8196341: Add JFR events for parallel phases of G1

0 /*
1  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
2  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
3  *
4  * This code is free software; you can redistribute it and/or modify it
5  * under the terms of the GNU General Public License version 2 only, as
6  * published by the Free Software Foundation.
7  *
8  * This code is distributed in the hope that it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
11  * version 2 for more details (a copy is included in the LICENSE file that
12  * accompanied this code).
13  *
14  * You should have received a copy of the GNU General Public License version
15  * 2 along with this work; if not, write to the Free Software Foundation,
16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17  *
18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19  * or visit www.oracle.com if you need additional information or have any
20  * questions.
21  *
22  */
23 
24 #include "precompiled.hpp"
25 #include "gc/g1/dirtyCardQueue.hpp"
26 #include "gc/g1/g1BarrierSet.hpp"
27 #include "gc/g1/g1BlockOffsetTable.inline.hpp"
28 #include "gc/g1/g1CardTable.inline.hpp"
29 #include "gc/g1/g1CollectedHeap.inline.hpp"
30 #include "gc/g1/g1ConcurrentRefine.hpp"
31 #include "gc/g1/g1FromCardCache.hpp"
32 #include "gc/g1/g1GCPhaseTimes.hpp"
33 #include "gc/g1/g1HotCardCache.hpp"
34 #include "gc/g1/g1OopClosures.inline.hpp"
35 #include "gc/g1/g1RootClosures.hpp"
36 #include "gc/g1/g1RemSet.hpp"
37 #include "gc/g1/heapRegion.inline.hpp"
38 #include "gc/g1/heapRegionManager.inline.hpp"
39 #include "gc/g1/heapRegionRemSet.hpp"
40 #include "gc/shared/gcTraceTime.inline.hpp"
41 #include "gc/shared/suspendibleThreadSet.hpp"
42 #include "jfr/jfrEvents.hpp"
43 #include "memory/iterator.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "oops/access.inline.hpp"
46 #include "oops/oop.inline.hpp"
47 #include "runtime/os.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/globalDefinitions.hpp"
50 #include "utilities/intHisto.hpp"
51 #include "utilities/stack.inline.hpp"
52 #include "utilities/ticks.hpp"
53 
54 // Collects information about the overall remembered set scan progress during an evacuation.
55 class G1RemSetScanState : public CHeapObj<mtGC> {
56 private:
57   class G1ClearCardTableTask : public AbstractGangTask {
58     G1CollectedHeap* _g1h;
59     uint* _dirty_region_list;
60     size_t _num_dirty_regions;
61     size_t _chunk_length;
62 
63     size_t volatile _cur_dirty_regions;
64   public:
65     G1ClearCardTableTask(G1CollectedHeap* g1h,
66                          uint* dirty_region_list,
67                          size_t num_dirty_regions,
68                          size_t chunk_length) :
69       AbstractGangTask("G1 Clear Card Table Task"),
70       _g1h(g1h),
71       _dirty_region_list(dirty_region_list),
72       _num_dirty_regions(num_dirty_regions),
73       _chunk_length(chunk_length),
74       _cur_dirty_regions(0) {
75 
76       assert(chunk_length > 0, "must be");
77     }
78 
79     static size_t chunk_size() { return M; }
80 
81     void work(uint worker_id) {
82       while (_cur_dirty_regions < _num_dirty_regions) {
83         size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
84         size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
85 
86         for (size_t i = next; i < max; i++) {
87           HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
88           if (!r->is_survivor()) {
89             r->clear_cardtable();
90           }
91         }
92       }
93     }
94   };
95 
96   size_t _max_regions;
97 
98   // Scan progress for the remembered set of a single region. Transitions from
99   // Unclaimed -> Claimed -> Complete.
100   // At each of the transitions the thread that does the transition needs to perform
101   // some special action once. This is the reason for the extra "Claimed" state.
102   typedef jint G1RemsetIterState;
103 
104   static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
105   static const G1RemsetIterState Claimed = 1;   // The remembered set is currently being scanned.
106   static const G1RemsetIterState Complete = 2;  // The remembered set has been completely scanned.
107 
108   G1RemsetIterState volatile* _iter_states;
109   // The current location where the next thread should continue scanning in a region's
110   // remembered set.
111   size_t volatile* _iter_claims;
112 
113   // Temporary buffer holding the regions we used to store remembered set scan duplicate
114   // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
115   uint* _dirty_region_buffer;
116 
117   typedef jbyte IsDirtyRegionState;
118   static const IsDirtyRegionState Clean = 0;
119   static const IsDirtyRegionState Dirty = 1;
120   // Holds a flag for every region whether it is in the _dirty_region_buffer already
121   // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools.
122   IsDirtyRegionState* _in_dirty_region_buffer;
123   size_t _cur_dirty_region;
124 
125   // Creates a snapshot of the current _top values at the start of collection to
126   // filter out card marks that we do not want to scan.
127   class G1ResetScanTopClosure : public HeapRegionClosure {
128   private:
129     HeapWord** _scan_top;
130   public:
131     G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
132 
133     virtual bool do_heap_region(HeapRegion* r) {
134       uint hrm_index = r->hrm_index();
135       if (!r->in_collection_set() && r->is_old_or_humongous_or_archive()) {
136         _scan_top[hrm_index] = r->top();
137       } else {
138         _scan_top[hrm_index] = r->bottom();
139       }
140       return false;
141     }
142   };
143 
144   // For each region, contains the maximum top() value to be used during this garbage
145   // collection. Subsumes common checks like filtering out everything but old and
146   // humongous regions outside the collection set.
147   // This is valid because we are not interested in scanning stray remembered set
148   // entries from free or archive regions.
149   HeapWord** _scan_top;
150 public:
151   G1RemSetScanState() :
152     _max_regions(0),
153     _iter_states(NULL),
154     _iter_claims(NULL),
155     _dirty_region_buffer(NULL),
156     _in_dirty_region_buffer(NULL),
157     _cur_dirty_region(0),
158     _scan_top(NULL) {
159   }
160 
161   ~G1RemSetScanState() {
162     if (_iter_states != NULL) {
163       FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
164     }
165     if (_iter_claims != NULL) {
166       FREE_C_HEAP_ARRAY(size_t, _iter_claims);
167     }
168     if (_dirty_region_buffer != NULL) {
169       FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
170     }
171     if (_in_dirty_region_buffer != NULL) {
172       FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
173     }
174     if (_scan_top != NULL) {
175       FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
176     }
177   }
178 
179   void initialize(uint max_regions) {
180     assert(_iter_states == NULL, "Must not be initialized twice");
181     assert(_iter_claims == NULL, "Must not be initialized twice");
182     _max_regions = max_regions;
183     _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
184     _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
185     _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
186     _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
187     _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
188   }
189 
190   void reset() {
191     for (uint i = 0; i < _max_regions; i++) {
192       _iter_states[i] = Unclaimed;
193     }
194 
195     G1ResetScanTopClosure cl(_scan_top);
196     G1CollectedHeap::heap()->heap_region_iterate(&cl);
197 
198     memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
199     memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState));
200     _cur_dirty_region = 0;
201   }
202 
203   // Attempt to claim the remembered set of the region for iteration. Returns true
204   // if this call caused the transition from Unclaimed to Claimed.
205   inline bool claim_iter(uint region) {
206     assert(region < _max_regions, "Tried to access invalid region %u", region);
207     if (_iter_states[region] != Unclaimed) {
208       return false;
209     }
210     G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
211     return (res == Unclaimed);
212   }
213 
214   // Try to atomically sets the iteration state to "complete". Returns true for the
215   // thread that caused the transition.
216   inline bool set_iter_complete(uint region) {
217     if (iter_is_complete(region)) {
218       return false;
219     }
220     G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
221     return (res == Claimed);
222   }
223 
224   // Returns true if the region's iteration is complete.
225   inline bool iter_is_complete(uint region) const {
226     assert(region < _max_regions, "Tried to access invalid region %u", region);
227     return _iter_states[region] == Complete;
228   }
229 
230   // The current position within the remembered set of the given region.
231   inline size_t iter_claimed(uint region) const {
232     assert(region < _max_regions, "Tried to access invalid region %u", region);
233     return _iter_claims[region];
234   }
235 
236   // Claim the next block of cards within the remembered set of the region with
237   // step size.
238   inline size_t iter_claimed_next(uint region, size_t step) {
239     return Atomic::add(step, &_iter_claims[region]) - step;
240   }
241 
242   void add_dirty_region(uint region) {
243     if (_in_dirty_region_buffer[region] == Dirty) {
244       return;
245     }
246 
247     bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean;
248     if (marked_as_dirty) {
249       size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
250       _dirty_region_buffer[allocated] = region;
251     }
252   }
253 
254   HeapWord* scan_top(uint region_idx) const {
255     return _scan_top[region_idx];
256   }
257 
258   // Clear the card table of "dirty" regions.
259   void clear_card_table(WorkGang* workers) {
260     if (_cur_dirty_region == 0) {
261       return;
262     }
263 
264     size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) 
265     uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
266     size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
267 
268     // Iterate over the dirty cards region list.
269     G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
270 
271     log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
272                         "units of work for " SIZE_FORMAT " regions.",
273                         cl.name(), num_workers, num_chunks, _cur_dirty_region);
274     workers->run_task(&cl, num_workers);
275 
276 #ifndef PRODUCT
277     G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
278 #endif
279   }
280 };
281 
282 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
283                    G1CardTable* ct,
284                    G1HotCardCache* hot_card_cache) :
285   _scan_state(new G1RemSetScanState()),
286   _prev_period_summary(),
287   _g1h(g1h),
288   _num_conc_refined_cards(0),
289   _ct(ct),
290   _g1p(_g1h->g1_policy()),
291   _hot_card_cache(hot_card_cache) {
292 }
293 
294 G1RemSet::~G1RemSet() {
295   if (_scan_state != NULL) {
296     delete _scan_state;
297   }
298 }
299 
300 uint G1RemSet::num_par_rem_sets() {
301   return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
302 }
303 
304 void G1RemSet::initialize(size_t capacity, uint max_regions) {
305   G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
306   _scan_state->initialize(max_regions);
307 }
308 
309 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
310                                                    G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
311                                                    G1ParScanThreadState* pss,
312                                                    uint worker_i) :
313   _g1h(G1CollectedHeap::heap()),
314   _ct(_g1h->card_table()),
315   _pss(pss),
316   _scan_objs_on_card_cl(scan_obj_on_card),
317   _scan_state(scan_state),
318   _worker_i(worker_i),
319   _cards_scanned(0),
320   _cards_claimed(0),
321   _cards_skipped(0),
322   _rem_set_root_scan_time(),
323   _rem_set_trim_partially_time(),
324   _strong_code_root_scan_time(),
325   _strong_code_trim_partially_time() {
326 }
327 
328 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
329   _ct->set_card_claimed(card_index);
330   _scan_state->add_dirty_region(region_idx_for_card);
331 }
332 
333 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
334   HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
335   _scan_objs_on_card_cl->set_region(card_region);
336   card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
337   _scan_objs_on_card_cl->trim_queue_partially();
338   _cards_scanned++;
339 }
340 
341 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
342   EventGCPhaseParallel event;
343   uint const region_idx = r->hrm_index();
344 
345   if (_scan_state->claim_iter(region_idx)) {
346     // If we ever free the collection set concurrently, we should also
347     // clear the card table concurrently therefore we won't need to
348     // add regions of the collection set to the dirty cards region.
349     _scan_state->add_dirty_region(region_idx);
350   }
351 
352   // We claim cards in blocks so as to reduce the contention.
353   size_t const block_size = G1RSetScanBlockSize;
354 
355   HeapRegionRemSetIterator iter(r->rem_set());
356   size_t card_index;
357 
358   size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
359   for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
360     if (current_card >= claimed_card_block + block_size) {
361       claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
362     }
363     if (current_card < claimed_card_block) {
364       _cards_skipped++;
365       continue;
366     }
367     _cards_claimed++;
368 
369     // If the card is dirty, then G1 will scan it during Update RS.
370     if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
371       continue;
372     }
373 
374     HeapWord* const card_start = _g1h->bot()->address_for_index(card_index);
375     uint const region_idx_for_card = _g1h->addr_to_region(card_start);
376 
377     assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start),
378            "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)-
379     HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
380     if (card_start >= top) {
381       continue;
382     }
383 
384     // We claim lazily (so races are possible but they're benign), which reduces the
385     // number of duplicate scans (the rsets of the regions in the cset can intersect).
386     // Claim the card after checking bounds above: the remembered set may contain
387     // random cards into current survivor, and we would then have an incorrectly
388     // claimed card in survivor space. Card table clear does not reset the card table
389     // of survivor space regions.
390     claim_card(card_index, region_idx_for_card);
391 
392     MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
393 
394     scan_card(mr, region_idx_for_card);
395   }
396   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::ScanRS));
397 }
398 
399 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
400   EventGCPhaseParallel event;
401   r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
402   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
403 }
404 
405 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
406   assert(r->in_collection_set(),
407          "Should only be called on elements of the collection set but region %u is not.",
408          r->hrm_index());
409   uint const region_idx = r->hrm_index();
410 
411   // Do an early out if we know we are complete.
412   if (_scan_state->iter_is_complete(region_idx)) {
413     return false;
414   }
415 
416   {
417     G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
418     scan_rem_set_roots(r);
419   }
420 
421   if (_scan_state->set_iter_complete(region_idx)) {
422     G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
423     // Scan the strong code root list attached to the current region
424     scan_strong_code_roots(r);
425   }
426   return false;
427 }
428 
429 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
430   G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
431   G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i);
432   _g1h->collection_set_iterate_from(&cl, worker_i);
433 
434   G1GCPhaseTimes* p = _g1p->phase_times();
435 
436   p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
437   p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
438 
439   p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
440   p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
441   p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
442 
443   p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
444   p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
445 }
446 
447 // Closure used for updating rem sets. Only called during an evacuation pause.
448 class G1RefineCardClosure: public CardTableEntryClosure {
449   G1RemSet* _g1rs;
450   G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
451 
452   size_t _cards_scanned;
453   size_t _cards_skipped;
454 public:
455   G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
456     _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
457   {}
458 
459   bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
460     // The only time we care about recording cards that
461     // contain references that point into the collection set
462     // is during RSet updating within an evacuation pause.
463     // In this case worker_i should be the id of a GC worker thread.
464     assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
465 
466     bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
467 
468     if (card_scanned) {
469       _update_rs_cl->trim_queue_partially();
470       _cards_scanned++;
471     } else {
472       _cards_skipped++;
473     }
474     return true;
475   }
476 
477   size_t cards_scanned() const { return _cards_scanned; }
478   size_t cards_skipped() const { return _cards_skipped; }
479 };
480 
481 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
482   G1GCPhaseTimes* p = _g1p->phase_times();
483 
484   // Apply closure to log entries in the HCC.
485   if (G1HotCardCache::default_use_cache()) {
486     G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
487 
488     G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss, worker_i);
489     G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
490     _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
491   }
492 
493   // Now apply the closure to all remaining log entries.
494   {
495     G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
496 
497     G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i);
498     G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
499     _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
500 
501     p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRS
502     p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRS
503   }
504 }
505 
506 void G1RemSet::cleanupHRRS() {
507   HeapRegionRemSet::cleanup();
508 }
509 
510 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
511   update_rem_set(pss, worker_i);
512   scan_rem_set(pss, worker_i);;
513 }
514 
515 void G1RemSet::prepare_for_oops_into_collection_set_do() {
516   DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
517   dcqs.concatenate_logs();
518 
519   _scan_state->reset();
520 }
521 
522 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
523   G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times();
524 
525   // Set all cards back to clean.
526   double start = os::elapsedTime();
527   _scan_state->clear_card_table(_g1h->workers());
528   phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
529 }
530 
531 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
532 #ifdef ASSERT
533   G1CollectedHeap* g1h = G1CollectedHeap::heap();
534   assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
535          "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
536          p2i(card_ptr),
537          ct->index_for(ct->addr_for(card_ptr)),
538          p2i(ct->addr_for(card_ptr)),
539          g1h->addr_to_region(ct->addr_for(card_ptr)));
540 #endif
541 }
542 
543 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
544                                         uint worker_i) {
545   assert(!_g1h->is_gc_active(), "Only call concurrently");
546 
547   check_card_ptr(card_ptr, _ct);
548 
549   // If the card is no longer dirty, nothing to do.
550   if (*card_ptr != G1CardTable::dirty_card_val()) {
551     return;
552   }
553 
554   // Construct the region representing the card.
555   HeapWord* start = _ct->addr_for(card_ptr);
556   // And find the region containing it.
557   HeapRegion* r = _g1h->heap_region_containing(start);
558 
559   // This check is needed for some uncommon cases where we should
560   // ignore the card.
561   //
562   // The region could be young.  Cards for young regions are
563   // distinctly marked (set to g1_young_gen), so the post-barrier will
564   // filter them out.  However, that marking is performed
565   // concurrently.  A write to a young object could occur before the
566   // card has been marked young, slipping past the filter.
567   //
568   // The card could be stale, because the region has been freed since
569   // the card was recorded. In this case the region type could be
570   // anything.  If (still) free or (reallocated) young, just ignore
571   // it.  If (reallocated) old or humongous, the later card trimming
572   // and additional checks in iteration may detect staleness.  At
573   // worst, we end up processing a stale card unnecessarily.
574   //
575   // In the normal (non-stale) case, the synchronization between the
576   // enqueueing of the card and processing it here will have ensured
577   // we see the up-to-date region type here.
578   if (!r->is_old_or_humongous_or_archive()) {
579     return;
580   }
581 
582   // The result from the hot card cache insert call is either:
583   //   * pointer to the current card
584   //     (implying that the current card is not 'hot'),
585   //   * null
586   //     (meaning we had inserted the card ptr into the "hot" card cache,
587   //     which had some headroom),
588   //   * a pointer to a "hot" card that was evicted from the "hot" cache.
589   //
590 
591   if (_hot_card_cache->use_cache()) {
592     assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
593 
594     const jbyte* orig_card_ptr = card_ptr;
595     card_ptr = _hot_card_cache->insert(card_ptr);
596     if (card_ptr == NULL) {
597       // There was no eviction. Nothing to do.
598       return;
599     } else if (card_ptr != orig_card_ptr) {
600       // Original card was inserted and an old card was evicted.
601       start = _ct->addr_for(card_ptr);
602       r = _g1h->heap_region_containing(start);
603 
604       // Check whether the region formerly in the cache should be
605       // ignored, as discussed earlier for the original card.  The
606       // region could have been freed while in the cache.
607       if (!r->is_old_or_humongous_or_archive()) {
608         return;
609       }
610     } // Else we still have the original card.
611   }
612 
613   // Trim the region designated by the card to what's been allocated
614   // in the region.  The card could be stale, or the card could cover
615   // (part of) an object at the end of the allocated space and extend
616   // beyond the end of allocation.
617 
618   // Non-humongous objects are only allocated in the old-gen during
619   // GC, so if region is old then top is stable.  Humongous object
620   // allocation sets top last; if top has not yet been set, this is
621   // a stale card and we'll end up with an empty intersection.  If
622   // this is not a stale card, the synchronization between the
623   // enqueuing of the card and processing it here will have ensured
624   // we see the up-to-date top here.
625   HeapWord* scan_limit = r->top();
626 
627   if (scan_limit <= start) {
628     // If the trimmed region is empty, the card must be stale.
629     return;
630   }
631 
632   // Okay to clean and process the card now.  There are still some
633   // stale card cases that may be detected by iteration and dealt with
634   // as iteration failure.
635   *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
636 
637   // This fence serves two purposes.  First, the card must be cleaned
638   // before processing the contents.  Second, we can't proceed with
639   // processing until after the read of top, for synchronization with
640   // possibly concurrent humongous object allocation.  It's okay that
641   // reading top and reading type were racy wrto each other.  We need
642   // both set, in any order, to proceed.
643   OrderAccess::fence();
644 
645   // Don't use addr_for(card_ptr + 1) which can ask for
646   // a card beyond the heap.
647   HeapWord* end = start + G1CardTable::card_size_in_words;
648   MemRegion dirty_region(start, MIN2(scan_limit, end));
649   assert(!dirty_region.is_empty(), "sanity");
650 
651   G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
652 
653   bool card_processed =
654     r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
655 
656   // If unable to process the card then we encountered an unparsable
657   // part of the heap (e.g. a partially allocated object) while
658   // processing a stale card.  Despite the card being stale, redirty
659   // and re-enqueue, because we've already cleaned the card.  Without
660   // this we could incorrectly discard a non-stale card.
661   if (!card_processed) {
662     // The card might have gotten re-dirtied and re-enqueued while we
663     // worked.  (In fact, it's pretty likely.)
664     if (*card_ptr != G1CardTable::dirty_card_val()) {
665       *card_ptr = G1CardTable::dirty_card_val();
666       MutexLockerEx x(Shared_DirtyCardQ_lock,
667                       Mutex::_no_safepoint_check_flag);
668       DirtyCardQueue* sdcq =
669         G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
670       sdcq->enqueue(card_ptr);
671     }
672   } else {
673     _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
674   }
675 }
676 
677 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
678                                      G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
679   assert(_g1h->is_gc_active(), "Only call during GC");
680 
681   check_card_ptr(card_ptr, _ct);
682 
683   // If the card is no longer dirty, nothing to do. This covers cards that were already
684   // scanned as parts of the remembered sets.
685   if (*card_ptr != G1CardTable::dirty_card_val()) {
686     return false;
687   }
688 
689   // We claim lazily (so races are possible but they're benign), which reduces the
690   // number of potential duplicate scans (multiple threads may enqueue the same card twice).
691   *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
692 
693   // Construct the region representing the card.
694   HeapWord* card_start = _ct->addr_for(card_ptr);
695   // And find the region containing it.
696   uint const card_region_idx = _g1h->addr_to_region(card_start);
697 
698   _scan_state->add_dirty_region(card_region_idx);
699   HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
700   if (scan_limit <= card_start) {
701     // If the card starts above the area in the region containing objects to scan, skip it.
702     return false;
703   }
704 
705   // Don't use addr_for(card_ptr + 1) which can ask for
706   // a card beyond the heap.
707   HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
708   MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
709   assert(!dirty_region.is_empty(), "sanity");
710 
711   HeapRegion* const card_region = _g1h->region_at(card_region_idx);
712   update_rs_cl->set_region(card_region);
713   bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
714   assert(card_processed, "must be");
715   return true;
716 }
717 
718 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
719   if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
720       (period_count % G1SummarizeRSetStatsPeriod == 0)) {
721 
722     G1RemSetSummary current(this);
723     _prev_period_summary.subtract_from(&current);
724 
725     Log(gc, remset) log;
726     log.trace("%s", header);
727     ResourceMark rm;
728     LogStream ls(log.trace());
729     _prev_period_summary.print_on(&ls);
730 
731     _prev_period_summary.set(&current);
732   }
733 }
734 
735 void G1RemSet::print_summary_info() {
736   Log(gc, remset, exit) log;
737   if (log.is_trace()) {
738     log.trace(" Cumulative RS summary");
739     G1RemSetSummary current(this);
740     ResourceMark rm;
741     LogStream ls(log.trace());
742     current.print_on(&ls);
743   }
744 }
745 
746 class G1RebuildRemSetTask: public AbstractGangTask {
747   // Aggregate the counting data that was constructed concurrently
748   // with marking.
749   class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
750     G1ConcurrentMark* _cm;
751     G1RebuildRemSetClosure _update_cl;
752 
753     // Applies _update_cl to the references of the given object, limiting objArrays
754     // to the given MemRegion. Returns the amount of words actually scanned.
755     size_t scan_for_references(oop const obj, MemRegion mr) {
756       size_t const obj_size = obj->size();
757       // All non-objArrays and objArrays completely within the mr
758       // can be scanned without passing the mr.
759       if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
760         obj->oop_iterate(&_update_cl);
761         return obj_size;
762       }
763       // This path is for objArrays crossing the given MemRegion. Only scan the
764       // area within the MemRegion.
765       obj->oop_iterate(&_update_cl, mr);
766       return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
767     }
768 
769     // A humongous object is live (with respect to the scanning) either
770     // a) it is marked on the bitmap as such
771     // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
772     bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
773       return bitmap->is_marked(humongous_obj) || (tars > tams);
774     }
775 
776     // Iterator over the live objects within the given MemRegion.
777     class LiveObjIterator : public StackObj {
778       const G1CMBitMap* const _bitmap;
779       const HeapWord* _tams;
780       const MemRegion _mr;
781       HeapWord* _current;
782 
783       bool is_below_tams() const {
784         return _current < _tams;
785       }
786 
787       bool is_live(HeapWord* obj) const {
788         return !is_below_tams() || _bitmap->is_marked(obj);
789       }
790 
791       HeapWord* bitmap_limit() const {
792         return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
793       }
794 
795       void move_if_below_tams() {
796         if (is_below_tams() && has_next()) {
797           _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
798         }
799       }
800     public:
801       LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr
802           _bitmap(bitmap),
803           _tams(tams),
804           _mr(mr),
805           _current(first_oop_into_mr) {
806 
807         assert(_current <= _mr.start(),
808                "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
809                p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
810 
811         // Step to the next live object within the MemRegion if needed.
812         if (is_live(_current)) {
813           // Non-objArrays were scanned by the previous part of that region.
814           if (_current < mr.start() && !oop(_current)->is_objArray()) {
815             _current += oop(_current)->size();
816             // We might have positioned _current on a non-live object. Reposition to the next
817             // live one if needed.
818             move_if_below_tams();
819           }
820         } else {
821           // The object at _current can only be dead if below TAMS, so we can use the bitmap.
822           // immediately.
823           _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
824           assert(_current == _mr.end() || is_live(_current),
825                  "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
826                  p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
827         }
828       }
829 
830       void move_to_next() {
831         _current += next()->size();
832         move_if_below_tams();
833       }
834 
835       oop next() const {
836         oop result = oop(_current);
837         assert(is_live(_current),
838                "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d
839                p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
840         return result;
841       }
842 
843       bool has_next() const {
844         return _current < _mr.end();
845       }
846     };
847 
848     // Rebuild remembered sets in the part of the region specified by mr and hr.
849     // Objects between the bottom of the region and the TAMS are checked for liveness
850     // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
851     // Returns the number of live words between bottom and TAMS.
852     size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
853                                      HeapWord* const top_at_mark_start,
854                                      HeapWord* const top_at_rebuild_start,
855                                      HeapRegion* hr,
856                                      MemRegion mr) {
857       size_t marked_words = 0;
858 
859       if (hr->is_humongous()) {
860         oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
861         if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
862           // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
863           // however in case of humongous objects it is sufficient to scan the encompassing
864           // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
865           // two areas will be zero sized. I.e. TAMS is either
866           // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
867           // value: this would mean that TAMS points somewhere into the object.
868           assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
869                  "More than one object in the humongous region?");
870           humongous_obj->oop_iterate(&_update_cl, mr);
871           return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->siz
872         } else {
873           return 0;
874         }
875       }
876 
877       for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()
878         oop obj = it.next();
879         size_t scanned_size = scan_for_references(obj, mr);
880         if ((HeapWord*)obj < top_at_mark_start) {
881           marked_words += scanned_size;
882         }
883       }
884 
885       return marked_words * HeapWordSize;
886     }
887 public:
888   G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
889                                    G1ConcurrentMark* cm,
890                                    uint worker_id) :
891     HeapRegionClosure(),
892     _cm(cm),
893     _update_cl(g1h, worker_id) { }
894 
895     bool do_heap_region(HeapRegion* hr) {
896       if (_cm->has_aborted()) {
897         return true;
898       }
899 
900       uint const region_idx = hr->hrm_index();
901       DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
902       assert(top_at_rebuild_start_check == NULL ||
903              top_at_rebuild_start_check > hr->bottom(),
904              "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
905              p2i(top_at_rebuild_start_check), p2i(hr->bottom()),  region_idx, hr->get_type_str());
906 
907       size_t total_marked_bytes = 0;
908       size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
909 
910       HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
911 
912       HeapWord* cur = hr->bottom();
913       while (cur < hr->end()) {
914         // After every iteration (yield point) we need to check whether the region's
915         // TARS changed due to e.g. eager reclaim.
916         HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
917         if (top_at_rebuild_start == NULL) {
918           return false;
919         }
920 
921         MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_word
922         if (next_chunk.is_empty()) {
923           break;
924         }
925 
926         const Ticks start = Ticks::now();
927         size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
928                                                         top_at_mark_start,
929                                                         top_at_rebuild_start,
930                                                         hr,
931                                                         next_chunk);
932         Tickspan time = Ticks::now() - start;
933 
934         log_trace(gc, remset, tracking)("Rebuilt region %u "
935                                         "live " SIZE_FORMAT " "
936                                         "time %.3fms "
937                                         "marked bytes " SIZE_FORMAT " "
938                                         "bot " PTR_FORMAT " "
939                                         "TAMS " PTR_FORMAT " "
940                                         "TARS " PTR_FORMAT,
941                                         region_idx,
942                                         _cm->liveness(region_idx) * HeapWordSize,
943                                         time.seconds() * 1000.0,
944                                         marked_bytes,
945                                         p2i(hr->bottom()),
946                                         p2i(top_at_mark_start),
947                                         p2i(top_at_rebuild_start));
948 
949         if (marked_bytes > 0) {
950           total_marked_bytes += marked_bytes;
951         }
952         cur += chunk_size_in_words;
953 
954         _cm->do_yield_check();
955         if (_cm->has_aborted()) {
956           return true;
957         }
958       }
959       // In the final iteration of the loop the region might have been eagerly reclaimed.
960       // Simply filter out those regions. We can not just use region type because there
961       // might have already been new allocations into these regions.
962       DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
963       assert(top_at_rebuild_start == NULL ||
964              total_marked_bytes == hr->marked_bytes(),
965              "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE
966              "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
967              total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
968              p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
969        // Abort state may have changed after the yield check.
970       return _cm->has_aborted();
971     }
972   };
973 
974   HeapRegionClaimer _hr_claimer;
975   G1ConcurrentMark* _cm;
976 
977   uint _worker_id_offset;
978 public:
979   G1RebuildRemSetTask(G1ConcurrentMark* cm,
980                       uint n_workers,
981                       uint worker_id_offset) :
982       AbstractGangTask("G1 Rebuild Remembered Set"),
983       _hr_claimer(n_workers),
984       _cm(cm),
985       _worker_id_offset(worker_id_offset) {
986   }
987 
988   void work(uint worker_id) {
989     SuspendibleThreadSetJoiner sts_join;
990 
991     G1CollectedHeap* g1h = G1CollectedHeap::heap();
992 
993     G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
994     g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
995   }
996 };
997 
998 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
999                                WorkGang* workers,
1000                                uint worker_id_offset) {
1001   uint num_workers = workers->active_workers();
1002 
1003   G1RebuildRemSetTask cl(cm,
1004                          num_workers,
1005                          worker_id_offset);
1006   workers->run_task(&cl, num_workers);
1007 }
--- EOF ---