1 /*
2 * Copyright (c) 2001, 2019, 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/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/g1DirtyCardQueue.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 "jfr/jfrEvents.hpp"
44 #include "memory/iterator.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "oops/access.inline.hpp"
47 #include "oops/oop.inline.hpp"
48 #include "runtime/os.hpp"
49 #include "utilities/align.hpp"
50 #include "utilities/globalDefinitions.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 // Flag for every region whether it is in the _dirty_region_buffer already
118 // to avoid duplicates.
119 bool volatile* _in_dirty_region_buffer;
120 size_t _cur_dirty_region;
121
122 // Creates a snapshot of the current _top values at the start of collection to
123 // filter out card marks that we do not want to scan.
124 class G1ResetScanTopClosure : public HeapRegionClosure {
125 private:
126 HeapWord** _scan_top;
127 public:
128 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
129
130 virtual bool do_heap_region(HeapRegion* r) {
131 uint hrm_index = r->hrm_index();
132 if (!r->in_collection_set() && r->is_old_or_humongous_or_archive() && !r->is_empty()) {
133 _scan_top[hrm_index] = r->top();
134 } else {
135 _scan_top[hrm_index] = NULL;
136 }
137 return false;
138 }
139 };
140
141 // For each region, contains the maximum top() value to be used during this garbage
142 // collection. Subsumes common checks like filtering out everything but old and
143 // humongous regions outside the collection set.
144 // This is valid because we are not interested in scanning stray remembered set
145 // entries from free or archive regions.
146 HeapWord** _scan_top;
147 public:
148 G1RemSetScanState() :
149 _max_regions(0),
150 _iter_states(NULL),
151 _iter_claims(NULL),
152 _dirty_region_buffer(NULL),
153 _in_dirty_region_buffer(NULL),
154 _cur_dirty_region(0),
155 _scan_top(NULL) {
156 }
157
158 ~G1RemSetScanState() {
159 if (_iter_states != NULL) {
160 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
161 }
162 if (_iter_claims != NULL) {
163 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
164 }
165 if (_dirty_region_buffer != NULL) {
166 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
167 }
168 if (_in_dirty_region_buffer != NULL) {
169 FREE_C_HEAP_ARRAY(bool, _in_dirty_region_buffer);
170 }
171 if (_scan_top != NULL) {
172 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
173 }
174 }
175
176 void initialize(uint max_regions) {
177 assert(_iter_states == NULL, "Must not be initialized twice");
178 assert(_iter_claims == NULL, "Must not be initialized twice");
179 _max_regions = max_regions;
180 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
181 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
182 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
183 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(bool, max_regions, mtGC);
184 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
185 }
186
187 void reset() {
188 for (uint i = 0; i < _max_regions; i++) {
189 _iter_states[i] = Unclaimed;
190 _scan_top[i] = NULL;
191 }
192
193 G1ResetScanTopClosure cl(_scan_top);
194 G1CollectedHeap::heap()->heap_region_iterate(&cl);
195
196 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
197 memset((void*)_in_dirty_region_buffer, false, _max_regions * sizeof(bool));
198 _cur_dirty_region = 0;
199 }
200
201 // Attempt to claim the remembered set of the region for iteration. Returns true
202 // if this call caused the transition from Unclaimed to Claimed.
203 inline bool claim_iter(uint region) {
204 assert(region < _max_regions, "Tried to access invalid region %u", region);
205 if (_iter_states[region] != Unclaimed) {
206 return false;
207 }
208 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
209 return (res == Unclaimed);
210 }
211
212 // Try to atomically sets the iteration state to "complete". Returns true for the
213 // thread that caused the transition.
214 inline bool set_iter_complete(uint region) {
215 if (iter_is_complete(region)) {
216 return false;
217 }
218 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
219 return (res == Claimed);
220 }
221
222 // Returns true if the region's iteration is complete.
223 inline bool iter_is_complete(uint region) const {
224 assert(region < _max_regions, "Tried to access invalid region %u", region);
225 return _iter_states[region] == Complete;
226 }
227
228 // The current position within the remembered set of the given region.
229 inline size_t iter_claimed(uint region) const {
230 assert(region < _max_regions, "Tried to access invalid region %u", region);
231 return _iter_claims[region];
232 }
233
234 // Claim the next block of cards within the remembered set of the region with
235 // step size.
236 inline size_t iter_claimed_next(uint region, size_t step) {
237 return Atomic::add(step, &_iter_claims[region]) - step;
238 }
239
240 void add_dirty_region(uint region) {
241 if (_in_dirty_region_buffer[region]) {
242 return;
243 }
244
245 if (!Atomic::cmpxchg(true, &_in_dirty_region_buffer[region], false)) {
246 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
247 _dirty_region_buffer[allocated] = region;
248 }
249 }
250
251 HeapWord* scan_top(uint region_idx) const {
252 return _scan_top[region_idx];
253 }
254
255 // Clear the card table of "dirty" regions.
256 void clear_card_table(WorkGang* workers) {
257 if (_cur_dirty_region == 0) {
258 return;
259 }
260
261 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
262 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
263 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
264
265 // Iterate over the dirty cards region list.
266 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
267
268 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
269 "units of work for " SIZE_FORMAT " regions.",
270 cl.name(), num_workers, num_chunks, _cur_dirty_region);
271 workers->run_task(&cl, num_workers);
272
273 #ifndef PRODUCT
274 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
275 #endif
276 }
277 };
278
279 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
280 G1CardTable* ct,
281 G1HotCardCache* hot_card_cache) :
282 _scan_state(new G1RemSetScanState()),
283 _prev_period_summary(),
284 _g1h(g1h),
285 _num_conc_refined_cards(0),
286 _ct(ct),
287 _g1p(_g1h->policy()),
288 _hot_card_cache(hot_card_cache) {
289 }
290
291 G1RemSet::~G1RemSet() {
292 if (_scan_state != NULL) {
293 delete _scan_state;
294 }
295 }
296
297 uint G1RemSet::num_par_rem_sets() {
298 return G1DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
299 }
300
301 void G1RemSet::initialize(size_t capacity, uint max_regions) {
302 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
303 _scan_state->initialize(max_regions);
304 }
305
306 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
307 G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
308 G1ParScanThreadState* pss,
309 G1GCPhaseTimes::GCParPhases phase,
310 uint worker_i) :
311 _g1h(G1CollectedHeap::heap()),
312 _ct(_g1h->card_table()),
313 _pss(pss),
314 _scan_objs_on_card_cl(scan_obj_on_card),
315 _scan_state(scan_state),
316 _phase(phase),
317 _worker_i(worker_i),
318 _cards_scanned(0),
319 _cards_claimed(0),
320 _cards_skipped(0),
321 _rem_set_root_scan_time(),
322 _rem_set_trim_partially_time(),
323 _strong_code_root_scan_time(),
324 _strong_code_trim_partially_time() {
325 }
326
327 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
328 _ct->set_card_claimed(card_index);
329 _scan_state->add_dirty_region(region_idx_for_card);
330 }
331
332 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
333 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
334 assert(!card_region->is_young(), "Should not scan card in young region %u", region_idx_for_card);
335 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
336 _scan_objs_on_card_cl->trim_queue_partially();
337 _cards_scanned++;
338 }
339
340 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
341 EventGCPhaseParallel event;
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 if (r->rem_set()->cardset_is_empty()) {
352 return;
353 }
354
355 // We claim cards in blocks so as to reduce the contention.
356 size_t const block_size = G1RSetScanBlockSize;
357
358 HeapRegionRemSetIterator iter(r->rem_set());
359 size_t card_index;
360
361 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
362 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
363 if (current_card >= claimed_card_block + block_size) {
364 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
365 }
366 if (current_card < claimed_card_block) {
367 _cards_skipped++;
368 continue;
369 }
370 _cards_claimed++;
371
372 HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
373 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
374
375 #ifdef ASSERT
376 HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
377 assert(hr == NULL || hr->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)->hrm_index());
379 #endif
380 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
381 if (card_start >= top) {
382 continue;
383 }
384
385 // If the card is dirty, then G1 will scan it during Update RS.
386 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
387 continue;
388 }
389
390 // We claim lazily (so races are possible but they're benign), which reduces the
391 // number of duplicate scans (the rsets of the regions in the cset can intersect).
392 // Claim the card after checking bounds above: the remembered set may contain
393 // random cards into current survivor, and we would then have an incorrectly
394 // claimed card in survivor space. Card table clear does not reset the card table
395 // of survivor space regions.
396 claim_card(card_index, region_idx_for_card);
397
398 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
399
400 scan_card(mr, region_idx_for_card);
401 }
402 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
403 }
404
405 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
406 EventGCPhaseParallel event;
407 // We pass a weak code blobs closure to the remembered set scanning because we want to avoid
408 // treating the nmethods visited to act as roots for concurrent marking.
409 // We only want to make sure that the oops in the nmethods are adjusted with regard to the
410 // objects copied by the current evacuation.
411 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
412 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
413 }
414
415 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
416 assert(r->in_collection_set(),
417 "Should only be called on elements of the collection set but region %u is not.",
418 r->hrm_index());
419 uint const region_idx = r->hrm_index();
420
421 // Do an early out if we know we are complete.
422 if (_scan_state->iter_is_complete(region_idx)) {
423 return false;
424 }
425
426 {
427 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
428 scan_rem_set_roots(r);
429 }
430
431 if (_scan_state->set_iter_complete(region_idx)) {
432 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
433 // Scan the strong code root list attached to the current region
434 scan_strong_code_roots(r);
435 }
436 return false;
437 }
438
439 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
440 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
441 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, G1GCPhaseTimes::ScanRS, worker_i);
442 _g1h->collection_set_iterate_from(&cl, worker_i);
443
444 G1GCPhaseTimes* p = _g1p->phase_times();
445
446 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
447 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
448
449 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
450 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
451 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
452
453 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
454 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
455 }
456
457 // Closure used for updating rem sets. Only called during an evacuation pause.
458 class G1RefineCardClosure: public G1CardTableEntryClosure {
459 G1RemSet* _g1rs;
460 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
461
462 size_t _cards_scanned;
463 size_t _cards_skipped;
464 public:
465 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
466 _g1rs(g1h->rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
467 {}
468
469 bool do_card_ptr(CardValue* card_ptr, uint worker_i) {
470 // The only time we care about recording cards that
471 // contain references that point into the collection set
472 // is during RSet updating within an evacuation pause.
473 // In this case worker_i should be the id of a GC worker thread.
474 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
475
476 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
477
478 if (card_scanned) {
479 _update_rs_cl->trim_queue_partially();
480 _cards_scanned++;
481 } else {
482 _cards_skipped++;
483 }
484 return true;
485 }
486
487 size_t cards_scanned() const { return _cards_scanned; }
488 size_t cards_skipped() const { return _cards_skipped; }
489 };
490
491 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
492 G1GCPhaseTimes* p = _g1p->phase_times();
493
494 // Apply closure to log entries in the HCC.
495 if (G1HotCardCache::default_use_cache()) {
496 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
497
498 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss);
499 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
500 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
501 }
502
503 // Now apply the closure to all remaining log entries.
504 {
505 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
506
507 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss);
508 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
509 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
510
511 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
512 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
513 }
514 }
515
516 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
517 update_rem_set(pss, worker_i);
518 scan_rem_set(pss, worker_i);;
519 }
520
521 void G1RemSet::prepare_for_oops_into_collection_set_do() {
522 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
523 dcqs.concatenate_logs();
524
525 _scan_state->reset();
526 }
527
528 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
529 G1GCPhaseTimes* phase_times = _g1h->phase_times();
530
531 // Set all cards back to clean.
532 double start = os::elapsedTime();
533 _scan_state->clear_card_table(_g1h->workers());
534 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
535 }
536
537 inline void check_card_ptr(CardTable::CardValue* card_ptr, G1CardTable* ct) {
538 #ifdef ASSERT
539 G1CollectedHeap* g1h = G1CollectedHeap::heap();
540 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
541 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
542 p2i(card_ptr),
543 ct->index_for(ct->addr_for(card_ptr)),
544 p2i(ct->addr_for(card_ptr)),
545 g1h->addr_to_region(ct->addr_for(card_ptr)));
546 #endif
547 }
548
549 void G1RemSet::refine_card_concurrently(CardValue* card_ptr,
550 uint worker_i) {
551 assert(!_g1h->is_gc_active(), "Only call concurrently");
552
553 // Construct the region representing the card.
554 HeapWord* start = _ct->addr_for(card_ptr);
555 // And find the region containing it.
556 HeapRegion* r = _g1h->heap_region_containing_or_null(start);
557
558 // If this is a (stale) card into an uncommitted region, exit.
559 if (r == NULL) {
560 return;
561 }
562
563 check_card_ptr(card_ptr, _ct);
564
565 // If the card is no longer dirty, nothing to do.
566 if (*card_ptr != G1CardTable::dirty_card_val()) {
567 return;
568 }
569
570 // This check is needed for some uncommon cases where we should
571 // ignore the card.
572 //
573 // The region could be young. Cards for young regions are
574 // distinctly marked (set to g1_young_gen), so the post-barrier will
575 // filter them out. However, that marking is performed
576 // concurrently. A write to a young object could occur before the
577 // card has been marked young, slipping past the filter.
578 //
579 // The card could be stale, because the region has been freed since
580 // the card was recorded. In this case the region type could be
581 // anything. If (still) free or (reallocated) young, just ignore
582 // it. If (reallocated) old or humongous, the later card trimming
583 // and additional checks in iteration may detect staleness. At
584 // worst, we end up processing a stale card unnecessarily.
585 //
586 // In the normal (non-stale) case, the synchronization between the
587 // enqueueing of the card and processing it here will have ensured
588 // we see the up-to-date region type here.
589 if (!r->is_old_or_humongous_or_archive()) {
590 return;
591 }
592
593 // The result from the hot card cache insert call is either:
594 // * pointer to the current card
595 // (implying that the current card is not 'hot'),
596 // * null
597 // (meaning we had inserted the card ptr into the "hot" card cache,
598 // which had some headroom),
599 // * a pointer to a "hot" card that was evicted from the "hot" cache.
600 //
601
602 if (_hot_card_cache->use_cache()) {
603 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
604
605 const CardValue* orig_card_ptr = card_ptr;
606 card_ptr = _hot_card_cache->insert(card_ptr);
607 if (card_ptr == NULL) {
608 // There was no eviction. Nothing to do.
609 return;
610 } else if (card_ptr != orig_card_ptr) {
611 // Original card was inserted and an old card was evicted.
612 start = _ct->addr_for(card_ptr);
613 r = _g1h->heap_region_containing(start);
614
615 // Check whether the region formerly in the cache should be
616 // ignored, as discussed earlier for the original card. The
617 // region could have been freed while in the cache.
618 if (!r->is_old_or_humongous_or_archive()) {
619 return;
620 }
621 } // Else we still have the original card.
622 }
623
624 // Trim the region designated by the card to what's been allocated
625 // in the region. The card could be stale, or the card could cover
626 // (part of) an object at the end of the allocated space and extend
627 // beyond the end of allocation.
628
629 // Non-humongous objects are only allocated in the old-gen during
630 // GC, so if region is old then top is stable. Humongous object
631 // allocation sets top last; if top has not yet been set, this is
632 // a stale card and we'll end up with an empty intersection. If
633 // this is not a stale card, the synchronization between the
634 // enqueuing of the card and processing it here will have ensured
635 // we see the up-to-date top here.
636 HeapWord* scan_limit = r->top();
637
638 if (scan_limit <= start) {
639 // If the trimmed region is empty, the card must be stale.
640 return;
641 }
642
643 // Okay to clean and process the card now. There are still some
644 // stale card cases that may be detected by iteration and dealt with
645 // as iteration failure.
646 *const_cast<volatile CardValue*>(card_ptr) = G1CardTable::clean_card_val();
647
648 // This fence serves two purposes. First, the card must be cleaned
649 // before processing the contents. Second, we can't proceed with
650 // processing until after the read of top, for synchronization with
651 // possibly concurrent humongous object allocation. It's okay that
652 // reading top and reading type were racy wrto each other. We need
653 // both set, in any order, to proceed.
654 OrderAccess::fence();
655
656 // Don't use addr_for(card_ptr + 1) which can ask for
657 // a card beyond the heap.
658 HeapWord* end = start + G1CardTable::card_size_in_words;
659 MemRegion dirty_region(start, MIN2(scan_limit, end));
660 assert(!dirty_region.is_empty(), "sanity");
661
662 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
663
664 bool card_processed =
665 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
666
667 // If unable to process the card then we encountered an unparsable
668 // part of the heap (e.g. a partially allocated object) while
669 // processing a stale card. Despite the card being stale, redirty
670 // and re-enqueue, because we've already cleaned the card. Without
671 // this we could incorrectly discard a non-stale card.
672 if (!card_processed) {
673 // The card might have gotten re-dirtied and re-enqueued while we
674 // worked. (In fact, it's pretty likely.)
675 if (*card_ptr != G1CardTable::dirty_card_val()) {
676 *card_ptr = G1CardTable::dirty_card_val();
677 MutexLockerEx x(Shared_DirtyCardQ_lock,
678 Mutex::_no_safepoint_check_flag);
679 G1DirtyCardQueue* sdcq =
680 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
681 sdcq->enqueue(card_ptr);
682 }
683 } else {
684 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
685 }
686 }
687
688 bool G1RemSet::refine_card_during_gc(CardValue* card_ptr,
689 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
690 assert(_g1h->is_gc_active(), "Only call during GC");
691
692 // Construct the region representing the card.
693 HeapWord* card_start = _ct->addr_for(card_ptr);
694 // And find the region containing it.
695 uint const card_region_idx = _g1h->addr_to_region(card_start);
696
697 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
698 if (scan_limit == NULL) {
699 // This is a card into an uncommitted region. We need to bail out early as we
700 // should not access the corresponding card table entry.
701 return false;
702 }
703
704 check_card_ptr(card_ptr, _ct);
705
706 // If the card is no longer dirty, nothing to do. This covers cards that were already
707 // scanned as parts of the remembered sets.
708 if (*card_ptr != G1CardTable::dirty_card_val()) {
709 return false;
710 }
711
712 // We claim lazily (so races are possible but they're benign), which reduces the
713 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
714 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
715
716 _scan_state->add_dirty_region(card_region_idx);
717 if (scan_limit <= card_start) {
718 // If the card starts above the area in the region containing objects to scan, skip it.
719 return false;
720 }
721
722 // Don't use addr_for(card_ptr + 1) which can ask for
723 // a card beyond the heap.
724 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
725 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
726 assert(!dirty_region.is_empty(), "sanity");
727
728 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
729 assert(!card_region->is_young(), "Should not scan card in young region %u", card_region_idx);
730 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
731 assert(card_processed, "must be");
732 return true;
733 }
734
735 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
736 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
737 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
738
739 G1RemSetSummary current(this);
740 _prev_period_summary.subtract_from(¤t);
741
742 Log(gc, remset) log;
743 log.trace("%s", header);
744 ResourceMark rm;
745 LogStream ls(log.trace());
746 _prev_period_summary.print_on(&ls);
747
748 _prev_period_summary.set(¤t);
749 }
750 }
751
752 void G1RemSet::print_summary_info() {
753 Log(gc, remset, exit) log;
754 if (log.is_trace()) {
755 log.trace(" Cumulative RS summary");
756 G1RemSetSummary current(this);
757 ResourceMark rm;
758 LogStream ls(log.trace());
759 current.print_on(&ls);
760 }
761 }
762
763 class G1RebuildRemSetTask: public AbstractGangTask {
764 // Aggregate the counting data that was constructed concurrently
765 // with marking.
766 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
767 G1ConcurrentMark* _cm;
768 G1RebuildRemSetClosure _update_cl;
769
770 // Applies _update_cl to the references of the given object, limiting objArrays
771 // to the given MemRegion. Returns the amount of words actually scanned.
772 size_t scan_for_references(oop const obj, MemRegion mr) {
773 size_t const obj_size = obj->size();
774 // All non-objArrays and objArrays completely within the mr
775 // can be scanned without passing the mr.
776 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
777 obj->oop_iterate(&_update_cl);
778 return obj_size;
779 }
780 // This path is for objArrays crossing the given MemRegion. Only scan the
781 // area within the MemRegion.
782 obj->oop_iterate(&_update_cl, mr);
783 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
784 }
785
786 // A humongous object is live (with respect to the scanning) either
787 // a) it is marked on the bitmap as such
788 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
789 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
790 return bitmap->is_marked(humongous_obj) || (tars > tams);
791 }
792
793 // Iterator over the live objects within the given MemRegion.
794 class LiveObjIterator : public StackObj {
795 const G1CMBitMap* const _bitmap;
796 const HeapWord* _tams;
797 const MemRegion _mr;
798 HeapWord* _current;
799
800 bool is_below_tams() const {
801 return _current < _tams;
802 }
803
804 bool is_live(HeapWord* obj) const {
805 return !is_below_tams() || _bitmap->is_marked(obj);
806 }
807
808 HeapWord* bitmap_limit() const {
809 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
810 }
811
812 void move_if_below_tams() {
813 if (is_below_tams() && has_next()) {
814 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
815 }
816 }
817 public:
818 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
819 _bitmap(bitmap),
820 _tams(tams),
821 _mr(mr),
822 _current(first_oop_into_mr) {
823
824 assert(_current <= _mr.start(),
825 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
826 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
827
828 // Step to the next live object within the MemRegion if needed.
829 if (is_live(_current)) {
830 // Non-objArrays were scanned by the previous part of that region.
831 if (_current < mr.start() && !oop(_current)->is_objArray()) {
832 _current += oop(_current)->size();
833 // We might have positioned _current on a non-live object. Reposition to the next
834 // live one if needed.
835 move_if_below_tams();
836 }
837 } else {
838 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
839 // immediately.
840 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
841 assert(_current == _mr.end() || is_live(_current),
842 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
843 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
844 }
845 }
846
847 void move_to_next() {
848 _current += next()->size();
849 move_if_below_tams();
850 }
851
852 oop next() const {
853 oop result = oop(_current);
854 assert(is_live(_current),
855 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
856 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
857 return result;
858 }
859
860 bool has_next() const {
861 return _current < _mr.end();
862 }
863 };
864
865 // Rebuild remembered sets in the part of the region specified by mr and hr.
866 // Objects between the bottom of the region and the TAMS are checked for liveness
867 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
868 // Returns the number of live words between bottom and TAMS.
869 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
870 HeapWord* const top_at_mark_start,
871 HeapWord* const top_at_rebuild_start,
872 HeapRegion* hr,
873 MemRegion mr) {
874 size_t marked_words = 0;
875
876 if (hr->is_humongous()) {
877 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
878 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
879 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
880 // however in case of humongous objects it is sufficient to scan the encompassing
881 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
882 // two areas will be zero sized. I.e. TAMS is either
883 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
884 // value: this would mean that TAMS points somewhere into the object.
885 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
886 "More than one object in the humongous region?");
887 humongous_obj->oop_iterate(&_update_cl, mr);
888 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
889 } else {
890 return 0;
891 }
892 }
893
894 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
895 oop obj = it.next();
896 size_t scanned_size = scan_for_references(obj, mr);
897 if ((HeapWord*)obj < top_at_mark_start) {
898 marked_words += scanned_size;
899 }
900 }
901
902 return marked_words * HeapWordSize;
903 }
904 public:
905 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
906 G1ConcurrentMark* cm,
907 uint worker_id) :
908 HeapRegionClosure(),
909 _cm(cm),
910 _update_cl(g1h, worker_id) { }
911
912 bool do_heap_region(HeapRegion* hr) {
913 if (_cm->has_aborted()) {
914 return true;
915 }
916
917 uint const region_idx = hr->hrm_index();
918 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
919 assert(top_at_rebuild_start_check == NULL ||
920 top_at_rebuild_start_check > hr->bottom(),
921 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
922 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
923
924 size_t total_marked_bytes = 0;
925 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
926
927 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
928
929 HeapWord* cur = hr->bottom();
930 while (cur < hr->end()) {
931 // After every iteration (yield point) we need to check whether the region's
932 // TARS changed due to e.g. eager reclaim.
933 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
934 if (top_at_rebuild_start == NULL) {
935 return false;
936 }
937
938 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
939 if (next_chunk.is_empty()) {
940 break;
941 }
942
943 const Ticks start = Ticks::now();
944 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
945 top_at_mark_start,
946 top_at_rebuild_start,
947 hr,
948 next_chunk);
949 Tickspan time = Ticks::now() - start;
950
951 log_trace(gc, remset, tracking)("Rebuilt region %u "
952 "live " SIZE_FORMAT " "
953 "time %.3fms "
954 "marked bytes " SIZE_FORMAT " "
955 "bot " PTR_FORMAT " "
956 "TAMS " PTR_FORMAT " "
957 "TARS " PTR_FORMAT,
958 region_idx,
959 _cm->liveness(region_idx) * HeapWordSize,
960 time.seconds() * 1000.0,
961 marked_bytes,
962 p2i(hr->bottom()),
963 p2i(top_at_mark_start),
964 p2i(top_at_rebuild_start));
965
966 if (marked_bytes > 0) {
967 total_marked_bytes += marked_bytes;
968 }
969 cur += chunk_size_in_words;
970
971 _cm->do_yield_check();
972 if (_cm->has_aborted()) {
973 return true;
974 }
975 }
976 // In the final iteration of the loop the region might have been eagerly reclaimed.
977 // Simply filter out those regions. We can not just use region type because there
978 // might have already been new allocations into these regions.
979 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
980 assert(top_at_rebuild_start == NULL ||
981 total_marked_bytes == hr->marked_bytes(),
982 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
983 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
984 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
985 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
986 // Abort state may have changed after the yield check.
987 return _cm->has_aborted();
988 }
989 };
990
991 HeapRegionClaimer _hr_claimer;
992 G1ConcurrentMark* _cm;
993
994 uint _worker_id_offset;
995 public:
996 G1RebuildRemSetTask(G1ConcurrentMark* cm,
997 uint n_workers,
998 uint worker_id_offset) :
999 AbstractGangTask("G1 Rebuild Remembered Set"),
1000 _hr_claimer(n_workers),
1001 _cm(cm),
1002 _worker_id_offset(worker_id_offset) {
1003 }
1004
1005 void work(uint worker_id) {
1006 SuspendibleThreadSetJoiner sts_join;
1007
1008 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1009
1010 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1011 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1012 }
1013 };
1014
1015 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1016 WorkGang* workers,
1017 uint worker_id_offset) {
1018 uint num_workers = workers->active_workers();
1019
1020 G1RebuildRemSetTask cl(cm,
1021 num_workers,
1022 worker_id_offset);
1023 workers->run_task(&cl, num_workers);
1024 }