(friend class ScanMarkedObjectsAgainClosure;) // assertion on _overflow_list
501 friend class ReleaseForegroundGC; // to access _foregroundGCShouldWait
502 friend class VM_CMS_Operation;
503 friend class VM_CMS_Initial_Mark;
504 friend class VM_CMS_Final_Remark;
506 private:
507 jlong _time_of_last_gc;
508 void update_time_of_last_gc(jlong now) {
509 _time_of_last_gc = now;
510 }
512 OopTaskQueueSet* _task_queues;
514 // Overflow list of grey objects, threaded through mark-word
515 // Manipulated with CAS in the parallel/multi-threaded case.
516 oop _overflow_list;
517 // The following array-pair keeps track of mark words
518 // displaced for accomodating overflow list above.
519 // This code will likely be revisited under RFE#4922830.
520 GrowableArray<oop>* _preserved_oop_stack;
521 GrowableArray<markOop>* _preserved_mark_stack;
523 int* _hash_seed;
525 // In support of multi-threaded concurrent phases
526 YieldingFlexibleWorkGang* _conc_workers;
528 // Performance Counters
529 CollectorCounters* _gc_counters;
531 // Initialization Errors
532 bool _completed_initialization;
534 // In support of ExplicitGCInvokesConcurrent
535 static bool _full_gc_requested;
536 unsigned int _collection_count_start;
538 // Should we unload classes this concurrent cycle?
539 bool _should_unload_classes;
540 unsigned int _concurrent_cycles_since_last_unload;
541 unsigned int concurrent_cycles_since_last_unload() const {
542 return _concurrent_cycles_since_last_unload;
543 }
544 // Did we (allow) unload classes in the previous concurrent cycle?
545 bool unloaded_classes_last_cycle() const {
546 return concurrent_cycles_since_last_unload() == 0;
547 }
549 // Verification support
550 CMSBitMap _verification_mark_bm;
551 void verify_after_remark_work_1();
552 void verify_after_remark_work_2();
554 // true if any verification flag is on.
555 bool _verifying;
556 bool verifying() const { return _verifying; }
557 void set_verifying(bool v) { _verifying = v; }
559 // Collector policy
560 ConcurrentMarkSweepPolicy* _collector_policy;
561 ConcurrentMarkSweepPolicy* collector_policy() { return _collector_policy; }
563 // Check whether the gc time limit has been
564 // exceeded and set the size policy flag
565 // appropriately.
566 void check_gc_time_limit();
567 // XXX Move these to CMSStats ??? FIX ME !!!
568 elapsedTimer _sweep_timer;
569 AdaptivePaddedAverage _sweep_estimate;
571 protected:
572 ConcurrentMarkSweepGeneration* _cmsGen; // old gen (CMS)
573 ConcurrentMarkSweepGeneration* _permGen; // perm gen
574 MemRegion _span; // span covering above two
575 CardTableRS* _ct; // card table
577 // CMS marking support structures
578 CMSBitMap _markBitMap;
579 CMSBitMap _modUnionTable;
580 CMSMarkStack _markStack;
581 CMSMarkStack _revisitStack; // used to keep track of klassKlass objects
582 // to revisit
583 CMSBitMap _perm_gen_verify_bit_map; // Mark bit map for perm gen verification support.
585 HeapWord* _restart_addr; // in support of marking stack overflow
586 void lower_restart_addr(HeapWord* low);
588 // Counters in support of marking stack / work queue overflow handling:
589 // a non-zero value indicates certain types of overflow events during
590 // the current CMS cycle and could lead to stack resizing efforts at
591 // an opportune future time.
592 size_t _ser_pmc_preclean_ovflw;
593 size_t _ser_pmc_remark_ovflw;
594 size_t _par_pmc_remark_ovflw;
595 size_t _ser_kac_preclean_ovflw;
596 size_t _ser_kac_ovflw;
597 size_t _par_kac_ovflw;
598 NOT_PRODUCT(size_t _num_par_pushes;)
600 // ("Weak") Reference processing support
601 ReferenceProcessor* _ref_processor;
602 CMSIsAliveClosure _is_alive_closure;
603 // keep this textually after _markBitMap and _span; c'tor dependency
605 ConcurrentMarkSweepThread* _cmsThread; // the thread doing the work
606 ModUnionClosure _modUnionClosure;
607 ModUnionClosurePar _modUnionClosurePar;
609 // CMS abstract state machine
610 // initial_state: Idling
611 // next_state(Idling) = {Marking}
612 // next_state(Marking) = {Precleaning, Sweeping}
613 // next_state(Precleaning) = {AbortablePreclean, FinalMarking}
614 // next_state(AbortablePreclean) = {FinalMarking}
615 // next_state(FinalMarking) = {Sweeping}
616 // next_state(Sweeping) = {Resizing}
617 // next_state(Resizing) = {Resetting}
618 // next_state(Resetting) = {Idling}
619 // The numeric values below are chosen so that:
620 // . _collectorState <= Idling == post-sweep && pre-mark
621 // . _collectorState in (Idling, Sweeping) == {initial,final}marking ||
622 // precleaning || abortablePrecleanb
623 enum CollectorState {
624 Resizing = 0,
625 Resetting = 1,
626 Idling = 2,
627 InitialMarking = 3,
628 Marking = 4,
629 Precleaning = 5,
630 AbortablePreclean = 6,
631 FinalMarking = 7,
632 Sweeping = 8
633 };
634 static CollectorState _collectorState;
636 // State related to prologue/epilogue invocation for my generations
637 bool _between_prologue_and_epilogue;
639 // Signalling/State related to coordination between fore- and backgroud GC
640 // Note: When the baton has been passed from background GC to foreground GC,
641 // _foregroundGCIsActive is true and _foregroundGCShouldWait is false.
642 static bool _foregroundGCIsActive; // true iff foreground collector is active or
643 // wants to go active
644 static bool _foregroundGCShouldWait; // true iff background GC is active and has not
645 // yet passed the baton to the foreground GC
647 // Support for CMSScheduleRemark (abortable preclean)
648 bool _abort_preclean;
649 bool _start_sampling;
651 int _numYields;
652 size_t _numDirtyCards;
653 uint _sweepCount;
654 // number of full gc's since the last concurrent gc.
655 uint _full_gcs_since_conc_gc;
657 // occupancy used for bootstrapping stats
658 double _bootstrap_occupancy;
660 // timer
661 elapsedTimer _timer;
663 // Timing, allocation and promotion statistics, used for scheduling.
664 CMSStats _stats;
666 // Allocation limits installed in the young gen, used only in
667 // CMSIncrementalMode. When an allocation in the young gen would cross one of
668 // these limits, the cms generation is notified and the cms thread is started
669 // or stopped, respectively.
670 HeapWord* _icms_start_limit;
671 HeapWord* _icms_stop_limit;
673 enum CMS_op_type {
674 CMS_op_checkpointRootsInitial,
675 CMS_op_checkpointRootsFinal
676 };
678 void do_CMS_operation(CMS_op_type op);
679 bool stop_world_and_do(CMS_op_type op);
681 OopTaskQueueSet* task_queues() { return _task_queues; }
682 int* hash_seed(int i) { return &_hash_seed[i]; }
683 YieldingFlexibleWorkGang* conc_workers() { return _conc_workers; }
685 // Support for parallelizing Eden rescan in CMS remark phase
686 void sample_eden(); // ... sample Eden space top
688 private:
689 // Support for parallelizing young gen rescan in CMS remark phase
690 Generation* _young_gen; // the younger gen
691 HeapWord** _top_addr; // ... Top of Eden
692 HeapWord** _end_addr; // ... End of Eden
693 HeapWord** _eden_chunk_array; // ... Eden partitioning array
694 size_t _eden_chunk_index; // ... top (exclusive) of array
695 size_t _eden_chunk_capacity; // ... max entries in array
697 // Support for parallelizing survivor space rescan
698 HeapWord** _survivor_chunk_array;
699 size_t _survivor_chunk_index;
700 size_t _survivor_chunk_capacity;
701 size_t* _cursor;
702 ChunkArray* _survivor_plab_array;
704 // Support for marking stack overflow handling
705 bool take_from_overflow_list(size_t num, CMSMarkStack* to_stack);
706 bool par_take_from_overflow_list(size_t num, OopTaskQueue* to_work_q);
707 void push_on_overflow_list(oop p);
708 void par_push_on_overflow_list(oop p);
709 // the following is, obviously, not, in general, "MT-stable"
710 bool overflow_list_is_empty() const;
712 void preserve_mark_if_necessary(oop p);
713 void par_preserve_mark_if_necessary(oop p);
714 void preserve_mark_work(oop p, markOop m);
715 void restore_preserved_marks_if_any();
716 NOT_PRODUCT(bool no_preserved_marks() const;)
717 // in support of testing overflow code
718 NOT_PRODUCT(int _overflow_counter;)
719 NOT_PRODUCT(bool simulate_overflow();) // sequential
720 NOT_PRODUCT(bool par_simulate_overflow();) // MT version
722 int _roots_scanning_options;
723 int roots_scanning_options() const { return _roots_scanning_options; }
724 void add_root_scanning_option(int o) { _roots_scanning_options |= o; }
725 void remove_root_scanning_option(int o) { _roots_scanning_options &= ~o; }
727 // CMS work methods
728 void checkpointRootsInitialWork(bool asynch); // initial checkpoint work
730 // a return value of false indicates failure due to stack overflow
731 bool markFromRootsWork(bool asynch); // concurrent marking work
733 public: // FIX ME!!! only for testing
734 bool do_marking_st(bool asynch); // single-threaded marking
735 bool do_marking_mt(bool asynch); // multi-threaded marking
737 private:
739 // concurrent precleaning work
740 size_t preclean_mod_union_table(ConcurrentMarkSweepGeneration* gen,
741 ScanMarkedObjectsAgainCarefullyClosure* cl);
742 size_t preclean_card_table(ConcurrentMarkSweepGeneration* gen,
743 ScanMarkedObjectsAgainCarefullyClosure* cl);
744 // Does precleaning work, returning a quantity indicative of
745 // the amount of "useful work" done.
746 size_t preclean_work(bool clean_refs, bool clean_survivors);
747 void abortable_preclean(); // Preclean while looking for possible abort
748 void initialize_sequential_subtasks_for_young_gen_rescan(int i);
749 // Helper function for above; merge-sorts the per-thread plab samples
750 void merge_survivor_plab_arrays(ContiguousSpace* surv);
751 // Resets (i.e. clears) the per-thread plab sample vectors
752 void reset_survivor_plab_arrays();
754 // final (second) checkpoint work
755 void checkpointRootsFinalWork(bool asynch, bool clear_all_soft_refs,
756 bool init_mark_was_synchronous);
757 // work routine for parallel version of remark
758 void do_remark_parallel();
759 // work routine for non-parallel version of remark
760 void do_remark_non_parallel();
761 // reference processing work routine (during second checkpoint)
762 void refProcessingWork(bool asynch, bool clear_all_soft_refs);
764 // concurrent sweeping work
765 void sweepWork(ConcurrentMarkSweepGeneration* gen, bool asynch);
767 // (concurrent) resetting of support data structures
768 void reset(bool asynch);
770 // Clear _expansion_cause fields of constituent generations
771 void clear_expansion_cause();
773 // An auxilliary method used to record the ends of
774 // used regions of each generation to limit the extent of sweep
775 void save_sweep_limits();
777 // Resize the generations included in the collector.
778 void compute_new_size();
780 // A work method used by foreground collection to determine
781 // what type of collection (compacting or not, continuing or fresh)
782 // it should do.
783 void decide_foreground_collection_type(bool clear_all_soft_refs,
784 bool* should_compact, bool* should_start_over);
786 // A work method used by the foreground collector to do
787 // a mark-sweep-compact.
788 void do_compaction_work(bool clear_all_soft_refs);
790 // A work method used by the foreground collector to do
791 // a mark-sweep, after taking over from a possibly on-going
792 // concurrent mark-sweep collection.
793 void do_mark_sweep_work(bool clear_all_soft_refs,
794 CollectorState first_state, bool should_start_over);
796 // If the backgrould GC is active, acquire control from the background
797 // GC and do the collection.
798 void acquire_control_and_collect(bool full, bool clear_all_soft_refs);
800 // For synchronizing passing of control from background to foreground
801 // GC. waitForForegroundGC() is called by the background
802 // collector. It if had to wait for a foreground collection,
803 // it returns true and the background collection should assume
804 // that the collection was finished by the foreground
805 // collector.
806 bool waitForForegroundGC();
808 // Incremental mode triggering: recompute the icms duty cycle and set the
809 // allocation limits in the young gen.
810 void icms_update_allocation_limits();
812 size_t block_size_using_printezis_bits(HeapWord* addr) const;
813 size_t block_size_if_printezis_bits(HeapWord* addr) const;
814 HeapWord* next_card_start_after_block(HeapWord* addr) const;
816 void setup_cms_unloading_and_verification_state();
817 public:
818 CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
819 ConcurrentMarkSweepGeneration* permGen,
820 CardTableRS* ct,
821 ConcurrentMarkSweepPolicy* cp);
822 ConcurrentMarkSweepThread* cmsThread() { return _cmsThread; }
824 ReferenceProcessor* ref_processor() { return _ref_processor; }
825 void ref_processor_init();
827 Mutex* bitMapLock() const { return _markBitMap.lock(); }
828 static CollectorState abstract_state() { return _collectorState; }
830 bool should_abort_preclean() const; // Whether preclean should be aborted.
831 size_t get_eden_used() const;
832 size_t get_eden_capacity() const;
834 ConcurrentMarkSweepGeneration* cmsGen() { return _cmsGen; }
836 // locking checks
837 NOT_PRODUCT(static bool have_cms_token();)
839 // XXXPERM bool should_collect(bool full, size_t size, bool tlab);
840 bool shouldConcurrentCollect();
842 void collect(bool full,
843 bool clear_all_soft_refs,
844 size_t size,
845 bool tlab);
846 void collect_in_background(bool clear_all_soft_refs);
847 void collect_in_foreground(bool clear_all_soft_refs);
849 // In support of ExplicitGCInvokesConcurrent
850 static void request_full_gc(unsigned int full_gc_count);
851 // Should we unload classes in a particular concurrent cycle?
852 bool should_unload_classes() const {
853 return _should_unload_classes;
854 }
855 bool update_should_unload_classes();
857 void direct_allocated(HeapWord* start, size_t size);
859 // Object is dead if not marked and current phase is sweeping.
860 bool is_dead_obj(oop obj) const;
862 // After a promotion (of "start"), do any necessary marking.
863 // If "par", then it's being done by a parallel GC thread.
864 // The last two args indicate if we need precise marking
865 // and if so the size of the object so it can be dirtied
866 // in its entirety.
867 void promoted(bool par, HeapWord* start,
868 bool is_obj_array, size_t obj_size);
870 HeapWord* allocation_limit_reached(Space* space, HeapWord* top,
871 size_t word_size);
873 void getFreelistLocks() const;
874 void releaseFreelistLocks() const;
875 bool haveFreelistLocks() const;
877 // GC prologue and epilogue
878 void gc_prologue(bool full);
879 void gc_epilogue(bool full);
881 jlong time_of_last_gc(jlong now) {
882 if (_collectorState <= Idling) {
883 // gc not in progress
884 return _time_of_last_gc;
885 } else {
886 // collection in progress
887 return now;
888 }
889 }
891 // Support for parallel remark of survivor space
892 void* get_data_recorder(int thr_num);
894 CMSBitMap* markBitMap() { return &_markBitMap; }
895 void directAllocated(HeapWord* start, size_t size);
897 // main CMS steps and related support
898 void checkpointRootsInitial(bool asynch);
899 bool markFromRoots(bool asynch); // a return value of false indicates failure
900 // due to stack overflow
901 void preclean();
902 void checkpointRootsFinal(bool asynch, bool clear_all_soft_refs,
903 bool init_mark_was_synchronous);
904 void sweep(bool asynch);
906 // Check that the currently executing thread is the expected
907 // one (foreground collector or background collector).
908 void check_correct_thread_executing() PRODUCT_RETURN;
909 // XXXPERM void print_statistics() PRODUCT_RETURN;
911 bool is_cms_reachable(HeapWord* addr);
913 // Performance Counter Support
914 CollectorCounters* counters() { return _gc_counters; }
916 // timer stuff
917 void startTimer() { assert(!_timer.is_active(), "Error"); _timer.start(); }
918 void stopTimer() { assert( _timer.is_active(), "Error"); _timer.stop(); }
919 void resetTimer() { assert(!_timer.is_active(), "Error"); _timer.reset(); }
920 double timerValue() { assert(!_timer.is_active(), "Error"); return _timer.seconds(); }
922 int yields() { return _numYields; }
923 void resetYields() { _numYields = 0; }
924 void incrementYields() { _numYields++; }
925 void resetNumDirtyCards() { _numDirtyCards = 0; }
926 void incrementNumDirtyCards(size_t num) { _numDirtyCards += num; }
927 size_t numDirtyCards() { return _numDirtyCards; }
929 static bool foregroundGCShouldWait() { return _foregroundGCShouldWait; }
930 static void set_foregroundGCShouldWait(bool v) { _foregroundGCShouldWait = v; }
931 static bool foregroundGCIsActive() { return _foregroundGCIsActive; }
932 static void set_foregroundGCIsActive(bool v) { _foregroundGCIsActive = v; }
933 uint sweepCount() const { return _sweepCount; }
934 void incrementSweepCount() { _sweepCount++; }
936 // Timers/stats for gc scheduling and incremental mode pacing.
937 CMSStats& stats() { return _stats; }
939 // Convenience methods that check whether CMSIncrementalMode is enabled and
940 // forward to the corresponding methods in ConcurrentMarkSweepThread.
941 static void start_icms();
942 static void stop_icms(); // Called at the end of the cms cycle.
943 static void disable_icms(); // Called before a foreground collection.
944 static void enable_icms(); // Called after a foreground collection.
945 void icms_wait(); // Called at yield points.
947 // Adaptive size policy
948 CMSAdaptiveSizePolicy* size_policy();
949 CMSGCAdaptivePolicyCounters* gc_adaptive_policy_counters();
951 // debugging
952 void verify(bool);
953 bool verify_after_remark();
954 void verify_ok_to_terminate() const PRODUCT_RETURN;
955 void verify_work_stacks_empty() const PRODUCT_RETURN;
956 void verify_overflow_empty() const PRODUCT_RETURN;
958 // convenience methods in support of debugging
959 static const size_t skip_header_HeapWords() PRODUCT_RETURN0;
960 HeapWord* block_start(const void* p) const PRODUCT_RETURN0;
962 // accessors
963 CMSMarkStack* verification_mark_stack() { return &_markStack; }
964 CMSBitMap* verification_mark_bm() { return &_verification_mark_bm; }
966 // Get the bit map with a perm gen "deadness" information.
967 CMSBitMap* perm_gen_verify_bit_map() { return &_perm_gen_verify_bit_map; }
969 // Initialization errors
970 bool completed_initialization() { return _completed_initialization; }
971 };
973 class CMSExpansionCause : public AllStatic {
974 public:
975 enum Cause {
976 _no_expansion,
977 _satisfy_free_ratio,
978 _satisfy_promotion,
979 _satisfy_allocation,
980 _allocate_par_lab,
981 _allocate_par_spooling_space,
982 _adaptive_size_policy
983 };
984 // Return a string describing the cause of the expansion.
985 static const char* to_string(CMSExpansionCause::Cause cause);
986 };
988 class ConcurrentMarkSweepGeneration: public CardGeneration {
989 friend class VMStructs;
990 friend class ConcurrentMarkSweepThread;
991 friend class ConcurrentMarkSweep;
992 friend class CMSCollector;
993 protected:
994 static CMSCollector* _collector; // the collector that collects us
995 CompactibleFreeListSpace* _cmsSpace; // underlying space (only one for now)
997 // Performance Counters
998 GenerationCounters* _gen_counters;
999 GSpaceCounters* _space_counters;
1001 // Words directly allocated, used by CMSStats.
1002 size_t _direct_allocated_words;
1004 // Non-product stat counters
1005 NOT_PRODUCT(
1006 int _numObjectsPromoted;
1007 int _numWordsPromoted;
1008 int _numObjectsAllocated;
1009 int _numWordsAllocated;
1010 )
1012 // Used for sizing decisions
1013 bool _incremental_collection_failed;
1014 bool incremental_collection_failed() {
1015 return _incremental_collection_failed;
1016 }
1017 void set_incremental_collection_failed() {
1018 _incremental_collection_failed = true;
1019 }
1020 void clear_incremental_collection_failed() {
1021 _incremental_collection_failed = false;
1022 }
1024 // accessors
1025 void set_expansion_cause(CMSExpansionCause::Cause v) { _expansion_cause = v;}
1026 CMSExpansionCause::Cause expansion_cause() const { return _expansion_cause; }
1028 private:
1029 // For parallel young-gen GC support.
1030 CMSParGCThreadState** _par_gc_thread_states;
1032 // Reason generation was expanded
1033 CMSExpansionCause::Cause _expansion_cause;
1035 // In support of MinChunkSize being larger than min object size
1036 const double _dilatation_factor;
1038 enum CollectionTypes {
1039 Concurrent_collection_type = 0,
1040 MS_foreground_collection_type = 1,
1041 MSC_foreground_collection_type = 2,
1042 Unknown_collection_type = 3
1043 };
1045 CollectionTypes _debug_collection_type;
1047 // Fraction of current occupancy at which to start a CMS collection which
1048 // will collect this generation (at least).
1049 double _initiating_occupancy;
1051 protected:
1052 // Shrink generation by specified size (returns false if unable to shrink)
1053 virtual void shrink_by(size_t bytes);
1055 // Update statistics for GC
1056 virtual void update_gc_stats(int level, bool full);
1058 // Maximum available space in the generation (including uncommitted)
1059 // space.
1060 size_t max_available() const;
1062 // getter and initializer for _initiating_occupancy field.
1063 double initiating_occupancy() const { return _initiating_occupancy; }
1064 void init_initiating_occupancy(intx io, intx tr);
1066 public:
1067 ConcurrentMarkSweepGeneration(ReservedSpace rs, size_t initial_byte_size,
1068 int level, CardTableRS* ct,
1069 bool use_adaptive_freelists,
1070 FreeBlockDictionary::DictionaryChoice);
1072 // Accessors
1073 CMSCollector* collector() const { return _collector; }
1074 static void set_collector(CMSCollector* collector) {
1075 assert(_collector == NULL, "already set");
1076 _collector = collector;
1077 }
1078 CompactibleFreeListSpace* cmsSpace() const { return _cmsSpace; }
1080 Mutex* freelistLock() const;
1082 virtual Generation::Name kind() { return Generation::ConcurrentMarkSweep; }
1084 // Adaptive size policy
1085 CMSAdaptiveSizePolicy* size_policy();
1087 bool refs_discovery_is_atomic() const { return false; }
1088 bool refs_discovery_is_mt() const {
1089 // Note: CMS does MT-discovery during the parallel-remark
1090 // phases. Use ReferenceProcessorMTMutator to make refs
1091 // discovery MT-safe during such phases or other parallel
1092 // discovery phases in the future. This may all go away
1093 // if/when we decide that refs discovery is sufficiently
1094 // rare that the cost of the CAS's involved is in the
1095 // noise. That's a measurement that should be done, and
1096 // the code simplified if that turns out to be the case.
1097 return false;
1098 }
1100 // Override
1101 virtual void ref_processor_init();
1103 // Grow generation by specified size (returns false if unable to grow)
1104 bool grow_by(size_t bytes);
1105 // Grow generation to reserved size.
1106 bool grow_to_reserved();
1108 void clear_expansion_cause() { _expansion_cause = CMSExpansionCause::_no_expansion; }
1110 // Space enquiries
1111 size_t capacity() const;
1112 size_t used() const;
1113 size_t free() const;
1114 double occupancy() const { return ((double)used())/((double)capacity()); }
1115 size_t contiguous_available() const;
1116 size_t unsafe_max_alloc_nogc() const;
1118 // over-rides
1119 MemRegion used_region() const;
1120 MemRegion used_region_at_save_marks() const;
1122 // Does a "full" (forced) collection invoked on this generation collect
1123 // all younger generations as well? Note that the second conjunct is a
1124 // hack to allow the collection of the younger gen first if the flag is
1125 // set. This is better than using th policy's should_collect_gen0_first()
1126 // since that causes us to do an extra unnecessary pair of restart-&-stop-world.
1127 virtual bool full_collects_younger_generations() const {
1128 return UseCMSCompactAtFullCollection && !CollectGen0First;
1129 }
1131 void space_iterate(SpaceClosure* blk, bool usedOnly = false);
1133 // Support for compaction
1134 CompactibleSpace* first_compaction_space() const;
1135 // Adjust quantites in the generation affected by
1136 // the compaction.
1137 void reset_after_compaction();
1139 // Allocation support
1140 HeapWord* allocate(size_t size, bool tlab);
1141 HeapWord* have_lock_and_allocate(size_t size, bool tlab);
1142 oop promote(oop obj, size_t obj_size);
1143 HeapWord* par_allocate(size_t size, bool tlab) {
1144 return allocate(size, tlab);
1145 }
1147 // Incremental mode triggering.
1148 HeapWord* allocation_limit_reached(Space* space, HeapWord* top,
1149 size_t word_size);
1151 // Used by CMSStats to track direct allocation. The value is sampled and
1152 // reset after each young gen collection.
1153 size_t direct_allocated_words() const { return _direct_allocated_words; }
1154 void reset_direct_allocated_words() { _direct_allocated_words = 0; }
1156 // Overrides for parallel promotion.
1157 virtual oop par_promote(int thread_num,
1158 oop obj, markOop m, size_t word_sz);
1159 // This one should not be called for CMS.
1160 virtual void par_promote_alloc_undo(int thread_num,
1161 HeapWord* obj, size_t word_sz);
1162 virtual void par_promote_alloc_done(int thread_num);
1163 virtual void par_oop_since_save_marks_iterate_done(int thread_num);
1165 virtual bool promotion_attempt_is_safe(size_t promotion_in_bytes,
1166 bool younger_handles_promotion_failure) const;
1168 bool should_collect(bool full, size_t size, bool tlab);
1169 virtual bool should_concurrent_collect() const;
1170 virtual bool is_too_full() const;
1171 void collect(bool full,
1172 bool clear_all_soft_refs,
1173 size_t size,
1174 bool tlab);
1176 HeapWord* expand_and_allocate(size_t word_size,
1177 bool tlab,
1178 bool parallel = false);
1180 // GC prologue and epilogue
1181 void gc_prologue(bool full);
1182 void gc_prologue_work(bool full, bool registerClosure,
1183 ModUnionClosure* modUnionClosure);
1184 void gc_epilogue(bool full);
1185 void gc_epilogue_work(bool full);
1187 // Time since last GC of this generation
1188 jlong time_of_last_gc(jlong now) {
1189 return collector()->time_of_last_gc(now);
1190 }
1191 void update_time_of_last_gc(jlong now) {
1192 collector()-> update_time_of_last_gc(now);
1193 }
1195 // Allocation failure
1196 void expand(size_t bytes, size_t expand_bytes,
1197 CMSExpansionCause::Cause cause);
1198 virtual bool expand(size_t bytes, size_t expand_bytes);
1199 void shrink(size_t bytes);
1200 HeapWord* expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz);
1201 bool expand_and_ensure_spooling_space(PromotionInfo* promo);
1203 // Iteration support and related enquiries
1204 void save_marks();
1205 bool no_allocs_since_save_marks();
1206 void object_iterate_since_last_GC(ObjectClosure* cl);
1207 void younger_refs_iterate(OopsInGenClosure* cl);
1209 // Iteration support specific to CMS generations
1210 void save_sweep_limit();
1212 // More iteration support
1213 virtual void oop_iterate(MemRegion mr, OopClosure* cl);
1214 virtual void oop_iterate(OopClosure* cl);
1215 virtual void object_iterate(ObjectClosure* cl);
1217 // Need to declare the full complement of closures, whether we'll
1218 // override them or not, or get message from the compiler:
1219 // oop_since_save_marks_iterate_nv hides virtual function...
1220 #define CMS_SINCE_SAVE_MARKS_DECL(OopClosureType, nv_suffix) \
1221 void oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl);
1222 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DECL)
1224 // Smart allocation XXX -- move to CFLSpace?
1225 void setNearLargestChunk();
1226 bool isNearLargestChunk(HeapWord* addr);
1228 // Get the chunk at the end of the space. Delagates to
1229 // the space.
1230 FreeChunk* find_chunk_at_end();
1232 // Overriding of unused functionality (sharing not yet supported with CMS)
1233 void pre_adjust_pointers();
1234 void post_compact();
1236 // Debugging
1237 void prepare_for_verify();
1238 void verify(bool allow_dirty);
1239 void print_statistics() PRODUCT_RETURN;
1241 // Performance Counters support
1242 virtual void update_counters();
1243 virtual void update_counters(size_t used);
1244 void initialize_performance_counters();
1245 CollectorCounters* counters() { return collector()->counters(); }
1247 // Support for parallel remark of survivor space
1248 void* get_data_recorder(int thr_num) {
1249 //Delegate to collector
1250 return collector()->get_data_recorder(thr_num);
1251 }
1253 // Printing
1254 const char* name() const;
1255 virtual const char* short_name() const { return "CMS"; }
1256 void print() const;
1257 void printOccupancy(const char* s);
1258 bool must_be_youngest() const { return false; }
1259 bool must_be_oldest() const { return true; }
1261 void compute_new_size();
1263 CollectionTypes debug_collection_type() { return _debug_collection_type; }
1264 void rotate_debug_collection_type();
1265 };
1267 class ASConcurrentMarkSweepGeneration : public ConcurrentMarkSweepGeneration {
1269 // Return the size policy from the heap's collector
1270 // policy casted to CMSAdaptiveSizePolicy*.
1271 CMSAdaptiveSizePolicy* cms_size_policy() const;
1273 // Resize the generation based on the adaptive size
1274 // policy.
1275 void resize(size_t cur_promo, size_t desired_promo);
1277 // Return the GC counters from the collector policy
1278 CMSGCAdaptivePolicyCounters* gc_adaptive_policy_counters();
1280 virtual void shrink_by(size_t bytes);
1282 public:
1283 virtual void compute_new_size();
1284 ASConcurrentMarkSweepGeneration(ReservedSpace rs, size_t initial_byte_size,
1285 int level, CardTableRS* ct,
1286 bool use_adaptive_freelists,
1287 FreeBlockDictionary::DictionaryChoice
1288 dictionaryChoice) :
1289 ConcurrentMarkSweepGeneration(rs, initial_byte_size, level, ct,
1290 use_adaptive_freelists, dictionaryChoice) {}
1292 virtual const char* short_name() const { return "ASCMS"; }
1293 virtual Generation::Name kind() { return Generation::ASConcurrentMarkSweep; }
1295 virtual void update_counters();
1296 virtual void update_counters(size_t used);
1297 };
1299 //
1300 // Closures of various sorts used by CMS to accomplish its work
1301 //
1303 // This closure is used to check that a certain set of oops is empty.
1304 class FalseClosure: public OopClosure {
1305 public:
1306 void do_oop(oop* p) { guarantee(false, "Should be an empty set"); }
1307 void do_oop(narrowOop* p) { guarantee(false, "Should be an empty set"); }
1308 };
1310 // This closure is used to do concurrent marking from the roots
1311 // following the first checkpoint.
1312 class MarkFromRootsClosure: public BitMapClosure {
1313 CMSCollector* _collector;
1314 MemRegion _span;
1315 CMSBitMap* _bitMap;
1316 CMSBitMap* _mut;
1317 CMSMarkStack* _markStack;
1318 CMSMarkStack* _revisitStack;
1319 bool _yield;
1320 int _skipBits;
1321 HeapWord* _finger;
1322 HeapWord* _threshold;
1323 DEBUG_ONLY(bool _verifying;)
1325 public:
1326 MarkFromRootsClosure(CMSCollector* collector, MemRegion span,
1327 CMSBitMap* bitMap,
1328 CMSMarkStack* markStack,
1329 CMSMarkStack* revisitStack,
1330 bool should_yield, bool verifying = false);
1331 bool do_bit(size_t offset);
1332 void reset(HeapWord* addr);
1333 inline void do_yield_check();
1335 private:
1336 void scanOopsInOop(HeapWord* ptr);
1337 void do_yield_work();
1338 };
1340 // This closure is used to do concurrent multi-threaded
1341 // marking from the roots following the first checkpoint.
1342 // XXX This should really be a subclass of The serial version
1343 // above, but i have not had the time to refactor things cleanly.
1344 // That willbe done for Dolphin.
1345 class Par_MarkFromRootsClosure: public BitMapClosure {
1346 CMSCollector* _collector;
1347 MemRegion _whole_span;
1348 MemRegion _span;
1349 CMSBitMap* _bit_map;
1350 CMSBitMap* _mut;
1351 OopTaskQueue* _work_queue;
1352 CMSMarkStack* _overflow_stack;
1353 CMSMarkStack* _revisit_stack;
1354 bool _yield;
1355 int _skip_bits;
1356 HeapWord* _finger;
1357 HeapWord* _threshold;
1358 CMSConcMarkingTask* _task;
1359 public:
1360 Par_MarkFromRootsClosure(CMSConcMarkingTask* task, CMSCollector* collector,
1361 MemRegion span,
1362 CMSBitMap* bit_map,
1363 OopTaskQueue* work_queue,
1364 CMSMarkStack* overflow_stack,
1365 CMSMarkStack* revisit_stack,
1366 bool should_yield);
1367 bool do_bit(size_t offset);
1368 inline void do_yield_check();
1370 private:
1371 void scan_oops_in_oop(HeapWord* ptr);
1372 void do_yield_work();
1373 bool get_work_from_overflow_stack();
1374 };
1376 // The following closures are used to do certain kinds of verification of
1377 // CMS marking.
1378 class PushAndMarkVerifyClosure: public OopClosure {
1379 CMSCollector* _collector;
1380 MemRegion _span;
1381 CMSBitMap* _verification_bm;
1382 CMSBitMap* _cms_bm;
1383 CMSMarkStack* _mark_stack;
1384 protected:
1385 void do_oop(oop p);
1386 template <class T> inline void do_oop_work(T *p) {
1387 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
1388 do_oop(obj);
1389 }
1390 public:
1391 PushAndMarkVerifyClosure(CMSCollector* cms_collector,
1392 MemRegion span,
1393 CMSBitMap* verification_bm,
1394 CMSBitMap* cms_bm,
1395 CMSMarkStack* mark_stack);
1396 void do_oop(oop* p);
1397 void do_oop(narrowOop* p);
1398 // Deal with a stack overflow condition
1399 void handle_stack_overflow(HeapWord* lost);
1400 };
1402 class MarkFromRootsVerifyClosure: public BitMapClosure {
1403 CMSCollector* _collector;
1404 MemRegion _span;
1405 CMSBitMap* _verification_bm;
1406 CMSBitMap* _cms_bm;
1407 CMSMarkStack* _mark_stack;
1408 HeapWord* _finger;
1409 PushAndMarkVerifyClosure _pam_verify_closure;
1410 public:
1411 MarkFromRootsVerifyClosure(CMSCollector* collector, MemRegion span,
1412 CMSBitMap* verification_bm,
1413 CMSBitMap* cms_bm,
1414 CMSMarkStack* mark_stack);
1415 bool do_bit(size_t offset);
1416 void reset(HeapWord* addr);
1417 };
1420 // This closure is used to check that a certain set of bits is
1421 // "empty" (i.e. the bit vector doesn't have any 1-bits).
1422 class FalseBitMapClosure: public BitMapClosure {
1423 public:
1424 bool do_bit(size_t offset) {
1425 guarantee(false, "Should not have a 1 bit");
1426 return true;
1427 }
1428 };
1430 // This closure is used during the second checkpointing phase
1431 // to rescan the marked objects on the dirty cards in the mod
1432 // union table and the card table proper. It's invoked via
1433 // MarkFromDirtyCardsClosure below. It uses either
1434 // [Par_]MarkRefsIntoAndScanClosure (Par_ in the parallel case)
1435 // declared in genOopClosures.hpp to accomplish some of its work.
1436 // In the parallel case the bitMap is shared, so access to
1437 // it needs to be suitably synchronized for updates by embedded
1438 // closures that update it; however, this closure itself only
1439 // reads the bit_map and because it is idempotent, is immune to
1440 // reading stale values.
1441 class ScanMarkedObjectsAgainClosure: public UpwardsObjectClosure {
1442 #ifdef ASSERT
1443 CMSCollector* _collector;
1444 MemRegion _span;
1445 union {
1446 CMSMarkStack* _mark_stack;
1447 OopTaskQueue* _work_queue;
1448 };
1449 #endif // ASSERT
1450 bool _parallel;
1451 CMSBitMap* _bit_map;
1452 union {
1453 MarkRefsIntoAndScanClosure* _scan_closure;
1454 Par_MarkRefsIntoAndScanClosure* _par_scan_closure;
1455 };
1457 public:
1458 ScanMarkedObjectsAgainClosure(CMSCollector* collector,
1459 MemRegion span,
1460 ReferenceProcessor* rp,
1461 CMSBitMap* bit_map,
1462 CMSMarkStack* mark_stack,
1463 CMSMarkStack* revisit_stack,
1464 MarkRefsIntoAndScanClosure* cl):
1465 #ifdef ASSERT
1466 _collector(collector),
1467 _span(span),
1468 _mark_stack(mark_stack),
1469 #endif // ASSERT
1470 _parallel(false),
1471 _bit_map(bit_map),
1472 _scan_closure(cl) { }
1474 ScanMarkedObjectsAgainClosure(CMSCollector* collector,
1475 MemRegion span,
1476 ReferenceProcessor* rp,
1477 CMSBitMap* bit_map,
1478 OopTaskQueue* work_queue,
1479 CMSMarkStack* revisit_stack,
1480 Par_MarkRefsIntoAndScanClosure* cl):
1481 #ifdef ASSERT
1482 _collector(collector),
1483 _span(span),
1484 _work_queue(work_queue),
1485 #endif // ASSERT
1486 _parallel(true),
1487 _bit_map(bit_map),
1488 _par_scan_closure(cl) { }
1490 void do_object(oop obj) {
1491 guarantee(false, "Call do_object_b(oop, MemRegion) instead");
1492 }
1493 bool do_object_b(oop obj) {
1494 guarantee(false, "Call do_object_b(oop, MemRegion) form instead");
1495 return false;
1496 }
1497 bool do_object_bm(oop p, MemRegion mr);
1498 };
1500 // This closure is used during the second checkpointing phase
1501 // to rescan the marked objects on the dirty cards in the mod
1502 // union table and the card table proper. It invokes
1503 // ScanMarkedObjectsAgainClosure above to accomplish much of its work.
1504 // In the parallel case, the bit map is shared and requires
1505 // synchronized access.
1506 class MarkFromDirtyCardsClosure: public MemRegionClosure {
1507 CompactibleFreeListSpace* _space;
1508 ScanMarkedObjectsAgainClosure _scan_cl;
1509 size_t _num_dirty_cards;
1511 public:
1512 MarkFromDirtyCardsClosure(CMSCollector* collector,
1513 MemRegion span,
1514 CompactibleFreeListSpace* space,
1515 CMSBitMap* bit_map,
1516 CMSMarkStack* mark_stack,
1517 CMSMarkStack* revisit_stack,
1518 MarkRefsIntoAndScanClosure* cl):
1519 _space(space),
1520 _num_dirty_cards(0),
1521 _scan_cl(collector, span, collector->ref_processor(), bit_map,
1522 mark_stack, revisit_stack, cl) { }
1524 MarkFromDirtyCardsClosure(CMSCollector* collector,
1525 MemRegion span,
1526 CompactibleFreeListSpace* space,
1527 CMSBitMap* bit_map,
1528 OopTaskQueue* work_queue,
1529 CMSMarkStack* revisit_stack,
1530 Par_MarkRefsIntoAndScanClosure* cl):
1531 _space(space),
1532 _num_dirty_cards(0),
1533 _scan_cl(collector, span, collector->ref_processor(), bit_map,
1534 work_queue, revisit_stack, cl) { }
1536 void do_MemRegion(MemRegion mr);
1537 void set_space(CompactibleFreeListSpace* space) { _space = space; }
1538 size_t num_dirty_cards() { return _num_dirty_cards; }
1539 };
1541 // This closure is used in the non-product build to check
1542 // that there are no MemRegions with a certain property.
1543 class FalseMemRegionClosure: public MemRegionClosure {
1544 void do_MemRegion(MemRegion mr) {
1545 guarantee(!mr.is_empty(), "Shouldn't be empty");
1546 guarantee(false, "Should never be here");
1547 }
1548 };
1550 // This closure is used during the precleaning phase
1551 // to "carefully" rescan marked objects on dirty cards.
1552 // It uses MarkRefsIntoAndScanClosure declared in genOopClosures.hpp
1553 // to accomplish some of its work.
1554 class ScanMarkedObjectsAgainCarefullyClosure: public ObjectClosureCareful {
1555 CMSCollector* _collector;
1556 MemRegion _span;
1557 bool _yield;
1558 Mutex* _freelistLock;
1559 CMSBitMap* _bitMap;
1560 CMSMarkStack* _markStack;
1561 MarkRefsIntoAndScanClosure* _scanningClosure;
1563 public:
1564 ScanMarkedObjectsAgainCarefullyClosure(CMSCollector* collector,
1565 MemRegion span,
1566 CMSBitMap* bitMap,
1567 CMSMarkStack* markStack,
1568 CMSMarkStack* revisitStack,
1569 MarkRefsIntoAndScanClosure* cl,
1570 bool should_yield):
1571 _collector(collector),
1572 _span(span),
1573 _yield(should_yield),
1574 _bitMap(bitMap),
1575 _markStack(markStack),
1576 _scanningClosure(cl) {
1577 }
1579 void do_object(oop p) {
1580 guarantee(false, "call do_object_careful instead");
1581 }
1583 size_t do_object_careful(oop p) {
1584 guarantee(false, "Unexpected caller");
1585 return 0;
1586 }
1588 size_t do_object_careful_m(oop p, MemRegion mr);
1590 void setFreelistLock(Mutex* m) {
1591 _freelistLock = m;
1592 _scanningClosure->set_freelistLock(m);
1593 }
1595 private:
1596 inline bool do_yield_check();
1598 void do_yield_work();
1599 };
1601 class SurvivorSpacePrecleanClosure: public ObjectClosureCareful {
1602 CMSCollector* _collector;
1603 MemRegion _span;
1604 bool _yield;
1605 CMSBitMap* _bit_map;
1606 CMSMarkStack* _mark_stack;
1607 PushAndMarkClosure* _scanning_closure;
1608 unsigned int _before_count;
1610 public:
1611 SurvivorSpacePrecleanClosure(CMSCollector* collector,
1612 MemRegion span,
1613 CMSBitMap* bit_map,
1614 CMSMarkStack* mark_stack,
1615 PushAndMarkClosure* cl,
1616 unsigned int before_count,
1617 bool should_yield):
1618 _collector(collector),
1619 _span(span),
1620 _yield(should_yield),
1621 _bit_map(bit_map),
1622 _mark_stack(mark_stack),
1623 _scanning_closure(cl),
1624 _before_count(before_count)
1625 { }
1627 void do_object(oop p) {
1628 guarantee(false, "call do_object_careful instead");
1629 }
1631 size_t do_object_careful(oop p);
1633 size_t do_object_careful_m(oop p, MemRegion mr) {
1634 guarantee(false, "Unexpected caller");
1635 return 0;
1636 }
1638 private:
1639 inline void do_yield_check();
1640 void do_yield_work();
1641 };
1643 // This closure is used to accomplish the sweeping work
1644 // after the second checkpoint but before the concurrent reset
1645 // phase.
1646 //
1647 // Terminology
1648 // left hand chunk (LHC) - block of one or more chunks currently being
1649 // coalesced. The LHC is available for coalescing with a new chunk.
1650 // right hand chunk (RHC) - block that is currently being swept that is
1651 // free or garbage that can be coalesced with the LHC.
1652 // _inFreeRange is true if there is currently a LHC
1653 // _lastFreeRangeCoalesced is true if the LHC consists of more than one chunk.
1654 // _freeRangeInFreeLists is true if the LHC is in the free lists.
1655 // _freeFinger is the address of the current LHC
1656 class SweepClosure: public BlkClosureCareful {
1657 CMSCollector* _collector; // collector doing the work
1658 ConcurrentMarkSweepGeneration* _g; // Generation being swept
1659 CompactibleFreeListSpace* _sp; // Space being swept
1660 HeapWord* _limit;
1661 Mutex* _freelistLock; // Free list lock (in space)
1662 CMSBitMap* _bitMap; // Marking bit map (in
1663 // generation)
1664 bool _inFreeRange; // Indicates if we are in the
1665 // midst of a free run
1666 bool _freeRangeInFreeLists;
1667 // Often, we have just found
1668 // a free chunk and started
1669 // a new free range; we do not
1670 // eagerly remove this chunk from
1671 // the free lists unless there is
1672 // a possibility of coalescing.
1673 // When true, this flag indicates
1674 // that the _freeFinger below
1675 // points to a potentially free chunk
1676 // that may still be in the free lists
1677 bool _lastFreeRangeCoalesced;
1678 // free range contains chunks
1679 // coalesced
1680 bool _yield;
1681 // Whether sweeping should be
1682 // done with yields. For instance
1683 // when done by the foreground
1684 // collector we shouldn't yield.
1685 HeapWord* _freeFinger; // When _inFreeRange is set, the
1686 // pointer to the "left hand
1687 // chunk"
1688 size_t _freeRangeSize;
1689 // When _inFreeRange is set, this
1690 // indicates the accumulated size
1691 // of the "left hand chunk"
1692 NOT_PRODUCT(
1693 size_t _numObjectsFreed;
1694 size_t _numWordsFreed;
1695 size_t _numObjectsLive;
1696 size_t _numWordsLive;
1697 size_t _numObjectsAlreadyFree;
1698 size_t _numWordsAlreadyFree;
1699 FreeChunk* _last_fc;
1700 )
1701 private:
1702 // Code that is common to a free chunk or garbage when
1703 // encountered during sweeping.
1704 void doPostIsFreeOrGarbageChunk(FreeChunk *fc,
1705 size_t chunkSize);
1706 // Process a free chunk during sweeping.
1707 void doAlreadyFreeChunk(FreeChunk *fc);
1708 // Process a garbage chunk during sweeping.
1709 size_t doGarbageChunk(FreeChunk *fc);
1710 // Process a live chunk during sweeping.
1711 size_t doLiveChunk(FreeChunk* fc);
1713 // Accessors.
1714 HeapWord* freeFinger() const { return _freeFinger; }
1715 void set_freeFinger(HeapWord* v) { _freeFinger = v; }
1716 size_t freeRangeSize() const { return _freeRangeSize; }
1717 void set_freeRangeSize(size_t v) { _freeRangeSize = v; }
1718 bool inFreeRange() const { return _inFreeRange; }
1719 void set_inFreeRange(bool v) { _inFreeRange = v; }
1720 bool lastFreeRangeCoalesced() const { return _lastFreeRangeCoalesced; }
1721 void set_lastFreeRangeCoalesced(bool v) { _lastFreeRangeCoalesced = v; }
1722 bool freeRangeInFreeLists() const { return _freeRangeInFreeLists; }
1723 void set_freeRangeInFreeLists(bool v) { _freeRangeInFreeLists = v; }
1725 // Initialize a free range.
1726 void initialize_free_range(HeapWord* freeFinger, bool freeRangeInFreeLists);
1727 // Return this chunk to the free lists.
1728 void flushCurFreeChunk(HeapWord* chunk, size_t size);
1730 // Check if we should yield and do so when necessary.
1731 inline void do_yield_check(HeapWord* addr);
1733 // Yield
1734 void do_yield_work(HeapWord* addr);
1736 // Debugging/Printing
1737 void record_free_block_coalesced(FreeChunk* fc) const PRODUCT_RETURN;
1739 public:
1740 SweepClosure(CMSCollector* collector, ConcurrentMarkSweepGeneration* g,
1741 CMSBitMap* bitMap, bool should_yield);
1742 ~SweepClosure();
1744 size_t do_blk_careful(HeapWord* addr);
1745 };
1747 // Closures related to weak references processing
1749 // During CMS' weak reference processing, this is a
1750 // work-routine/closure used to complete transitive
1751 // marking of objects as live after a certain point
1752 // in which an initial set has been completely accumulated.
1753 // This closure is currently used both during the final
1754 // remark stop-world phase, as well as during the concurrent
1755 // precleaning of the discovered reference lists.
1756 class CMSDrainMarkingStackClosure: public VoidClosure {
1757 CMSCollector* _collector;
1758 MemRegion _span;
1759 CMSMarkStack* _mark_stack;
1760 CMSBitMap* _bit_map;
1761 CMSKeepAliveClosure* _keep_alive;
1762 bool _concurrent_precleaning;
1763 public:
1764 CMSDrainMarkingStackClosure(CMSCollector* collector, MemRegion span,
1765 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
1766 CMSKeepAliveClosure* keep_alive,
1767 bool cpc):
1768 _collector(collector),
1769 _span(span),
1770 _bit_map(bit_map),
1771 _mark_stack(mark_stack),
1772 _keep_alive(keep_alive),
1773 _concurrent_precleaning(cpc) {
1774 assert(_concurrent_precleaning == _keep_alive->concurrent_precleaning(),
1775 "Mismatch");
1776 }
1778 void do_void();
1779 };
1781 // A parallel version of CMSDrainMarkingStackClosure above.
1782 class CMSParDrainMarkingStackClosure: public VoidClosure {
1783 CMSCollector* _collector;
1784 MemRegion _span;
1785 OopTaskQueue* _work_queue;
1786 CMSBitMap* _bit_map;
1787 CMSInnerParMarkAndPushClosure _mark_and_push;
1789 public:
1790 CMSParDrainMarkingStackClosure(CMSCollector* collector,
1791 MemRegion span, CMSBitMap* bit_map,
1792 OopTaskQueue* work_queue):
1793 _collector(collector),
1794 _span(span),
1795 _bit_map(bit_map),
1796 _work_queue(work_queue),
1797 _mark_and_push(collector, span, bit_map, work_queue) { }
1799 public:
1800 void trim_queue(uint max);
1801 void do_void();
1802 };
1804 // Allow yielding or short-circuiting of reference list
1805 // prelceaning work.
1806 class CMSPrecleanRefsYieldClosure: public YieldClosure {
1807 CMSCollector* _collector;
1808 void do_yield_work();
1809 public:
1810 CMSPrecleanRefsYieldClosure(CMSCollector* collector):
1811 _collector(collector) {}
1812 virtual bool should_return();
1813 };
1816 // Convenience class that locks free list locks for given CMS collector
1817 class FreelistLocker: public StackObj {
1818 private:
1819 CMSCollector* _collector;
1820 public:
1821 FreelistLocker(CMSCollector* collector):
1822 _collector(collector) {
1823 _collector->getFreelistLocks();
1824 }
1826 ~FreelistLocker() {
1827 _collector->releaseFreelistLocks();
1828 }
1829 };
1831 // Mark all dead objects in a given space.
1832 class MarkDeadObjectsClosure: public BlkClosure {
1833 const CMSCollector* _collector;
1834 const CompactibleFreeListSpace* _sp;
1835 CMSBitMap* _live_bit_map;
1836 CMSBitMap* _dead_bit_map;
1837 public:
1838 MarkDeadObjectsClosure(const CMSCollector* collector,
1839 const CompactibleFreeListSpace* sp,
1840 CMSBitMap *live_bit_map,
1841 CMSBitMap *dead_bit_map) :
1842 _collector(collector),
1843 _sp(sp),
1844 _live_bit_map(live_bit_map),
1845 _dead_bit_map(dead_bit_map) {}
1846 size_t do_blk(HeapWord* addr);
1847 };