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 "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/intHisto.hpp"
52 #include "utilities/stack.inline.hpp"
53 #include "utilities/ticks.hpp"
54
55 // Collects information about the overall remembered set scan progress during an evacuation.
56 class G1RemSetScanState : public CHeapObj<mtGC> {
57 private:
58 class G1ClearCardTableTask : public AbstractGangTask {
59 G1CollectedHeap* _g1h;
60 uint* _dirty_region_list;
61 size_t _num_dirty_regions;
62 size_t _chunk_length;
63
64 size_t volatile _cur_dirty_regions;
65 public:
66 G1ClearCardTableTask(G1CollectedHeap* g1h,
67 uint* dirty_region_list,
68 size_t num_dirty_regions,
69 size_t chunk_length) :
70 AbstractGangTask("G1 Clear Card Table Task"),
71 _g1h(g1h),
72 _dirty_region_list(dirty_region_list),
73 _num_dirty_regions(num_dirty_regions),
74 _chunk_length(chunk_length),
75 _cur_dirty_regions(0) {
76
77 assert(chunk_length > 0, "must be");
78 }
79
80 static size_t chunk_size() { return M; }
81
82 void work(uint worker_id) {
83 while (_cur_dirty_regions < _num_dirty_regions) {
84 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
85 size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
86
87 for (size_t i = next; i < max; i++) {
88 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
89 if (!r->is_survivor()) {
90 r->clear_cardtable();
91 }
92 }
93 }
94 }
95 };
96
97 size_t _max_regions;
98
99 // Scan progress for the remembered set of a single region. Transitions from
100 // Unclaimed -> Claimed -> Complete.
101 // At each of the transitions the thread that does the transition needs to perform
102 // some special action once. This is the reason for the extra "Claimed" state.
103 typedef jint G1RemsetIterState;
104
105 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
106 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned.
107 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned.
108
109 G1RemsetIterState volatile* _iter_states;
110 // The current location where the next thread should continue scanning in a region's
111 // remembered set.
112 size_t volatile* _iter_claims;
113
114 // Temporary buffer holding the regions we used to store remembered set scan duplicate
115 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
116 uint* _dirty_region_buffer;
117
118 typedef jbyte IsDirtyRegionState;
119 static const IsDirtyRegionState Clean = 0;
120 static const IsDirtyRegionState Dirty = 1;
121 // Holds a flag for every region whether it is in the _dirty_region_buffer already
122 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools.
123 IsDirtyRegionState* _in_dirty_region_buffer;
124 size_t _cur_dirty_region;
125
126 // Creates a snapshot of the current _top values at the start of collection to
127 // filter out card marks that we do not want to scan.
128 class G1ResetScanTopClosure : public HeapRegionClosure {
129 private:
130 HeapWord** _scan_top;
131 public:
132 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
133
134 virtual bool do_heap_region(HeapRegion* r) {
135 uint hrm_index = r->hrm_index();
136 if (!r->in_collection_set() && r->is_old_or_humongous_or_archive() && !r->is_empty()) {
137 _scan_top[hrm_index] = r->top();
138 } else {
139 _scan_top[hrm_index] = NULL;
140 }
141 return false;
142 }
143 };
144
145 // For each region, contains the maximum top() value to be used during this garbage
146 // collection. Subsumes common checks like filtering out everything but old and
147 // humongous regions outside the collection set.
148 // This is valid because we are not interested in scanning stray remembered set
149 // entries from free or archive regions.
150 HeapWord** _scan_top;
151 public:
152 G1RemSetScanState() :
153 _max_regions(0),
154 _iter_states(NULL),
155 _iter_claims(NULL),
156 _dirty_region_buffer(NULL),
157 _in_dirty_region_buffer(NULL),
158 _cur_dirty_region(0),
159 _scan_top(NULL) {
160 }
161
162 ~G1RemSetScanState() {
163 if (_iter_states != NULL) {
164 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
165 }
166 if (_iter_claims != NULL) {
167 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
168 }
169 if (_dirty_region_buffer != NULL) {
170 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
171 }
172 if (_in_dirty_region_buffer != NULL) {
173 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
174 }
175 if (_scan_top != NULL) {
176 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
177 }
178 }
179
180 void initialize(uint max_regions) {
181 assert(_iter_states == NULL, "Must not be initialized twice");
182 assert(_iter_claims == NULL, "Must not be initialized twice");
183 _max_regions = max_regions;
184 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
185 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
186 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
187 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
188 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
189 }
190
191 void reset() {
192 for (uint i = 0; i < _max_regions; i++) {
193 _iter_states[i] = Unclaimed;
194 _scan_top[i] = NULL;
195 }
196
197 G1ResetScanTopClosure cl(_scan_top);
198 G1CollectedHeap::heap()->heap_region_iterate(&cl);
199
200 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
201 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState));
202 _cur_dirty_region = 0;
203 }
204
205 // Attempt to claim the remembered set of the region for iteration. Returns true
206 // if this call caused the transition from Unclaimed to Claimed.
207 inline bool claim_iter(uint region) {
208 assert(region < _max_regions, "Tried to access invalid region %u", region);
209 if (_iter_states[region] != Unclaimed) {
210 return false;
211 }
212 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
213 return (res == Unclaimed);
214 }
215
216 // Try to atomically sets the iteration state to "complete". Returns true for the
217 // thread that caused the transition.
218 inline bool set_iter_complete(uint region) {
219 if (iter_is_complete(region)) {
220 return false;
221 }
222 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
223 return (res == Claimed);
224 }
225
226 // Returns true if the region's iteration is complete.
227 inline bool iter_is_complete(uint region) const {
228 assert(region < _max_regions, "Tried to access invalid region %u", region);
229 return _iter_states[region] == Complete;
230 }
231
232 // The current position within the remembered set of the given region.
233 inline size_t iter_claimed(uint region) const {
234 assert(region < _max_regions, "Tried to access invalid region %u", region);
235 return _iter_claims[region];
236 }
237
238 // Claim the next block of cards within the remembered set of the region with
239 // step size.
240 inline size_t iter_claimed_next(uint region, size_t step) {
241 return Atomic::add(step, &_iter_claims[region]) - step;
242 }
243
244 void add_dirty_region(uint region) {
245 if (_in_dirty_region_buffer[region] == Dirty) {
246 return;
247 }
248
249 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean;
250 if (marked_as_dirty) {
251 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
252 _dirty_region_buffer[allocated] = region;
253 }
254 }
255
256 HeapWord* scan_top(uint region_idx) const {
257 return _scan_top[region_idx];
258 }
259
260 // Clear the card table of "dirty" regions.
261 void clear_card_table(WorkGang* workers) {
262 if (_cur_dirty_region == 0) {
263 return;
264 }
265
266 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
267 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
268 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
269
270 // Iterate over the dirty cards region list.
271 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
272
273 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
274 "units of work for " SIZE_FORMAT " regions.",
275 cl.name(), num_workers, num_chunks, _cur_dirty_region);
276 workers->run_task(&cl, num_workers);
277
278 #ifndef PRODUCT
279 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
280 #endif
281 }
282 };
283
284 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
285 G1CardTable* ct,
286 G1HotCardCache* hot_card_cache) :
287 _scan_state(new G1RemSetScanState()),
288 _prev_period_summary(),
289 _g1h(g1h),
290 _num_conc_refined_cards(0),
291 _ct(ct),
292 _g1p(_g1h->g1_policy()),
293 _hot_card_cache(hot_card_cache) {
294 }
295
296 G1RemSet::~G1RemSet() {
297 if (_scan_state != NULL) {
298 delete _scan_state;
299 }
300 }
301
302 uint G1RemSet::num_par_rem_sets() {
303 return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
304 }
305
306 void G1RemSet::initialize(size_t capacity, uint max_regions) {
307 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
308 _scan_state->initialize(max_regions);
309 }
310
311 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
312 G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
313 G1ParScanThreadState* pss,
314 uint worker_i) :
315 _g1h(G1CollectedHeap::heap()),
316 _ct(_g1h->card_table()),
317 _pss(pss),
318 _scan_objs_on_card_cl(scan_obj_on_card),
319 _scan_state(scan_state),
320 _worker_i(worker_i),
321 _cards_scanned(0),
322 _cards_claimed(0),
323 _cards_skipped(0),
324 _rem_set_root_scan_time(),
325 _rem_set_trim_partially_time(),
326 _strong_code_root_scan_time(),
327 _strong_code_trim_partially_time() {
328 }
329
330 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
331 _ct->set_card_claimed(card_index);
332 _scan_state->add_dirty_region(region_idx_for_card);
333 }
334
335 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
336 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
337 assert(!card_region->is_young(), "Should not scan card in young region %u", region_idx_for_card);
338 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
339 _scan_objs_on_card_cl->trim_queue_partially();
340 _cards_scanned++;
341 }
342
343 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
344 EventGCPhaseParallel event;
345 uint const region_idx = r->hrm_index();
346
347 if (_scan_state->claim_iter(region_idx)) {
348 // If we ever free the collection set concurrently, we should also
349 // clear the card table concurrently therefore we won't need to
350 // add regions of the collection set to the dirty cards region.
351 _scan_state->add_dirty_region(region_idx);
352 }
353
354 if (r->rem_set()->cardset_is_empty()) {
355 return;
356 }
357
358 // We claim cards in blocks so as to reduce the contention.
359 size_t const block_size = G1RSetScanBlockSize;
360
361 HeapRegionRemSetIterator iter(r->rem_set());
362 size_t card_index;
363
364 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
365 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
366 if (current_card >= claimed_card_block + block_size) {
367 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
368 }
369 if (current_card < claimed_card_block) {
370 _cards_skipped++;
371 continue;
372 }
373 _cards_claimed++;
374
375 HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
376 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
377
378 #ifdef ASSERT
379 HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
380 assert(hr == NULL || hr->is_in_reserved(card_start),
381 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
382 #endif
383 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
384 if (card_start >= top) {
385 continue;
386 }
387
388 // If the card is dirty, then G1 will scan it during Update RS.
389 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
390 continue;
391 }
392
393 // We claim lazily (so races are possible but they're benign), which reduces the
394 // number of duplicate scans (the rsets of the regions in the cset can intersect).
395 // Claim the card after checking bounds above: the remembered set may contain
396 // random cards into current survivor, and we would then have an incorrectly
397 // claimed card in survivor space. Card table clear does not reset the card table
398 // of survivor space regions.
399 claim_card(card_index, region_idx_for_card);
400
401 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
402
403 scan_card(mr, region_idx_for_card);
404 }
405 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::ScanRS));
406 }
407
408 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
409 EventGCPhaseParallel event;
410 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
411 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
412 }
413
414 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
415 assert(r->in_collection_set(),
416 "Should only be called on elements of the collection set but region %u is not.",
417 r->hrm_index());
418 uint const region_idx = r->hrm_index();
419
420 // Do an early out if we know we are complete.
421 if (_scan_state->iter_is_complete(region_idx)) {
422 return false;
423 }
424
425 {
426 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
427 scan_rem_set_roots(r);
428 }
429
430 if (_scan_state->set_iter_complete(region_idx)) {
431 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
432 // Scan the strong code root list attached to the current region
433 scan_strong_code_roots(r);
434 }
435 return false;
436 }
437
438 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
439 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
440 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i);
441 _g1h->collection_set_iterate_from(&cl, worker_i);
442
443 G1GCPhaseTimes* p = _g1p->phase_times();
444
445 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
446 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
447
448 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
449 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
450 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
451
452 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
453 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
454 }
455
456 // Closure used for updating rem sets. Only called during an evacuation pause.
457 class G1RefineCardClosure: public CardTableEntryClosure {
458 G1RemSet* _g1rs;
459 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
460
461 size_t _cards_scanned;
462 size_t _cards_skipped;
463 public:
464 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
465 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
466 {}
467
468 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
469 // The only time we care about recording cards that
470 // contain references that point into the collection set
471 // is during RSet updating within an evacuation pause.
472 // In this case worker_i should be the id of a GC worker thread.
473 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
474
475 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
476
477 if (card_scanned) {
478 _update_rs_cl->trim_queue_partially();
479 _cards_scanned++;
480 } else {
481 _cards_skipped++;
482 }
483 return true;
484 }
485
486 size_t cards_scanned() const { return _cards_scanned; }
487 size_t cards_skipped() const { return _cards_skipped; }
488 };
489
490 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
491 G1GCPhaseTimes* p = _g1p->phase_times();
492
493 // Apply closure to log entries in the HCC.
494 if (G1HotCardCache::default_use_cache()) {
495 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
496
497 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss);
498 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
499 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
500 }
501
502 // Now apply the closure to all remaining log entries.
503 {
504 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
505
506 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss);
507 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
508 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
509
510 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
511 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
512 }
513 }
514
515 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
516 update_rem_set(pss, worker_i);
517 scan_rem_set(pss, worker_i);;
518 }
519
520 void G1RemSet::prepare_for_oops_into_collection_set_do() {
521 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
522 dcqs.concatenate_logs();
523
524 _scan_state->reset();
525 }
526
527 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
528 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times();
529
530 // Set all cards back to clean.
531 double start = os::elapsedTime();
532 _scan_state->clear_card_table(_g1h->workers());
533 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
534 }
535
536 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
537 #ifdef ASSERT
538 G1CollectedHeap* g1h = G1CollectedHeap::heap();
539 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
540 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
541 p2i(card_ptr),
542 ct->index_for(ct->addr_for(card_ptr)),
543 p2i(ct->addr_for(card_ptr)),
544 g1h->addr_to_region(ct->addr_for(card_ptr)));
545 #endif
546 }
547
548 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
549 uint worker_i) {
550 assert(!_g1h->is_gc_active(), "Only call concurrently");
551
552 // Construct the region representing the card.
553 HeapWord* start = _ct->addr_for(card_ptr);
554 // And find the region containing it.
555 HeapRegion* r = _g1h->heap_region_containing_or_null(start);
556
557 // If this is a (stale) card into an uncommitted region, exit.
558 if (r == NULL) {
559 return;
560 }
561
562 check_card_ptr(card_ptr, _ct);
563
564 // If the card is no longer dirty, nothing to do.
565 if (*card_ptr != G1CardTable::dirty_card_val()) {
566 return;
567 }
568
569 // This check is needed for some uncommon cases where we should
570 // ignore the card.
571 //
572 // The region could be young. Cards for young regions are
573 // distinctly marked (set to g1_young_gen), so the post-barrier will
574 // filter them out. However, that marking is performed
575 // concurrently. A write to a young object could occur before the
576 // card has been marked young, slipping past the filter.
577 //
578 // The card could be stale, because the region has been freed since
579 // the card was recorded. In this case the region type could be
580 // anything. If (still) free or (reallocated) young, just ignore
581 // it. If (reallocated) old or humongous, the later card trimming
582 // and additional checks in iteration may detect staleness. At
583 // worst, we end up processing a stale card unnecessarily.
584 //
585 // In the normal (non-stale) case, the synchronization between the
586 // enqueueing of the card and processing it here will have ensured
587 // we see the up-to-date region type here.
588 if (!r->is_old_or_humongous_or_archive()) {
589 return;
590 }
591
592 // The result from the hot card cache insert call is either:
593 // * pointer to the current card
594 // (implying that the current card is not 'hot'),
595 // * null
596 // (meaning we had inserted the card ptr into the "hot" card cache,
597 // which had some headroom),
598 // * a pointer to a "hot" card that was evicted from the "hot" cache.
599 //
600
601 if (_hot_card_cache->use_cache()) {
602 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
603
604 const jbyte* orig_card_ptr = card_ptr;
605 card_ptr = _hot_card_cache->insert(card_ptr);
606 if (card_ptr == NULL) {
607 // There was no eviction. Nothing to do.
608 return;
609 } else if (card_ptr != orig_card_ptr) {
610 // Original card was inserted and an old card was evicted.
611 start = _ct->addr_for(card_ptr);
612 r = _g1h->heap_region_containing(start);
613
614 // Check whether the region formerly in the cache should be
615 // ignored, as discussed earlier for the original card. The
616 // region could have been freed while in the cache.
617 if (!r->is_old_or_humongous_or_archive()) {
618 return;
619 }
620 } // Else we still have the original card.
621 }
622
623 // Trim the region designated by the card to what's been allocated
624 // in the region. The card could be stale, or the card could cover
625 // (part of) an object at the end of the allocated space and extend
626 // beyond the end of allocation.
627
628 // Non-humongous objects are only allocated in the old-gen during
629 // GC, so if region is old then top is stable. Humongous object
630 // allocation sets top last; if top has not yet been set, this is
631 // a stale card and we'll end up with an empty intersection. If
632 // this is not a stale card, the synchronization between the
633 // enqueuing of the card and processing it here will have ensured
634 // we see the up-to-date top here.
635 HeapWord* scan_limit = r->top();
636
637 if (scan_limit <= start) {
638 // If the trimmed region is empty, the card must be stale.
639 return;
640 }
641
642 // Okay to clean and process the card now. There are still some
643 // stale card cases that may be detected by iteration and dealt with
644 // as iteration failure.
645 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
646
647 // This fence serves two purposes. First, the card must be cleaned
648 // before processing the contents. Second, we can't proceed with
649 // processing until after the read of top, for synchronization with
650 // possibly concurrent humongous object allocation. It's okay that
651 // reading top and reading type were racy wrto each other. We need
652 // both set, in any order, to proceed.
653 OrderAccess::fence();
654
655 // Don't use addr_for(card_ptr + 1) which can ask for
656 // a card beyond the heap.
657 HeapWord* end = start + G1CardTable::card_size_in_words;
658 MemRegion dirty_region(start, MIN2(scan_limit, end));
659 assert(!dirty_region.is_empty(), "sanity");
660
661 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
662
663 bool card_processed =
664 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
665
666 // If unable to process the card then we encountered an unparsable
667 // part of the heap (e.g. a partially allocated object) while
668 // processing a stale card. Despite the card being stale, redirty
669 // and re-enqueue, because we've already cleaned the card. Without
670 // this we could incorrectly discard a non-stale card.
671 if (!card_processed) {
672 // The card might have gotten re-dirtied and re-enqueued while we
673 // worked. (In fact, it's pretty likely.)
674 if (*card_ptr != G1CardTable::dirty_card_val()) {
675 *card_ptr = G1CardTable::dirty_card_val();
676 MutexLockerEx x(Shared_DirtyCardQ_lock,
677 Mutex::_no_safepoint_check_flag);
678 DirtyCardQueue* sdcq =
679 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
680 sdcq->enqueue(card_ptr);
681 }
682 } else {
683 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
684 }
685 }
686
687 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
688 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
689 assert(_g1h->is_gc_active(), "Only call during GC");
690
691 // Construct the region representing the card.
692 HeapWord* card_start = _ct->addr_for(card_ptr);
693 // And find the region containing it.
694 uint const card_region_idx = _g1h->addr_to_region(card_start);
695
696 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
697 if (scan_limit == NULL) {
698 // This is a card into an uncommitted region. We need to bail out early as we
699 // should not access the corresponding card table entry.
700 return false;
701 }
702
703 check_card_ptr(card_ptr, _ct);
704
705 // If the card is no longer dirty, nothing to do. This covers cards that were already
706 // scanned as parts of the remembered sets.
707 if (*card_ptr != G1CardTable::dirty_card_val()) {
708 return false;
709 }
710
711 // We claim lazily (so races are possible but they're benign), which reduces the
712 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
713 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
714
715 _scan_state->add_dirty_region(card_region_idx);
716 if (scan_limit <= card_start) {
717 // If the card starts above the area in the region containing objects to scan, skip it.
718 return false;
719 }
720
721 // Don't use addr_for(card_ptr + 1) which can ask for
722 // a card beyond the heap.
723 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
724 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
725 assert(!dirty_region.is_empty(), "sanity");
726
727 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
728 assert(!card_region->is_young(), "Should not scan card in young region %u", card_region_idx);
729 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
730 assert(card_processed, "must be");
731 return true;
732 }
733
734 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
735 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
736 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
737
738 G1RemSetSummary current(this);
739 _prev_period_summary.subtract_from(¤t);
740
741 Log(gc, remset) log;
742 log.trace("%s", header);
743 ResourceMark rm;
744 LogStream ls(log.trace());
745 _prev_period_summary.print_on(&ls);
746
747 _prev_period_summary.set(¤t);
748 }
749 }
750
751 void G1RemSet::print_summary_info() {
752 Log(gc, remset, exit) log;
753 if (log.is_trace()) {
754 log.trace(" Cumulative RS summary");
755 G1RemSetSummary current(this);
756 ResourceMark rm;
757 LogStream ls(log.trace());
758 current.print_on(&ls);
759 }
760 }
761
762 class G1RebuildRemSetTask: public AbstractGangTask {
763 // Aggregate the counting data that was constructed concurrently
764 // with marking.
765 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
766 G1ConcurrentMark* _cm;
767 G1RebuildRemSetClosure _update_cl;
768
769 // Applies _update_cl to the references of the given object, limiting objArrays
770 // to the given MemRegion. Returns the amount of words actually scanned.
771 size_t scan_for_references(oop const obj, MemRegion mr) {
772 size_t const obj_size = obj->size();
773 // All non-objArrays and objArrays completely within the mr
774 // can be scanned without passing the mr.
775 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
776 obj->oop_iterate(&_update_cl);
777 return obj_size;
778 }
779 // This path is for objArrays crossing the given MemRegion. Only scan the
780 // area within the MemRegion.
781 obj->oop_iterate(&_update_cl, mr);
782 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
783 }
784
785 // A humongous object is live (with respect to the scanning) either
786 // a) it is marked on the bitmap as such
787 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
788 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
789 return bitmap->is_marked(humongous_obj) || (tars > tams);
790 }
791
792 // Iterator over the live objects within the given MemRegion.
793 class LiveObjIterator : public StackObj {
794 const G1CMBitMap* const _bitmap;
795 const HeapWord* _tams;
796 const MemRegion _mr;
797 HeapWord* _current;
798
799 bool is_below_tams() const {
800 return _current < _tams;
801 }
802
803 bool is_live(HeapWord* obj) const {
804 return !is_below_tams() || _bitmap->is_marked(obj);
805 }
806
807 HeapWord* bitmap_limit() const {
808 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
809 }
810
811 void move_if_below_tams() {
812 if (is_below_tams() && has_next()) {
813 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
814 }
815 }
816 public:
817 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
818 _bitmap(bitmap),
819 _tams(tams),
820 _mr(mr),
821 _current(first_oop_into_mr) {
822
823 assert(_current <= _mr.start(),
824 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
825 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
826
827 // Step to the next live object within the MemRegion if needed.
828 if (is_live(_current)) {
829 // Non-objArrays were scanned by the previous part of that region.
830 if (_current < mr.start() && !oop(_current)->is_objArray()) {
831 _current += oop(_current)->size();
832 // We might have positioned _current on a non-live object. Reposition to the next
833 // live one if needed.
834 move_if_below_tams();
835 }
836 } else {
837 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
838 // immediately.
839 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
840 assert(_current == _mr.end() || is_live(_current),
841 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
842 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
843 }
844 }
845
846 void move_to_next() {
847 _current += next()->size();
848 move_if_below_tams();
849 }
850
851 oop next() const {
852 oop result = oop(_current);
853 assert(is_live(_current),
854 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
855 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
856 return result;
857 }
858
859 bool has_next() const {
860 return _current < _mr.end();
861 }
862 };
863
864 // Rebuild remembered sets in the part of the region specified by mr and hr.
865 // Objects between the bottom of the region and the TAMS are checked for liveness
866 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
867 // Returns the number of live words between bottom and TAMS.
868 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
869 HeapWord* const top_at_mark_start,
870 HeapWord* const top_at_rebuild_start,
871 HeapRegion* hr,
872 MemRegion mr) {
873 size_t marked_words = 0;
874
875 if (hr->is_humongous()) {
876 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
877 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
878 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
879 // however in case of humongous objects it is sufficient to scan the encompassing
880 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
881 // two areas will be zero sized. I.e. TAMS is either
882 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
883 // value: this would mean that TAMS points somewhere into the object.
884 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
885 "More than one object in the humongous region?");
886 humongous_obj->oop_iterate(&_update_cl, mr);
887 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
888 } else {
889 return 0;
890 }
891 }
892
893 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
894 oop obj = it.next();
895 size_t scanned_size = scan_for_references(obj, mr);
896 if ((HeapWord*)obj < top_at_mark_start) {
897 marked_words += scanned_size;
898 }
899 }
900
901 return marked_words * HeapWordSize;
902 }
903 public:
904 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
905 G1ConcurrentMark* cm,
906 uint worker_id) :
907 HeapRegionClosure(),
908 _cm(cm),
909 _update_cl(g1h, worker_id) { }
910
911 bool do_heap_region(HeapRegion* hr) {
912 if (_cm->has_aborted()) {
913 return true;
914 }
915
916 uint const region_idx = hr->hrm_index();
917 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
918 assert(top_at_rebuild_start_check == NULL ||
919 top_at_rebuild_start_check > hr->bottom(),
920 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
921 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
922
923 size_t total_marked_bytes = 0;
924 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
925
926 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
927
928 HeapWord* cur = hr->bottom();
929 while (cur < hr->end()) {
930 // After every iteration (yield point) we need to check whether the region's
931 // TARS changed due to e.g. eager reclaim.
932 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
933 if (top_at_rebuild_start == NULL) {
934 return false;
935 }
936
937 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
938 if (next_chunk.is_empty()) {
939 break;
940 }
941
942 const Ticks start = Ticks::now();
943 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
944 top_at_mark_start,
945 top_at_rebuild_start,
946 hr,
947 next_chunk);
948 Tickspan time = Ticks::now() - start;
949
950 log_trace(gc, remset, tracking)("Rebuilt region %u "
951 "live " SIZE_FORMAT " "
952 "time %.3fms "
953 "marked bytes " SIZE_FORMAT " "
954 "bot " PTR_FORMAT " "
955 "TAMS " PTR_FORMAT " "
956 "TARS " PTR_FORMAT,
957 region_idx,
958 _cm->liveness(region_idx) * HeapWordSize,
959 time.seconds() * 1000.0,
960 marked_bytes,
961 p2i(hr->bottom()),
962 p2i(top_at_mark_start),
963 p2i(top_at_rebuild_start));
964
965 if (marked_bytes > 0) {
966 total_marked_bytes += marked_bytes;
967 }
968 cur += chunk_size_in_words;
969
970 _cm->do_yield_check();
971 if (_cm->has_aborted()) {
972 return true;
973 }
974 }
975 // In the final iteration of the loop the region might have been eagerly reclaimed.
976 // Simply filter out those regions. We can not just use region type because there
977 // might have already been new allocations into these regions.
978 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
979 assert(top_at_rebuild_start == NULL ||
980 total_marked_bytes == hr->marked_bytes(),
981 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
982 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
983 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
984 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
985 // Abort state may have changed after the yield check.
986 return _cm->has_aborted();
987 }
988 };
989
990 HeapRegionClaimer _hr_claimer;
991 G1ConcurrentMark* _cm;
992
993 uint _worker_id_offset;
994 public:
995 G1RebuildRemSetTask(G1ConcurrentMark* cm,
996 uint n_workers,
997 uint worker_id_offset) :
998 AbstractGangTask("G1 Rebuild Remembered Set"),
999 _hr_claimer(n_workers),
1000 _cm(cm),
1001 _worker_id_offset(worker_id_offset) {
1002 }
1003
1004 void work(uint worker_id) {
1005 SuspendibleThreadSetJoiner sts_join;
1006
1007 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1008
1009 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1010 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1011 }
1012 };
1013
1014 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1015 WorkGang* workers,
1016 uint worker_id_offset) {
1017 uint num_workers = workers->active_workers();
1018
1019 G1RebuildRemSetTask cl(cm,
1020 num_workers,
1021 worker_id_offset);
1022 workers->run_task(&cl, num_workers);
1023 }