1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/dirtyCardQueue.hpp"
27 #include "gc/g1/g1BarrierSet.hpp"
28 #include "gc/g1/g1BlockOffsetTable.inline.hpp"
29 #include "gc/g1/g1CardTable.inline.hpp"
30 #include "gc/g1/g1CollectedHeap.inline.hpp"
31 #include "gc/g1/g1ConcurrentRefine.hpp"
32 #include "gc/g1/g1FromCardCache.hpp"
33 #include "gc/g1/g1GCPhaseTimes.hpp"
34 #include "gc/g1/g1HotCardCache.hpp"
35 #include "gc/g1/g1OopClosures.inline.hpp"
36 #include "gc/g1/g1RootClosures.hpp"
37 #include "gc/g1/g1RemSet.hpp"
38 #include "gc/g1/heapRegion.inline.hpp"
39 #include "gc/g1/heapRegionManager.inline.hpp"
40 #include "gc/g1/heapRegionRemSet.hpp"
41 #include "gc/shared/gcTraceTime.inline.hpp"
42 #include "gc/shared/suspendibleThreadSet.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()) / 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 uint const region_idx = r->hrm_index();
343
344 if (_scan_state->claim_iter(region_idx)) {
345 // If we ever free the collection set concurrently, we should also
346 // clear the card table concurrently therefore we won't need to
347 // add regions of the collection set to the dirty cards region.
348 _scan_state->add_dirty_region(region_idx);
349 }
350
351 // We claim cards in blocks so as to reduce the contention.
352 size_t const block_size = G1RSetScanBlockSize;
353
354 HeapRegionRemSetIterator iter(r->rem_set());
355 size_t card_index;
356
357 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
358 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
359 if (current_card >= claimed_card_block + block_size) {
360 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
361 }
362 if (current_card < claimed_card_block) {
363 _cards_skipped++;
364 continue;
365 }
366 _cards_claimed++;
367
368 // If the card is dirty, then G1 will scan it during Update RS.
369 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
370 continue;
371 }
372
373 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index);
374 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
375
376 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start),
377 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
378 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
379 if (card_start >= top) {
380 continue;
381 }
382
383 // We claim lazily (so races are possible but they're benign), which reduces the
384 // number of duplicate scans (the rsets of the regions in the cset can intersect).
385 // Claim the card after checking bounds above: the remembered set may contain
386 // random cards into current survivor, and we would then have an incorrectly
387 // claimed card in survivor space. Card table clear does not reset the card table
388 // of survivor space regions.
389 claim_card(card_index, region_idx_for_card);
390
391 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
392
393 scan_card(mr, region_idx_for_card);
394 }
395 }
396
397 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
398 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
399 }
400
401 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
402 assert(r->in_collection_set(),
403 "Should only be called on elements of the collection set but region %u is not.",
404 r->hrm_index());
405 uint const region_idx = r->hrm_index();
406
407 // Do an early out if we know we are complete.
408 if (_scan_state->iter_is_complete(region_idx)) {
409 return false;
410 }
411
412 {
413 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
414 scan_rem_set_roots(r);
415 }
416
417 if (_scan_state->set_iter_complete(region_idx)) {
418 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
419 // Scan the strong code root list attached to the current region
420 scan_strong_code_roots(r);
421 }
422 return false;
423 }
424
425 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
426 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
427 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i);
428 _g1h->collection_set_iterate_from(&cl, worker_i);
429
430 G1GCPhaseTimes* p = _g1p->phase_times();
431
432 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
433 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
434
435 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
436 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
437 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
438
439 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
440 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
441 }
442
443 // Closure used for updating rem sets. Only called during an evacuation pause.
444 class G1RefineCardClosure: public CardTableEntryClosure {
445 G1RemSet* _g1rs;
446 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
447
448 size_t _cards_scanned;
449 size_t _cards_skipped;
450 public:
451 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
452 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
453 {}
454
455 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
456 // The only time we care about recording cards that
457 // contain references that point into the collection set
458 // is during RSet updating within an evacuation pause.
459 // In this case worker_i should be the id of a GC worker thread.
460 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
461
462 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
463
464 if (card_scanned) {
465 _update_rs_cl->trim_queue_partially();
466 _cards_scanned++;
467 } else {
468 _cards_skipped++;
469 }
470 return true;
471 }
472
473 size_t cards_scanned() const { return _cards_scanned; }
474 size_t cards_skipped() const { return _cards_skipped; }
475 };
476
477 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
478 G1GCPhaseTimes* p = _g1p->phase_times();
479
480 // Apply closure to log entries in the HCC.
481 if (G1HotCardCache::default_use_cache()) {
482 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
483
484 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss, worker_i);
485 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
486 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
487 }
488
489 // Now apply the closure to all remaining log entries.
490 {
491 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
492
493 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i);
494 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
495 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
496
497 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
498 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
499 }
500 }
501
502 void G1RemSet::cleanupHRRS() {
503 HeapRegionRemSet::cleanup();
504 }
505
506 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
507 update_rem_set(pss, worker_i);
508 scan_rem_set(pss, worker_i);;
509 }
510
511 void G1RemSet::prepare_for_oops_into_collection_set_do() {
512 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
513 dcqs.concatenate_logs();
514
515 _scan_state->reset();
516 }
517
518 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
519 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times();
520
521 // Set all cards back to clean.
522 double start = os::elapsedTime();
523 _scan_state->clear_card_table(_g1h->workers());
524 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
525 }
526
527 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
528 #ifdef ASSERT
529 G1CollectedHeap* g1h = G1CollectedHeap::heap();
530 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
531 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
532 p2i(card_ptr),
533 ct->index_for(ct->addr_for(card_ptr)),
534 p2i(ct->addr_for(card_ptr)),
535 g1h->addr_to_region(ct->addr_for(card_ptr)));
536 #endif
537 }
538
539 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
540 uint worker_i) {
541 assert(!_g1h->is_gc_active(), "Only call concurrently");
542
543 check_card_ptr(card_ptr, _ct);
544
545 // If the card is no longer dirty, nothing to do.
546 if (*card_ptr != G1CardTable::dirty_card_val()) {
547 return;
548 }
549
550 // Construct the region representing the card.
551 HeapWord* start = _ct->addr_for(card_ptr);
552 // And find the region containing it.
553 HeapRegion* r = _g1h->heap_region_containing(start);
554
555 // This check is needed for some uncommon cases where we should
556 // ignore the card.
557 //
558 // The region could be young. Cards for young regions are
559 // distinctly marked (set to g1_young_gen), so the post-barrier will
560 // filter them out. However, that marking is performed
561 // concurrently. A write to a young object could occur before the
562 // card has been marked young, slipping past the filter.
563 //
564 // The card could be stale, because the region has been freed since
565 // the card was recorded. In this case the region type could be
566 // anything. If (still) free or (reallocated) young, just ignore
567 // it. If (reallocated) old or humongous, the later card trimming
568 // and additional checks in iteration may detect staleness. At
569 // worst, we end up processing a stale card unnecessarily.
570 //
571 // In the normal (non-stale) case, the synchronization between the
572 // enqueueing of the card and processing it here will have ensured
573 // we see the up-to-date region type here.
574 if (!r->is_old_or_humongous_or_archive()) {
575 return;
576 }
577
578 // The result from the hot card cache insert call is either:
579 // * pointer to the current card
580 // (implying that the current card is not 'hot'),
581 // * null
582 // (meaning we had inserted the card ptr into the "hot" card cache,
583 // which had some headroom),
584 // * a pointer to a "hot" card that was evicted from the "hot" cache.
585 //
586
587 if (_hot_card_cache->use_cache()) {
588 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
589
590 const jbyte* orig_card_ptr = card_ptr;
591 card_ptr = _hot_card_cache->insert(card_ptr);
592 if (card_ptr == NULL) {
593 // There was no eviction. Nothing to do.
594 return;
595 } else if (card_ptr != orig_card_ptr) {
596 // Original card was inserted and an old card was evicted.
597 start = _ct->addr_for(card_ptr);
598 r = _g1h->heap_region_containing(start);
599
600 // Check whether the region formerly in the cache should be
601 // ignored, as discussed earlier for the original card. The
602 // region could have been freed while in the cache.
603 if (!r->is_old_or_humongous_or_archive()) {
604 return;
605 }
606 } // Else we still have the original card.
607 }
608
609 // Trim the region designated by the card to what's been allocated
610 // in the region. The card could be stale, or the card could cover
611 // (part of) an object at the end of the allocated space and extend
612 // beyond the end of allocation.
613
614 // Non-humongous objects are only allocated in the old-gen during
615 // GC, so if region is old then top is stable. Humongous object
616 // allocation sets top last; if top has not yet been set, this is
617 // a stale card and we'll end up with an empty intersection. If
618 // this is not a stale card, the synchronization between the
619 // enqueuing of the card and processing it here will have ensured
620 // we see the up-to-date top here.
621 HeapWord* scan_limit = r->top();
622
623 if (scan_limit <= start) {
624 // If the trimmed region is empty, the card must be stale.
625 return;
626 }
627
628 // Okay to clean and process the card now. There are still some
629 // stale card cases that may be detected by iteration and dealt with
630 // as iteration failure.
631 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
632
633 // This fence serves two purposes. First, the card must be cleaned
634 // before processing the contents. Second, we can't proceed with
635 // processing until after the read of top, for synchronization with
636 // possibly concurrent humongous object allocation. It's okay that
637 // reading top and reading type were racy wrto each other. We need
638 // both set, in any order, to proceed.
639 OrderAccess::fence();
640
641 // Don't use addr_for(card_ptr + 1) which can ask for
642 // a card beyond the heap.
643 HeapWord* end = start + G1CardTable::card_size_in_words;
644 MemRegion dirty_region(start, MIN2(scan_limit, end));
645 assert(!dirty_region.is_empty(), "sanity");
646
647 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
648
649 bool card_processed =
650 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
651
652 // If unable to process the card then we encountered an unparsable
653 // part of the heap (e.g. a partially allocated object) while
654 // processing a stale card. Despite the card being stale, redirty
655 // and re-enqueue, because we've already cleaned the card. Without
656 // this we could incorrectly discard a non-stale card.
657 if (!card_processed) {
658 // The card might have gotten re-dirtied and re-enqueued while we
659 // worked. (In fact, it's pretty likely.)
660 if (*card_ptr != G1CardTable::dirty_card_val()) {
661 *card_ptr = G1CardTable::dirty_card_val();
662 MutexLockerEx x(Shared_DirtyCardQ_lock,
663 Mutex::_no_safepoint_check_flag);
664 DirtyCardQueue* sdcq =
665 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
666 sdcq->enqueue(card_ptr);
667 }
668 } else {
669 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
670 }
671 }
672
673 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
674 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
675 assert(_g1h->is_gc_active(), "Only call during GC");
676
677 check_card_ptr(card_ptr, _ct);
678
679 // If the card is no longer dirty, nothing to do. This covers cards that were already
680 // scanned as parts of the remembered sets.
681 if (*card_ptr != G1CardTable::dirty_card_val()) {
682 return false;
683 }
684
685 // We claim lazily (so races are possible but they're benign), which reduces the
686 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
687 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
688
689 // Construct the region representing the card.
690 HeapWord* card_start = _ct->addr_for(card_ptr);
691 // And find the region containing it.
692 uint const card_region_idx = _g1h->addr_to_region(card_start);
693
694 _scan_state->add_dirty_region(card_region_idx);
695 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
696 if (scan_limit <= card_start) {
697 // If the card starts above the area in the region containing objects to scan, skip it.
698 return false;
699 }
700
701 // Don't use addr_for(card_ptr + 1) which can ask for
702 // a card beyond the heap.
703 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
704 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
705 assert(!dirty_region.is_empty(), "sanity");
706
707 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
708 update_rs_cl->set_region(card_region);
709 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
710 assert(card_processed, "must be");
711 return true;
712 }
713
714 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
715 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
716 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
717
718 G1RemSetSummary current(this);
719 _prev_period_summary.subtract_from(¤t);
720
721 Log(gc, remset) log;
722 log.trace("%s", header);
723 ResourceMark rm;
724 LogStream ls(log.trace());
725 _prev_period_summary.print_on(&ls);
726
727 _prev_period_summary.set(¤t);
728 }
729 }
730
731 void G1RemSet::print_summary_info() {
732 Log(gc, remset, exit) log;
733 if (log.is_trace()) {
734 log.trace(" Cumulative RS summary");
735 G1RemSetSummary current(this);
736 ResourceMark rm;
737 LogStream ls(log.trace());
738 current.print_on(&ls);
739 }
740 }
741
742 class G1RebuildRemSetTask: public AbstractGangTask {
743 // Aggregate the counting data that was constructed concurrently
744 // with marking.
745 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
746 G1ConcurrentMark* _cm;
747 G1RebuildRemSetClosure _update_cl;
748
749 // Applies _update_cl to the references of the given object, limiting objArrays
750 // to the given MemRegion. Returns the amount of words actually scanned.
751 size_t scan_for_references(oop const obj, MemRegion mr) {
752 size_t const obj_size = obj->size();
753 // All non-objArrays and objArrays completely within the mr
754 // can be scanned without passing the mr.
755 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
756 obj->oop_iterate(&_update_cl);
757 return obj_size;
758 }
759 // This path is for objArrays crossing the given MemRegion. Only scan the
760 // area within the MemRegion.
761 obj->oop_iterate(&_update_cl, mr);
762 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
763 }
764
765 // A humongous object is live (with respect to the scanning) either
766 // a) it is marked on the bitmap as such
767 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
768 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
769 return bitmap->is_marked(humongous_obj) || (tars > tams);
770 }
771
772 // Iterator over the live objects within the given MemRegion.
773 class LiveObjIterator : public StackObj {
774 const G1CMBitMap* const _bitmap;
775 const HeapWord* _tams;
776 const MemRegion _mr;
777 HeapWord* _current;
778
779 bool is_below_tams() const {
780 return _current < _tams;
781 }
782
783 bool is_live(HeapWord* obj) const {
784 return !is_below_tams() || _bitmap->is_marked(obj);
785 }
786
787 HeapWord* bitmap_limit() const {
788 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
789 }
790
791 void move_if_below_tams() {
792 if (is_below_tams() && has_next()) {
793 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
794 }
795 }
796 public:
797 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
798 _bitmap(bitmap),
799 _tams(tams),
800 _mr(mr),
801 _current(first_oop_into_mr) {
802
803 assert(_current <= _mr.start(),
804 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
805 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
806
807 // Step to the next live object within the MemRegion if needed.
808 if (is_live(_current)) {
809 // Non-objArrays were scanned by the previous part of that region.
810 if (_current < mr.start() && !oop(_current)->is_objArray()) {
811 _current += oop(_current)->size();
812 // We might have positioned _current on a non-live object. Reposition to the next
813 // live one if needed.
814 move_if_below_tams();
815 }
816 } else {
817 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
818 // immediately.
819 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
820 assert(_current == _mr.end() || is_live(_current),
821 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
822 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
823 }
824 }
825
826 void move_to_next() {
827 _current += next()->size();
828 move_if_below_tams();
829 }
830
831 oop next() const {
832 oop result = oop(_current);
833 assert(is_live(_current),
834 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
835 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
836 return result;
837 }
838
839 bool has_next() const {
840 return _current < _mr.end();
841 }
842 };
843
844 // Rebuild remembered sets in the part of the region specified by mr and hr.
845 // Objects between the bottom of the region and the TAMS are checked for liveness
846 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
847 // Returns the number of live words between bottom and TAMS.
848 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
849 HeapWord* const top_at_mark_start,
850 HeapWord* const top_at_rebuild_start,
851 HeapRegion* hr,
852 MemRegion mr) {
853 size_t marked_words = 0;
854
855 if (hr->is_humongous()) {
856 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
857 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
858 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
859 // however in case of humongous objects it is sufficient to scan the encompassing
860 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
861 // two areas will be zero sized. I.e. TAMS is either
862 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
863 // value: this would mean that TAMS points somewhere into the object.
864 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
865 "More than one object in the humongous region?");
866 humongous_obj->oop_iterate(&_update_cl, mr);
867 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
868 } else {
869 return 0;
870 }
871 }
872
873 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
874 oop obj = it.next();
875 size_t scanned_size = scan_for_references(obj, mr);
876 if ((HeapWord*)obj < top_at_mark_start) {
877 marked_words += scanned_size;
878 }
879 }
880
881 return marked_words * HeapWordSize;
882 }
883 public:
884 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
885 G1ConcurrentMark* cm,
886 uint worker_id) :
887 HeapRegionClosure(),
888 _cm(cm),
889 _update_cl(g1h, worker_id) { }
890
891 bool do_heap_region(HeapRegion* hr) {
892 if (_cm->has_aborted()) {
893 return true;
894 }
895
896 uint const region_idx = hr->hrm_index();
897 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
898 assert(top_at_rebuild_start_check == NULL ||
899 top_at_rebuild_start_check > hr->bottom(),
900 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
901 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
902
903 size_t total_marked_bytes = 0;
904 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
905
906 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
907
908 HeapWord* cur = hr->bottom();
909 while (cur < hr->end()) {
910 // After every iteration (yield point) we need to check whether the region's
911 // TARS changed due to e.g. eager reclaim.
912 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
913 if (top_at_rebuild_start == NULL) {
914 return false;
915 }
916
917 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
918 if (next_chunk.is_empty()) {
919 break;
920 }
921
922 const Ticks start = Ticks::now();
923 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
924 top_at_mark_start,
925 top_at_rebuild_start,
926 hr,
927 next_chunk);
928 Tickspan time = Ticks::now() - start;
929
930 log_trace(gc, remset, tracking)("Rebuilt region %u "
931 "live " SIZE_FORMAT " "
932 "time %.3fms "
933 "marked bytes " SIZE_FORMAT " "
934 "bot " PTR_FORMAT " "
935 "TAMS " PTR_FORMAT " "
936 "TARS " PTR_FORMAT,
937 region_idx,
938 _cm->liveness(region_idx) * HeapWordSize,
939 time.seconds() * 1000.0,
940 marked_bytes,
941 p2i(hr->bottom()),
942 p2i(top_at_mark_start),
943 p2i(top_at_rebuild_start));
944
945 if (marked_bytes > 0) {
946 total_marked_bytes += marked_bytes;
947 }
948 cur += chunk_size_in_words;
949
950 _cm->do_yield_check();
951 if (_cm->has_aborted()) {
952 return true;
953 }
954 }
955 // In the final iteration of the loop the region might have been eagerly reclaimed.
956 // Simply filter out those regions. We can not just use region type because there
957 // might have already been new allocations into these regions.
958 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
959 assert(top_at_rebuild_start == NULL ||
960 total_marked_bytes == hr->marked_bytes(),
961 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
962 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
963 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
964 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
965 // Abort state may have changed after the yield check.
966 return _cm->has_aborted();
967 }
968 };
969
970 HeapRegionClaimer _hr_claimer;
971 G1ConcurrentMark* _cm;
972
973 uint _worker_id_offset;
974 public:
975 G1RebuildRemSetTask(G1ConcurrentMark* cm,
976 uint n_workers,
977 uint worker_id_offset) :
978 AbstractGangTask("G1 Rebuild Remembered Set"),
979 _hr_claimer(n_workers),
980 _cm(cm),
981 _worker_id_offset(worker_id_offset) {
982 }
983
984 void work(uint worker_id) {
985 SuspendibleThreadSetJoiner sts_join;
986
987 G1CollectedHeap* g1h = G1CollectedHeap::heap();
988
989 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
990 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
991 }
992 };
993
994 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
995 WorkGang* workers,
996 uint worker_id_offset) {
997 uint num_workers = workers->active_workers();
998
999 G1RebuildRemSetTask cl(cm,
1000 num_workers,
1001 worker_id_offset);
1002 workers->run_task(&cl, num_workers);
1003 }