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/g1CardTableEntryClosure.hpp"
30 #include "gc/g1/g1CollectedHeap.inline.hpp"
31 #include "gc/g1/g1ConcurrentRefine.hpp"
32 #include "gc/g1/g1DirtyCardQueue.hpp"
33 #include "gc/g1/g1FromCardCache.hpp"
34 #include "gc/g1/g1GCPhaseTimes.hpp"
35 #include "gc/g1/g1HotCardCache.hpp"
36 #include "gc/g1/g1OopClosures.inline.hpp"
37 #include "gc/g1/g1RootClosures.hpp"
38 #include "gc/g1/g1RemSet.hpp"
39 #include "gc/g1/g1SharedDirtyCardQueue.hpp"
40 #include "gc/g1/heapRegion.inline.hpp"
41 #include "gc/g1/heapRegionManager.inline.hpp"
42 #include "gc/g1/heapRegionRemSet.inline.hpp"
43 #include "gc/g1/sparsePRT.hpp"
44 #include "gc/shared/gcTraceTime.inline.hpp"
45 #include "gc/shared/ptrQueue.hpp"
46 #include "gc/shared/suspendibleThreadSet.hpp"
47 #include "jfr/jfrEvents.hpp"
48 #include "memory/iterator.hpp"
49 #include "memory/resourceArea.hpp"
50 #include "oops/access.inline.hpp"
51 #include "oops/oop.inline.hpp"
52 #include "runtime/os.hpp"
53 #include "utilities/align.hpp"
54 #include "utilities/globalDefinitions.hpp"
55 #include "utilities/stack.inline.hpp"
56 #include "utilities/ticks.hpp"
57
58 // Collects information about the overall heap root scan progress during an evacuation.
59 //
60 // Scanning the remembered sets works by first merging all sources of cards to be
61 // scanned (log buffers, hcc, remembered sets) into a single data structure to remove
62 // duplicates and simplify work distribution.
63 //
64 // During the following card scanning we not only scan this combined set of cards, but
65 // also remember that these were completely scanned. The following evacuation passes
66 // do not scan these cards again, and so need to be preserved across increments.
67 //
68 // The representation for all the cards to scan is the card table: cards can have
69 // one of three states during GC:
70 // - clean: these cards will not be scanned in this pass
71 // - dirty: these cards will be scanned in this pass
72 // - scanned: these cards have already been scanned in a previous pass
73 //
74 // After all evacuation is done, we reset the card table to clean.
75 //
76 // Work distribution occurs on "chunk" basis, i.e. contiguous ranges of cards. As an
77 // additional optimization, during card merging we remember which regions and which
78 // chunks actually contain cards to be scanned. Threads iterate only across these
79 // regions, and only compete for chunks containing any cards.
80 //
81 // Within these chunks, a worker scans the card table on "blocks" of cards, i.e.
82 // contiguous ranges of dirty cards to be scanned. These blocks are converted to actual
83 // memory ranges and then passed on to actual scanning.
84 class G1RemSetScanState : public CHeapObj<mtGC> {
85 class G1DirtyRegions;
86
87 size_t _max_regions;
88
89 // Has this region that is part of the regions in the collection set been processed yet.
90 typedef bool G1RemsetIterState;
91
92 G1RemsetIterState volatile* _collection_set_iter_state;
93
94 // Card table iteration claim for each heap region, from 0 (completely unscanned)
95 // to (>=) HeapRegion::CardsPerRegion (completely scanned).
96 uint volatile* _card_table_scan_state;
97
98 // Return "optimal" number of chunks per region we want to use for claiming areas
99 // within a region to claim. Dependent on the region size as proxy for the heap
100 // size, we limit the total number of chunks to limit memory usage and maintenance
101 // effort of that table vs. granularity of distributing scanning work.
102 // Testing showed that 8 for 1M/2M region, 16 for 4M/8M regions, 32 for 16/32M regions
103 // seems to be such a good trade-off.
104 static uint get_chunks_per_region(uint log_region_size) {
105 // Limit the expected input values to current known possible values of the
106 // (log) region size. Adjust as necessary after testing if changing the permissible
107 // values for region size.
108 assert(log_region_size >= 20 && log_region_size <= 25,
109 "expected value in [20,25], but got %u", log_region_size);
110 return 1u << (log_region_size / 2 - 7);
111 }
112
113 uint _scan_chunks_per_region; // Number of chunks per region.
114 uint8_t _log_scan_chunks_per_region; // Log of number of chunks per region.
115 bool* _region_scan_chunks;
116 size_t _num_total_scan_chunks; // Total number of elements in _region_scan_chunks.
117 uint8_t _scan_chunks_shift; // For conversion between card index and chunk index.
118 public:
119 uint scan_chunk_size() const { return (uint)1 << _scan_chunks_shift; }
120
121 // Returns whether the chunk corresponding to the given region/card in region contain a
122 // dirty card, i.e. actually needs scanning.
123 bool chunk_needs_scan(uint const region_idx, uint const card_in_region) const {
124 size_t const idx = ((size_t)region_idx << _log_scan_chunks_per_region) + (card_in_region >> _scan_chunks_shift);
125 assert(idx < _num_total_scan_chunks, "Index " SIZE_FORMAT " out of bounds " SIZE_FORMAT,
126 idx, _num_total_scan_chunks);
127 return _region_scan_chunks[idx];
128 }
129
130 private:
131 // The complete set of regions which card table needs to be cleared at the end of GC because
132 // we scribbled all over them.
133 G1DirtyRegions* _all_dirty_regions;
134 // The set of regions which card table needs to be scanned for new dirty cards
135 // in the current evacuation pass.
136 G1DirtyRegions* _next_dirty_regions;
137
138 // Set of (unique) regions that can be added to concurrently.
139 class G1DirtyRegions : public CHeapObj<mtGC> {
140 uint* _buffer;
141 uint _cur_idx;
142 size_t _max_regions;
143
144 bool* _contains;
145
146 public:
147 G1DirtyRegions(size_t max_regions) :
148 _buffer(NEW_C_HEAP_ARRAY(uint, max_regions, mtGC)),
149 _cur_idx(0),
150 _max_regions(max_regions),
151 _contains(NEW_C_HEAP_ARRAY(bool, max_regions, mtGC)) {
152
153 reset();
154 }
155
156 static size_t chunk_size() { return M; }
157
158 ~G1DirtyRegions() {
159 FREE_C_HEAP_ARRAY(uint, _buffer);
160 FREE_C_HEAP_ARRAY(bool, _contains);
161 }
162
163 void reset() {
164 _cur_idx = 0;
165 ::memset(_contains, false, _max_regions * sizeof(bool));
166 }
167
168 uint size() const { return _cur_idx; }
169
170 uint at(uint idx) const {
171 assert(idx < _cur_idx, "Index %u beyond valid regions", idx);
172 return _buffer[idx];
173 }
174
175 void add_dirty_region(uint region) {
176 if (_contains[region]) {
177 return;
178 }
179
180 bool marked_as_dirty = Atomic::cmpxchg(true, &_contains[region], false) == false;
181 if (marked_as_dirty) {
182 uint allocated = Atomic::add(1u, &_cur_idx) - 1;
183 _buffer[allocated] = region;
184 }
185 }
186
187 // Creates the union of this and the other G1DirtyRegions.
188 void merge(const G1DirtyRegions* other) {
189 for (uint i = 0; i < other->size(); i++) {
190 uint region = other->at(i);
191 if (!_contains[region]) {
192 _buffer[_cur_idx++] = region;
193 _contains[region] = true;
194 }
195 }
196 }
197 };
198
199 // For each region, contains the maximum top() value to be used during this garbage
200 // collection. Subsumes common checks like filtering out everything but old and
201 // humongous regions outside the collection set.
202 // This is valid because we are not interested in scanning stray remembered set
203 // entries from free or archive regions.
204 HeapWord** _scan_top;
205
206 class G1ClearCardTableTask : public AbstractGangTask {
207 G1CollectedHeap* _g1h;
208 G1DirtyRegions* _regions;
209 uint _chunk_length;
210
211 uint volatile _cur_dirty_regions;
212
213 G1RemSetScanState* _scan_state;
214
215 public:
216 G1ClearCardTableTask(G1CollectedHeap* g1h,
217 G1DirtyRegions* regions,
218 uint chunk_length,
219 G1RemSetScanState* scan_state) :
220 AbstractGangTask("G1 Clear Card Table Task"),
221 _g1h(g1h),
222 _regions(regions),
223 _chunk_length(chunk_length),
224 _cur_dirty_regions(0),
225 _scan_state(scan_state) {
226
227 assert(chunk_length > 0, "must be");
228 }
229
230 static uint chunk_size() { return M; }
231
232 void work(uint worker_id) {
233 while (_cur_dirty_regions < _regions->size()) {
234 uint next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
235 uint max = MIN2(next + _chunk_length, _regions->size());
236
237 for (uint i = next; i < max; i++) {
238 HeapRegion* r = _g1h->region_at(_regions->at(i));
239 if (!r->is_survivor()) {
240 r->clear_cardtable();
241 }
242 }
243 }
244 }
245 };
246
247 // Clear the card table of "dirty" regions.
248 void clear_card_table(WorkGang* workers) {
249 uint num_regions = _all_dirty_regions->size();
250
251 if (num_regions == 0) {
252 return;
253 }
254
255 uint const num_chunks = (uint)(align_up((size_t)num_regions << HeapRegion::LogCardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size());
256 uint const num_workers = MIN2(num_chunks, workers->active_workers());
257 uint const chunk_length = G1ClearCardTableTask::chunk_size() / (uint)HeapRegion::CardsPerRegion;
258
259 // Iterate over the dirty cards region list.
260 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _all_dirty_regions, chunk_length, this);
261
262 log_debug(gc, ergo)("Running %s using %u workers for %u "
263 "units of work for %u regions.",
264 cl.name(), num_workers, num_chunks, num_regions);
265 workers->run_task(&cl, num_workers);
266
267 #ifndef PRODUCT
268 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
269 #endif
270 }
271
272 public:
273 G1RemSetScanState() :
274 _max_regions(0),
275 _collection_set_iter_state(NULL),
276 _card_table_scan_state(NULL),
277 _scan_chunks_per_region(get_chunks_per_region(HeapRegion::LogOfHRGrainBytes)),
278 _log_scan_chunks_per_region(log2_uint(_scan_chunks_per_region)),
279 _region_scan_chunks(NULL),
280 _num_total_scan_chunks(0),
281 _scan_chunks_shift(0),
282 _all_dirty_regions(NULL),
283 _next_dirty_regions(NULL),
284 _scan_top(NULL) {
285 }
286
287 ~G1RemSetScanState() {
288 FREE_C_HEAP_ARRAY(G1RemsetIterState, _collection_set_iter_state);
289 FREE_C_HEAP_ARRAY(uint, _card_table_scan_state);
290 FREE_C_HEAP_ARRAY(bool, _region_scan_chunks);
291 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
292 }
293
294 void initialize(size_t max_regions) {
295 assert(_collection_set_iter_state == NULL, "Must not be initialized twice");
296 _max_regions = max_regions;
297 _collection_set_iter_state = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
298 _card_table_scan_state = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
299 _num_total_scan_chunks = max_regions * _scan_chunks_per_region;
300 _region_scan_chunks = NEW_C_HEAP_ARRAY(bool, _num_total_scan_chunks, mtGC);
301
302 _scan_chunks_shift = (uint8_t)log2_intptr(HeapRegion::CardsPerRegion / _scan_chunks_per_region);
303 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
304 }
305
306 void prepare() {
307 _all_dirty_regions = new G1DirtyRegions(_max_regions);
308 _next_dirty_regions = new G1DirtyRegions(_max_regions);
309 }
310
311 void prepare_for_merge_heap_roots() {
312 _all_dirty_regions->merge(_next_dirty_regions);
313
314 _next_dirty_regions->reset();
315 for (size_t i = 0; i < _max_regions; i++) {
316 _card_table_scan_state[i] = 0;
317 }
318
319 ::memset(_region_scan_chunks, false, _num_total_scan_chunks * sizeof(*_region_scan_chunks));
320 }
321
322 // Returns whether the given region contains cards we need to scan. The remembered
323 // set and other sources may contain cards that
324 // - are in uncommitted regions
325 // - are located in the collection set
326 // - are located in free regions
327 // as we do not clean up remembered sets before merging heap roots.
328 bool contains_cards_to_process(uint const region_idx) const {
329 HeapRegion* hr = G1CollectedHeap::heap()->region_at_or_null(region_idx);
330 return (hr != NULL && !hr->in_collection_set() && hr->is_old_or_humongous_or_archive());
331 }
332
333 size_t num_visited_cards() const {
334 size_t result = 0;
335 for (uint i = 0; i < _num_total_scan_chunks; i++) {
336 if (_region_scan_chunks[i]) {
337 result++;
338 }
339 }
340 return result * (HeapRegion::CardsPerRegion / _scan_chunks_per_region);
341 }
342
343 size_t num_cards_in_dirty_regions() const {
344 return _next_dirty_regions->size() * HeapRegion::CardsPerRegion;
345 }
346
347 void set_chunk_region_dirty(size_t const region_card_idx) {
348 size_t chunk_idx = region_card_idx >> _scan_chunks_shift;
349 for (uint i = 0; i < _scan_chunks_per_region; i++) {
350 _region_scan_chunks[chunk_idx++] = true;
351 }
352 }
353
354 void set_chunk_dirty(size_t const card_idx) {
355 assert((card_idx >> _scan_chunks_shift) < _num_total_scan_chunks,
356 "Trying to access index " SIZE_FORMAT " out of bounds " SIZE_FORMAT,
357 card_idx >> _scan_chunks_shift, _num_total_scan_chunks);
358 size_t const chunk_idx = card_idx >> _scan_chunks_shift;
359 if (!_region_scan_chunks[chunk_idx]) {
360 _region_scan_chunks[chunk_idx] = true;
361 }
362 }
363
364 void cleanup(WorkGang* workers) {
365 _all_dirty_regions->merge(_next_dirty_regions);
366
367 clear_card_table(workers);
368
369 delete _all_dirty_regions;
370 _all_dirty_regions = NULL;
371
372 delete _next_dirty_regions;
373 _next_dirty_regions = NULL;
374 }
375
376 void iterate_dirty_regions_from(HeapRegionClosure* cl, uint worker_id) {
377 uint num_regions = _next_dirty_regions->size();
378
379 if (num_regions == 0) {
380 return;
381 }
382
383 G1CollectedHeap* g1h = G1CollectedHeap::heap();
384
385 WorkGang* workers = g1h->workers();
386 uint const max_workers = workers->active_workers();
387
388 uint const start_pos = num_regions * worker_id / max_workers;
389 uint cur = start_pos;
390
391 do {
392 bool result = cl->do_heap_region(g1h->region_at(_next_dirty_regions->at(cur)));
393 guarantee(!result, "Not allowed to ask for early termination.");
394 cur++;
395 if (cur == _next_dirty_regions->size()) {
396 cur = 0;
397 }
398 } while (cur != start_pos);
399 }
400
401 void reset_region_claim(uint region_idx) {
402 _collection_set_iter_state[region_idx] = false;
403 }
404
405 // Attempt to claim the given region in the collection set for iteration. Returns true
406 // if this call caused the transition from Unclaimed to Claimed.
407 inline bool claim_collection_set_region(uint region) {
408 assert(region < _max_regions, "Tried to access invalid region %u", region);
409 if (_collection_set_iter_state[region]) {
410 return false;
411 }
412 return !Atomic::cmpxchg(true, &_collection_set_iter_state[region], false);
413 }
414
415 bool has_cards_to_scan(uint region) {
416 assert(region < _max_regions, "Tried to access invalid region %u", region);
417 return _card_table_scan_state[region] < HeapRegion::CardsPerRegion;
418 }
419
420 uint claim_cards_to_scan(uint region, uint increment) {
421 assert(region < _max_regions, "Tried to access invalid region %u", region);
422 return Atomic::add(increment, &_card_table_scan_state[region]) - increment;
423 }
424
425 void add_dirty_region(uint const region) {
426 #ifdef ASSERT
427 HeapRegion* hr = G1CollectedHeap::heap()->region_at(region);
428 assert(!hr->in_collection_set() && hr->is_old_or_humongous_or_archive(),
429 "Region %u is not suitable for scanning, is %sin collection set or %s",
430 hr->hrm_index(), hr->in_collection_set() ? "" : "not ", hr->get_short_type_str());
431 #endif
432 _next_dirty_regions->add_dirty_region(region);
433 }
434
435 void add_all_dirty_region(uint region) {
436 #ifdef ASSERT
437 HeapRegion* hr = G1CollectedHeap::heap()->region_at(region);
438 assert(hr->in_collection_set(),
439 "Only add young regions to all dirty regions directly but %u is %s",
440 hr->hrm_index(), hr->get_short_type_str());
441 #endif
442 _all_dirty_regions->add_dirty_region(region);
443 }
444
445 void set_scan_top(uint region_idx, HeapWord* value) {
446 _scan_top[region_idx] = value;
447 }
448
449 HeapWord* scan_top(uint region_idx) const {
450 return _scan_top[region_idx];
451 }
452
453 void clear_scan_top(uint region_idx) {
454 set_scan_top(region_idx, NULL);
455 }
456 };
457
458 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
459 G1CardTable* ct,
460 G1HotCardCache* hot_card_cache) :
461 _scan_state(new G1RemSetScanState()),
462 _prev_period_summary(false),
463 _g1h(g1h),
464 _ct(ct),
465 _g1p(_g1h->policy()),
466 _hot_card_cache(hot_card_cache) {
467 }
468
469 G1RemSet::~G1RemSet() {
470 delete _scan_state;
471 }
472
473 uint G1RemSet::num_par_rem_sets() {
474 return G1DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
475 }
476
477 void G1RemSet::initialize(size_t capacity, uint max_regions) {
478 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
479 _scan_state->initialize(max_regions);
480 }
481
482 // Helper class to scan and detect ranges of cards that need to be scanned on the
483 // card table.
484 class G1CardTableScanner : public StackObj {
485 public:
486 typedef CardTable::CardValue CardValue;
487
488 private:
489 CardValue* const _base_addr;
490
491 CardValue* _cur_addr;
492 CardValue* const _end_addr;
493
494 static const size_t ToScanMask = G1CardTable::g1_card_already_scanned;
495 static const size_t ExpandedToScanMask = G1CardTable::WordAlreadyScanned;
496
497 bool cur_addr_aligned() const {
498 return ((uintptr_t)_cur_addr) % sizeof(size_t) == 0;
499 }
500
501 bool cur_card_is_dirty() const {
502 CardValue value = *_cur_addr;
503 return (value & ToScanMask) == 0;
504 }
505
506 bool cur_word_of_cards_contains_any_dirty_card() const {
507 assert(cur_addr_aligned(), "Current address should be aligned");
508 size_t const value = *(size_t*)_cur_addr;
509 return (~value & ExpandedToScanMask) != 0;
510 }
511
512 bool cur_word_of_cards_all_dirty_cards() const {
513 size_t const value = *(size_t*)_cur_addr;
514 return value == G1CardTable::WordAllDirty;
515 }
516
517 size_t get_and_advance_pos() {
518 _cur_addr++;
519 return pointer_delta(_cur_addr, _base_addr, sizeof(CardValue)) - 1;
520 }
521
522 public:
523 G1CardTableScanner(CardValue* start_card, size_t size) :
524 _base_addr(start_card),
525 _cur_addr(start_card),
526 _end_addr(start_card + size) {
527
528 assert(is_aligned(start_card, sizeof(size_t)), "Unaligned start addr " PTR_FORMAT, p2i(start_card));
529 assert(is_aligned(size, sizeof(size_t)), "Unaligned size " SIZE_FORMAT, size);
530 }
531
532 size_t find_next_dirty() {
533 while (!cur_addr_aligned()) {
534 if (cur_card_is_dirty()) {
535 return get_and_advance_pos();
536 }
537 _cur_addr++;
538 }
539
540 assert(cur_addr_aligned(), "Current address should be aligned now.");
541 while (_cur_addr != _end_addr) {
542 if (cur_word_of_cards_contains_any_dirty_card()) {
543 for (size_t i = 0; i < sizeof(size_t); i++) {
544 if (cur_card_is_dirty()) {
545 return get_and_advance_pos();
546 }
547 _cur_addr++;
548 }
549 assert(false, "Should not reach here given we detected a dirty card in the word.");
550 }
551 _cur_addr += sizeof(size_t);
552 }
553 return get_and_advance_pos();
554 }
555
556 size_t find_next_non_dirty() {
557 assert(_cur_addr <= _end_addr, "Not allowed to search for marks after area.");
558
559 while (!cur_addr_aligned()) {
560 if (!cur_card_is_dirty()) {
561 return get_and_advance_pos();
562 }
563 _cur_addr++;
564 }
565
566 assert(cur_addr_aligned(), "Current address should be aligned now.");
567 while (_cur_addr != _end_addr) {
568 if (!cur_word_of_cards_all_dirty_cards()) {
569 for (size_t i = 0; i < sizeof(size_t); i++) {
570 if (!cur_card_is_dirty()) {
571 return get_and_advance_pos();
572 }
573 _cur_addr++;
574 }
575 assert(false, "Should not reach here given we detected a non-dirty card in the word.");
576 }
577 _cur_addr += sizeof(size_t);
578 }
579 return get_and_advance_pos();
580 }
581 };
582
583 // Helper class to claim dirty chunks within the card table.
584 class G1CardTableChunkClaimer {
585 G1RemSetScanState* _scan_state;
586 uint _region_idx;
587 uint _cur_claim;
588
589 public:
590 G1CardTableChunkClaimer(G1RemSetScanState* scan_state, uint region_idx) :
591 _scan_state(scan_state),
592 _region_idx(region_idx),
593 _cur_claim(0) {
594 guarantee(size() <= HeapRegion::CardsPerRegion, "Should not claim more space than possible.");
595 }
596
597 bool has_next() {
598 while (true) {
599 _cur_claim = _scan_state->claim_cards_to_scan(_region_idx, size());
600 if (_cur_claim >= HeapRegion::CardsPerRegion) {
601 return false;
602 }
603 if (_scan_state->chunk_needs_scan(_region_idx, _cur_claim)) {
604 return true;
605 }
606 }
607 }
608
609 uint value() const { return _cur_claim; }
610 uint size() const { return _scan_state->scan_chunk_size(); }
611 };
612
613 // Scans a heap region for dirty cards.
614 class G1ScanHRForRegionClosure : public HeapRegionClosure {
615 G1CollectedHeap* _g1h;
616 G1CardTable* _ct;
617 G1BlockOffsetTable* _bot;
618
619 G1ParScanThreadState* _pss;
620
621 G1RemSetScanState* _scan_state;
622
623 G1GCPhaseTimes::GCParPhases _phase;
624
625 uint _worker_id;
626
627 size_t _cards_scanned;
628 size_t _blocks_scanned;
629 size_t _chunks_claimed;
630
631 Tickspan _rem_set_root_scan_time;
632 Tickspan _rem_set_trim_partially_time;
633
634 // The address to which this thread already scanned (walked the heap) up to during
635 // card scanning (exclusive).
636 HeapWord* _scanned_to;
637
638 HeapWord* scan_memregion(uint region_idx_for_card, MemRegion mr) {
639 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
640 G1ScanCardClosure card_cl(_g1h, _pss);
641
642 HeapWord* const scanned_to = card_region->oops_on_memregion_seq_iterate_careful<true>(mr, &card_cl);
643 assert(scanned_to != NULL, "Should be able to scan range");
644 assert(scanned_to >= mr.end(), "Scanned to " PTR_FORMAT " less than range " PTR_FORMAT, p2i(scanned_to), p2i(mr.end()));
645
646 _pss->trim_queue_partially();
647 return scanned_to;
648 }
649
650 void do_claimed_block(uint const region_idx_for_card, size_t const first_card, size_t const num_cards) {
651 HeapWord* const card_start = _bot->address_for_index_raw(first_card);
652 #ifdef ASSERT
653 HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
654 assert(hr == NULL || hr->is_in_reserved(card_start),
655 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
656 #endif
657 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
658 if (card_start >= top) {
659 return;
660 }
661
662 HeapWord* scan_end = MIN2(card_start + (num_cards << BOTConstants::LogN_words), top);
663 if (_scanned_to >= scan_end) {
664 return;
665 }
666 MemRegion mr(MAX2(card_start, _scanned_to), scan_end);
667 _scanned_to = scan_memregion(region_idx_for_card, mr);
668
669 _cards_scanned += num_cards;
670 }
671
672 ALWAYSINLINE void do_card_block(uint const region_idx, size_t const first_card, size_t const num_cards) {
673 _ct->mark_as_scanned(first_card, num_cards);
674 do_claimed_block(region_idx, first_card, num_cards);
675 _blocks_scanned++;
676 }
677
678 void scan_heap_roots(HeapRegion* r) {
679 EventGCPhaseParallel event;
680 uint const region_idx = r->hrm_index();
681
682 ResourceMark rm;
683
684 G1CardTableChunkClaimer claim(_scan_state, region_idx);
685
686 // Set the current scan "finger" to NULL for every heap region to scan. Since
687 // the claim value is monotonically increasing, the check to not scan below this
688 // will filter out objects spanning chunks within the region too then, as opposed
689 // to resetting this value for every claim.
690 _scanned_to = NULL;
691
692 while (claim.has_next()) {
693 size_t const region_card_base_idx = ((size_t)region_idx << HeapRegion::LogCardsPerRegion) + claim.value();
694 CardTable::CardValue* const base_addr = _ct->byte_for_index(region_card_base_idx);
695
696 G1CardTableScanner scan(base_addr, claim.size());
697
698 size_t first_scan_idx = scan.find_next_dirty();
699 while (first_scan_idx != claim.size()) {
700 assert(*_ct->byte_for_index(region_card_base_idx + first_scan_idx) <= 0x1, "is %d at region %u idx " SIZE_FORMAT, *_ct->byte_for_index(region_card_base_idx + first_scan_idx), region_idx, first_scan_idx);
701
702 size_t const last_scan_idx = scan.find_next_non_dirty();
703 size_t const len = last_scan_idx - first_scan_idx;
704
705 do_card_block(region_idx, region_card_base_idx + first_scan_idx, len);
706
707 if (last_scan_idx == claim.size()) {
708 break;
709 }
710
711 first_scan_idx = scan.find_next_dirty();
712 }
713 _chunks_claimed++;
714 }
715
716 event.commit(GCId::current(), _worker_id, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::ScanHR));
717 }
718
719 public:
720 G1ScanHRForRegionClosure(G1RemSetScanState* scan_state,
721 G1ParScanThreadState* pss,
722 uint worker_id,
723 G1GCPhaseTimes::GCParPhases phase) :
724 _g1h(G1CollectedHeap::heap()),
725 _ct(_g1h->card_table()),
726 _bot(_g1h->bot()),
727 _pss(pss),
728 _scan_state(scan_state),
729 _phase(phase),
730 _worker_id(worker_id),
731 _cards_scanned(0),
732 _blocks_scanned(0),
733 _chunks_claimed(0),
734 _rem_set_root_scan_time(),
735 _rem_set_trim_partially_time(),
736 _scanned_to(NULL) {
737 }
738
739 bool do_heap_region(HeapRegion* r) {
740 assert(!r->in_collection_set() && r->is_old_or_humongous_or_archive(),
741 "Should only be called on old gen non-collection set regions but region %u is not.",
742 r->hrm_index());
743 uint const region_idx = r->hrm_index();
744
745 if (_scan_state->has_cards_to_scan(region_idx)) {
746 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
747 scan_heap_roots(r);
748 }
749 return false;
750 }
751
752 Tickspan rem_set_root_scan_time() const { return _rem_set_root_scan_time; }
753 Tickspan rem_set_trim_partially_time() const { return _rem_set_trim_partially_time; }
754
755 size_t cards_scanned() const { return _cards_scanned; }
756 size_t blocks_scanned() const { return _blocks_scanned; }
757 size_t chunks_claimed() const { return _chunks_claimed; }
758 };
759
760 void G1RemSet::scan_heap_roots(G1ParScanThreadState* pss,
761 uint worker_id,
762 G1GCPhaseTimes::GCParPhases scan_phase,
763 G1GCPhaseTimes::GCParPhases objcopy_phase) {
764 G1ScanHRForRegionClosure cl(_scan_state, pss, worker_id, scan_phase);
765 _scan_state->iterate_dirty_regions_from(&cl, worker_id);
766
767 G1GCPhaseTimes* p = _g1p->phase_times();
768
769 p->record_or_add_time_secs(objcopy_phase, worker_id, cl.rem_set_trim_partially_time().seconds());
770
771 p->record_or_add_time_secs(scan_phase, worker_id, cl.rem_set_root_scan_time().seconds());
772 p->record_or_add_thread_work_item(scan_phase, worker_id, cl.cards_scanned(), G1GCPhaseTimes::ScanHRScannedCards);
773 p->record_or_add_thread_work_item(scan_phase, worker_id, cl.blocks_scanned(), G1GCPhaseTimes::ScanHRScannedBlocks);
774 p->record_or_add_thread_work_item(scan_phase, worker_id, cl.chunks_claimed(), G1GCPhaseTimes::ScanHRClaimedChunks);
775 }
776
777 // Heap region closure to be applied to all regions in the current collection set
778 // increment to fix up non-card related roots.
779 class G1ScanCollectionSetRegionClosure : public HeapRegionClosure {
780 G1ParScanThreadState* _pss;
781 G1RemSetScanState* _scan_state;
782
783 G1GCPhaseTimes::GCParPhases _scan_phase;
784 G1GCPhaseTimes::GCParPhases _code_roots_phase;
785
786 uint _worker_id;
787
788 size_t _opt_refs_scanned;
789 size_t _opt_refs_memory_used;
790
791 Tickspan _strong_code_root_scan_time;
792 Tickspan _strong_code_trim_partially_time;
793
794 Tickspan _rem_set_opt_root_scan_time;
795 Tickspan _rem_set_opt_trim_partially_time;
796
797 void scan_opt_rem_set_roots(HeapRegion* r) {
798 EventGCPhaseParallel event;
799
800 G1OopStarChunkedList* opt_rem_set_list = _pss->oops_into_optional_region(r);
801
802 G1ScanCardClosure scan_cl(G1CollectedHeap::heap(), _pss);
803 G1ScanRSForOptionalClosure cl(G1CollectedHeap::heap(), &scan_cl);
804 _opt_refs_scanned += opt_rem_set_list->oops_do(&cl, _pss->closures()->strong_oops());
805 _opt_refs_memory_used += opt_rem_set_list->used_memory();
806
807 event.commit(GCId::current(), _worker_id, G1GCPhaseTimes::phase_name(_scan_phase));
808 }
809
810 public:
811 G1ScanCollectionSetRegionClosure(G1RemSetScanState* scan_state,
812 G1ParScanThreadState* pss,
813 uint worker_id,
814 G1GCPhaseTimes::GCParPhases scan_phase,
815 G1GCPhaseTimes::GCParPhases code_roots_phase) :
816 _pss(pss),
817 _scan_state(scan_state),
818 _scan_phase(scan_phase),
819 _code_roots_phase(code_roots_phase),
820 _worker_id(worker_id),
821 _opt_refs_scanned(0),
822 _opt_refs_memory_used(0),
823 _strong_code_root_scan_time(),
824 _strong_code_trim_partially_time(),
825 _rem_set_opt_root_scan_time(),
826 _rem_set_opt_trim_partially_time() { }
827
828 bool do_heap_region(HeapRegion* r) {
829 uint const region_idx = r->hrm_index();
830
831 // The individual references for the optional remembered set are per-worker, so we
832 // always need to scan them.
833 if (r->has_index_in_opt_cset()) {
834 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_opt_root_scan_time, _rem_set_opt_trim_partially_time);
835 scan_opt_rem_set_roots(r);
836 }
837
838 if (_scan_state->claim_collection_set_region(region_idx)) {
839 EventGCPhaseParallel event;
840
841 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
842 // Scan the strong code root list attached to the current region
843 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
844
845 event.commit(GCId::current(), _worker_id, G1GCPhaseTimes::phase_name(_code_roots_phase));
846 }
847
848 return false;
849 }
850
851 Tickspan strong_code_root_scan_time() const { return _strong_code_root_scan_time; }
852 Tickspan strong_code_root_trim_partially_time() const { return _strong_code_trim_partially_time; }
853
854 Tickspan rem_set_opt_root_scan_time() const { return _rem_set_opt_root_scan_time; }
855 Tickspan rem_set_opt_trim_partially_time() const { return _rem_set_opt_trim_partially_time; }
856
857 size_t opt_refs_scanned() const { return _opt_refs_scanned; }
858 size_t opt_refs_memory_used() const { return _opt_refs_memory_used; }
859 };
860
861 void G1RemSet::scan_collection_set_regions(G1ParScanThreadState* pss,
862 uint worker_id,
863 G1GCPhaseTimes::GCParPhases scan_phase,
864 G1GCPhaseTimes::GCParPhases coderoots_phase,
865 G1GCPhaseTimes::GCParPhases objcopy_phase) {
866 G1ScanCollectionSetRegionClosure cl(_scan_state, pss, worker_id, scan_phase, coderoots_phase);
867 _g1h->collection_set_iterate_increment_from(&cl, worker_id);
868
869 G1GCPhaseTimes* p = _g1h->phase_times();
870
871 p->record_or_add_time_secs(scan_phase, worker_id, cl.rem_set_opt_root_scan_time().seconds());
872 p->record_or_add_time_secs(scan_phase, worker_id, cl.rem_set_opt_trim_partially_time().seconds());
873
874 p->record_or_add_time_secs(coderoots_phase, worker_id, cl.strong_code_root_scan_time().seconds());
875 p->add_time_secs(objcopy_phase, worker_id, cl.strong_code_root_trim_partially_time().seconds());
876
877 // At this time we record some metrics only for the evacuations after the initial one.
878 if (scan_phase == G1GCPhaseTimes::OptScanHR) {
879 p->record_or_add_thread_work_item(scan_phase, worker_id, cl.opt_refs_scanned(), G1GCPhaseTimes::ScanHRScannedOptRefs);
880 p->record_or_add_thread_work_item(scan_phase, worker_id, cl.opt_refs_memory_used(), G1GCPhaseTimes::ScanHRUsedMemory);
881 }
882 }
883
884 void G1RemSet::prepare_region_for_scan(HeapRegion* region) {
885 uint hrm_index = region->hrm_index();
886
887 _scan_state->reset_region_claim(hrm_index);
888 if (region->in_collection_set()) {
889 // Young regions had their card table marked as young at their allocation;
890 // we need to make sure that these marks are cleared at the end of GC, *but*
891 // they should not be scanned for cards.
892 // So directly add them to the "all_dirty_regions".
893 // Same for regions in the (initial) collection set: they may contain cards from
894 // the log buffers, make sure they are cleaned.
895 _scan_state->clear_scan_top(hrm_index);
896 _scan_state->add_all_dirty_region(hrm_index);
897 } else {
898 assert(region->is_old_or_humongous_or_archive(), "All other regions should be in the collection set");
899 _scan_state->set_scan_top(hrm_index, region->top());
900 }
901 }
902
903 void G1RemSet::prepare_for_scan_heap_roots() {
904 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
905 dcqs.concatenate_logs();
906
907 _scan_state->prepare();
908 }
909
910 class G1MergeHeapRootsTask : public AbstractGangTask {
911
912 // Visitor for remembered sets, dropping entries onto the card table.
913 class G1MergeCardSetClosure : public HeapRegionClosure {
914 G1RemSetScanState* _scan_state;
915 G1CardTable* _ct;
916
917 uint _merged_sparse;
918 uint _merged_fine;
919 uint _merged_coarse;
920
921 // Returns if the region contains cards we need to scan. If so, remember that
922 // region in the current set of dirty regions.
923 bool remember_if_interesting(uint const region_idx) {
924 if (!_scan_state->contains_cards_to_process(region_idx)) {
925 return false;
926 }
927 _scan_state->add_dirty_region(region_idx);
928 return true;
929 }
930 public:
931 G1MergeCardSetClosure(G1RemSetScanState* scan_state) :
932 _scan_state(scan_state),
933 _ct(G1CollectedHeap::heap()->card_table()),
934 _merged_sparse(0),
935 _merged_fine(0),
936 _merged_coarse(0) { }
937
938 void next_coarse_prt(uint const region_idx) {
939 if (!remember_if_interesting(region_idx)) {
940 return;
941 }
942
943 _merged_coarse++;
944
945 size_t region_base_idx = (size_t)region_idx << HeapRegion::LogCardsPerRegion;
946 _ct->mark_region_dirty(region_base_idx, HeapRegion::CardsPerRegion);
947 _scan_state->set_chunk_region_dirty(region_base_idx);
948 }
949
950 void next_fine_prt(uint const region_idx, BitMap* bm) {
951 if (!remember_if_interesting(region_idx)) {
952 return;
953 }
954
955 _merged_fine++;
956
957 size_t const region_base_idx = (size_t)region_idx << HeapRegion::LogCardsPerRegion;
958 BitMap::idx_t cur = bm->get_next_one_offset(0);
959 while (cur != bm->size()) {
960 _ct->mark_clean_as_dirty(region_base_idx + cur);
961 _scan_state->set_chunk_dirty(region_base_idx + cur);
962 cur = bm->get_next_one_offset(cur + 1);
963 }
964 }
965
966 void next_sparse_prt(uint const region_idx, SparsePRTEntry::card_elem_t* cards, uint const num_cards) {
967 if (!remember_if_interesting(region_idx)) {
968 return;
969 }
970
971 _merged_sparse++;
972
973 size_t const region_base_idx = (size_t)region_idx << HeapRegion::LogCardsPerRegion;
974 for (uint i = 0; i < num_cards; i++) {
975 size_t card_idx = region_base_idx + cards[i];
976 _ct->mark_clean_as_dirty(card_idx);
977 _scan_state->set_chunk_dirty(card_idx);
978 }
979 }
980
981 virtual bool do_heap_region(HeapRegion* r) {
982 assert(r->in_collection_set() || r->is_starts_humongous(), "must be");
983
984 HeapRegionRemSet* rem_set = r->rem_set();
985 if (!rem_set->is_empty()) {
986 rem_set->iterate_prts(*this);
987 }
988
989 return false;
990 }
991
992 size_t merged_sparse() const { return _merged_sparse; }
993 size_t merged_fine() const { return _merged_fine; }
994 size_t merged_coarse() const { return _merged_coarse; }
995 };
996
997 // Visitor for the remembered sets of humongous candidate regions to merge their
998 // remembered set into the card table.
999 class G1FlushHumongousCandidateRemSets : public HeapRegionClosure {
1000 G1MergeCardSetClosure _cl;
1001
1002 public:
1003 G1FlushHumongousCandidateRemSets(G1RemSetScanState* scan_state) : _cl(scan_state) { }
1004
1005 virtual bool do_heap_region(HeapRegion* r) {
1006 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1007
1008 if (!r->is_starts_humongous() ||
1009 !g1h->region_attr(r->hrm_index()).is_humongous() ||
1010 r->rem_set()->is_empty()) {
1011 return false;
1012 }
1013
1014 guarantee(r->rem_set()->occupancy_less_or_equal_than(G1RSetSparseRegionEntries),
1015 "Found a not-small remembered set here. This is inconsistent with previous assumptions.");
1016
1017 _cl.do_heap_region(r);
1018
1019 // We should only clear the card based remembered set here as we will not
1020 // implicitly rebuild anything else during eager reclaim. Note that at the moment
1021 // (and probably never) we do not enter this path if there are other kind of
1022 // remembered sets for this region.
1023 r->rem_set()->clear_locked(true /* only_cardset */);
1024 // Clear_locked() above sets the state to Empty. However we want to continue
1025 // collecting remembered set entries for humongous regions that were not
1026 // reclaimed.
1027 r->rem_set()->set_state_complete();
1028 #ifdef ASSERT
1029 G1HeapRegionAttr region_attr = g1h->region_attr(r->hrm_index());
1030 assert(region_attr.needs_remset_update(), "must be");
1031 #endif
1032 assert(r->rem_set()->is_empty(), "At this point any humongous candidate remembered set must be empty.");
1033
1034 return false;
1035 }
1036
1037 size_t merged_sparse() const { return _cl.merged_sparse(); }
1038 size_t merged_fine() const { return _cl.merged_fine(); }
1039 size_t merged_coarse() const { return _cl.merged_coarse(); }
1040 };
1041
1042 // Visitor for the log buffer entries to merge them into the card table.
1043 class G1MergeLogBufferCardsClosure : public G1CardTableEntryClosure {
1044 G1RemSetScanState* _scan_state;
1045 G1CardTable* _ct;
1046
1047 size_t _cards_dirty;
1048 size_t _cards_skipped;
1049 public:
1050 G1MergeLogBufferCardsClosure(G1CollectedHeap* g1h, G1RemSetScanState* scan_state) :
1051 _scan_state(scan_state), _ct(g1h->card_table()), _cards_dirty(0), _cards_skipped(0)
1052 {}
1053
1054 void do_card_ptr(CardValue* card_ptr, uint worker_id) {
1055 // The only time we care about recording cards that
1056 // contain references that point into the collection set
1057 // is during RSet updating within an evacuation pause.
1058 // In this case worker_id should be the id of a GC worker thread.
1059 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
1060
1061 uint const region_idx = _ct->region_idx_for(card_ptr);
1062
1063 // The second clause must come after - the log buffers might contain cards to uncommited
1064 // regions.
1065 // This code may count duplicate entries in the log buffers (even if rare) multiple
1066 // times.
1067 if (_scan_state->contains_cards_to_process(region_idx) && (*card_ptr == G1CardTable::dirty_card_val())) {
1068 _scan_state->add_dirty_region(region_idx);
1069 _scan_state->set_chunk_dirty(_ct->index_for_cardvalue(card_ptr));
1070 _cards_dirty++;
1071 } else {
1072 // We may have had dirty cards in the (initial) collection set (or the
1073 // young regions which are always in the initial collection set). We do
1074 // not fix their cards here: we already added these regions to the set of
1075 // regions to clear the card table at the end during the prepare() phase.
1076 _cards_skipped++;
1077 }
1078 }
1079
1080 size_t cards_dirty() const { return _cards_dirty; }
1081 size_t cards_skipped() const { return _cards_skipped; }
1082 };
1083
1084 HeapRegionClaimer _hr_claimer;
1085 G1RemSetScanState* _scan_state;
1086 BufferNode::Stack _dirty_card_buffers;
1087 bool _initial_evacuation;
1088
1089 volatile bool _fast_reclaim_handled;
1090
1091 void apply_closure_to_dirty_card_buffers(G1MergeLogBufferCardsClosure* cl, uint worker_id) {
1092 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
1093 size_t buffer_size = dcqs.buffer_size();
1094 while (BufferNode* node = _dirty_card_buffers.pop()) {
1095 cl->apply_to_buffer(node, buffer_size, worker_id);
1096 dcqs.deallocate_buffer(node);
1097 }
1098 }
1099
1100 public:
1101 G1MergeHeapRootsTask(G1RemSetScanState* scan_state, uint num_workers, bool initial_evacuation) :
1102 AbstractGangTask("G1 Merge Heap Roots"),
1103 _hr_claimer(num_workers),
1104 _scan_state(scan_state),
1105 _dirty_card_buffers(),
1106 _initial_evacuation(initial_evacuation),
1107 _fast_reclaim_handled(false)
1108 {
1109 if (initial_evacuation) {
1110 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
1111 G1BufferNodeList buffers = dcqs.take_all_completed_buffers();
1112 if (buffers._entry_count != 0) {
1113 _dirty_card_buffers.prepend(*buffers._head, *buffers._tail);
1114 }
1115 }
1116 }
1117
1118 virtual void work(uint worker_id) {
1119 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1120 G1GCPhaseTimes* p = g1h->phase_times();
1121
1122 G1GCPhaseTimes::GCParPhases merge_remset_phase = _initial_evacuation ?
1123 G1GCPhaseTimes::MergeRS :
1124 G1GCPhaseTimes::OptMergeRS;
1125
1126 // We schedule flushing the remembered sets of humongous fast reclaim candidates
1127 // onto the card table first to allow the remaining parallelized tasks hide it.
1128 if (_initial_evacuation &&
1129 p->fast_reclaim_humongous_candidates() > 0 &&
1130 !_fast_reclaim_handled &&
1131 !Atomic::cmpxchg(true, &_fast_reclaim_handled, false)) {
1132
1133 G1GCParPhaseTimesTracker x(p, G1GCPhaseTimes::MergeER, worker_id);
1134
1135 G1FlushHumongousCandidateRemSets cl(_scan_state);
1136 g1h->heap_region_iterate(&cl);
1137
1138 p->record_or_add_thread_work_item(merge_remset_phase, worker_id, cl.merged_sparse(), G1GCPhaseTimes::MergeRSMergedSparse);
1139 p->record_or_add_thread_work_item(merge_remset_phase, worker_id, cl.merged_fine(), G1GCPhaseTimes::MergeRSMergedFine);
1140 p->record_or_add_thread_work_item(merge_remset_phase, worker_id, cl.merged_coarse(), G1GCPhaseTimes::MergeRSMergedCoarse);
1141 }
1142
1143 // Merge remembered sets of current candidates.
1144 {
1145 G1GCParPhaseTimesTracker x(p, merge_remset_phase, worker_id, _initial_evacuation /* must_record */);
1146 G1MergeCardSetClosure cl(_scan_state);
1147 g1h->collection_set_iterate_increment_from(&cl, &_hr_claimer, worker_id);
1148
1149 p->record_or_add_thread_work_item(merge_remset_phase, worker_id, cl.merged_sparse(), G1GCPhaseTimes::MergeRSMergedSparse);
1150 p->record_or_add_thread_work_item(merge_remset_phase, worker_id, cl.merged_fine(), G1GCPhaseTimes::MergeRSMergedFine);
1151 p->record_or_add_thread_work_item(merge_remset_phase, worker_id, cl.merged_coarse(), G1GCPhaseTimes::MergeRSMergedCoarse);
1152 }
1153
1154 // Apply closure to log entries in the HCC.
1155 if (_initial_evacuation && G1HotCardCache::default_use_cache()) {
1156 assert(merge_remset_phase == G1GCPhaseTimes::MergeRS, "Wrong merge phase");
1157 G1GCParPhaseTimesTracker x(p, G1GCPhaseTimes::MergeHCC, worker_id);
1158 G1MergeLogBufferCardsClosure cl(g1h, _scan_state);
1159 g1h->iterate_hcc_closure(&cl, worker_id);
1160
1161 p->record_thread_work_item(G1GCPhaseTimes::MergeHCC, worker_id, cl.cards_dirty(), G1GCPhaseTimes::MergeHCCDirtyCards);
1162 p->record_thread_work_item(G1GCPhaseTimes::MergeHCC, worker_id, cl.cards_skipped(), G1GCPhaseTimes::MergeHCCSkippedCards);
1163 }
1164
1165 // Now apply the closure to all remaining log entries.
1166 if (_initial_evacuation) {
1167 assert(merge_remset_phase == G1GCPhaseTimes::MergeRS, "Wrong merge phase");
1168 G1GCParPhaseTimesTracker x(p, G1GCPhaseTimes::MergeLB, worker_id);
1169
1170 G1MergeLogBufferCardsClosure cl(g1h, _scan_state);
1171 apply_closure_to_dirty_card_buffers(&cl, worker_id);
1172
1173 p->record_thread_work_item(G1GCPhaseTimes::MergeLB, worker_id, cl.cards_dirty(), G1GCPhaseTimes::MergeLBDirtyCards);
1174 p->record_thread_work_item(G1GCPhaseTimes::MergeLB, worker_id, cl.cards_skipped(), G1GCPhaseTimes::MergeLBSkippedCards);
1175 }
1176 }
1177 };
1178
1179 void G1RemSet::print_merge_heap_roots_stats() {
1180 size_t num_visited_cards = _scan_state->num_visited_cards();
1181
1182 size_t total_dirty_region_cards = _scan_state->num_cards_in_dirty_regions();
1183
1184 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1185 size_t total_old_region_cards =
1186 (g1h->num_regions() - (g1h->num_free_regions() - g1h->collection_set()->cur_length())) * HeapRegion::CardsPerRegion;
1187
1188 log_debug(gc,remset)("Visited cards " SIZE_FORMAT " Total dirty " SIZE_FORMAT " (%.2lf%%) Total old " SIZE_FORMAT " (%.2lf%%)",
1189 num_visited_cards,
1190 total_dirty_region_cards,
1191 percent_of(num_visited_cards, total_dirty_region_cards),
1192 total_old_region_cards,
1193 percent_of(num_visited_cards, total_old_region_cards));
1194 }
1195
1196 void G1RemSet::merge_heap_roots(bool initial_evacuation) {
1197 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1198
1199 {
1200 Ticks start = Ticks::now();
1201
1202 _scan_state->prepare_for_merge_heap_roots();
1203
1204 Tickspan total = Ticks::now() - start;
1205 if (initial_evacuation) {
1206 g1h->phase_times()->record_prepare_merge_heap_roots_time(total.seconds() * 1000.0);
1207 } else {
1208 g1h->phase_times()->record_or_add_optional_prepare_merge_heap_roots_time(total.seconds() * 1000.0);
1209 }
1210 }
1211
1212 WorkGang* workers = g1h->workers();
1213 size_t const increment_length = g1h->collection_set()->increment_length();
1214
1215 uint const num_workers = initial_evacuation ? workers->active_workers() :
1216 MIN2(workers->active_workers(), (uint)increment_length);
1217
1218 {
1219 G1MergeHeapRootsTask cl(_scan_state, num_workers, initial_evacuation);
1220 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " regions",
1221 cl.name(), num_workers, increment_length);
1222 workers->run_task(&cl, num_workers);
1223 }
1224
1225 if (log_is_enabled(Debug, gc, remset)) {
1226 print_merge_heap_roots_stats();
1227 }
1228 }
1229
1230 void G1RemSet::prepare_for_scan_heap_roots(uint region_idx) {
1231 _scan_state->clear_scan_top(region_idx);
1232 }
1233
1234 void G1RemSet::cleanup_after_scan_heap_roots() {
1235 G1GCPhaseTimes* phase_times = _g1h->phase_times();
1236
1237 // Set all cards back to clean.
1238 double start = os::elapsedTime();
1239 _scan_state->cleanup(_g1h->workers());
1240 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
1241 }
1242
1243 inline void check_card_ptr(CardTable::CardValue* card_ptr, G1CardTable* ct) {
1244 #ifdef ASSERT
1245 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1246 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
1247 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
1248 p2i(card_ptr),
1249 ct->index_for(ct->addr_for(card_ptr)),
1250 p2i(ct->addr_for(card_ptr)),
1251 g1h->addr_to_region(ct->addr_for(card_ptr)));
1252 #endif
1253 }
1254
1255 void G1RemSet::refine_card_concurrently(CardValue* card_ptr,
1256 uint worker_id) {
1257 assert(!_g1h->is_gc_active(), "Only call concurrently");
1258
1259 // Construct the region representing the card.
1260 HeapWord* start = _ct->addr_for(card_ptr);
1261 // And find the region containing it.
1262 HeapRegion* r = _g1h->heap_region_containing_or_null(start);
1263
1264 // If this is a (stale) card into an uncommitted region, exit.
1265 if (r == NULL) {
1266 return;
1267 }
1268
1269 check_card_ptr(card_ptr, _ct);
1270
1271 // If the card is no longer dirty, nothing to do.
1272 if (*card_ptr != G1CardTable::dirty_card_val()) {
1273 return;
1274 }
1275
1276 // This check is needed for some uncommon cases where we should
1277 // ignore the card.
1278 //
1279 // The region could be young. Cards for young regions are
1280 // distinctly marked (set to g1_young_gen), so the post-barrier will
1281 // filter them out. However, that marking is performed
1282 // concurrently. A write to a young object could occur before the
1283 // card has been marked young, slipping past the filter.
1284 //
1285 // The card could be stale, because the region has been freed since
1286 // the card was recorded. In this case the region type could be
1287 // anything. If (still) free or (reallocated) young, just ignore
1288 // it. If (reallocated) old or humongous, the later card trimming
1289 // and additional checks in iteration may detect staleness. At
1290 // worst, we end up processing a stale card unnecessarily.
1291 //
1292 // In the normal (non-stale) case, the synchronization between the
1293 // enqueueing of the card and processing it here will have ensured
1294 // we see the up-to-date region type here.
1295 if (!r->is_old_or_humongous_or_archive()) {
1296 return;
1297 }
1298
1299 // The result from the hot card cache insert call is either:
1300 // * pointer to the current card
1301 // (implying that the current card is not 'hot'),
1302 // * null
1303 // (meaning we had inserted the card ptr into the "hot" card cache,
1304 // which had some headroom),
1305 // * a pointer to a "hot" card that was evicted from the "hot" cache.
1306 //
1307
1308 if (_hot_card_cache->use_cache()) {
1309 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
1310
1311 const CardValue* orig_card_ptr = card_ptr;
1312 card_ptr = _hot_card_cache->insert(card_ptr);
1313 if (card_ptr == NULL) {
1314 // There was no eviction. Nothing to do.
1315 return;
1316 } else if (card_ptr != orig_card_ptr) {
1317 // Original card was inserted and an old card was evicted.
1318 start = _ct->addr_for(card_ptr);
1319 r = _g1h->heap_region_containing(start);
1320
1321 // Check whether the region formerly in the cache should be
1322 // ignored, as discussed earlier for the original card. The
1323 // region could have been freed while in the cache.
1324 if (!r->is_old_or_humongous_or_archive()) {
1325 return;
1326 }
1327 } // Else we still have the original card.
1328 }
1329
1330 // Trim the region designated by the card to what's been allocated
1331 // in the region. The card could be stale, or the card could cover
1332 // (part of) an object at the end of the allocated space and extend
1333 // beyond the end of allocation.
1334
1335 // Non-humongous objects are only allocated in the old-gen during
1336 // GC, so if region is old then top is stable. Humongous object
1337 // allocation sets top last; if top has not yet been set, this is
1338 // a stale card and we'll end up with an empty intersection. If
1339 // this is not a stale card, the synchronization between the
1340 // enqueuing of the card and processing it here will have ensured
1341 // we see the up-to-date top here.
1342 HeapWord* scan_limit = r->top();
1343
1344 if (scan_limit <= start) {
1345 // If the trimmed region is empty, the card must be stale.
1346 return;
1347 }
1348
1349 // Okay to clean and process the card now. There are still some
1350 // stale card cases that may be detected by iteration and dealt with
1351 // as iteration failure.
1352 *const_cast<volatile CardValue*>(card_ptr) = G1CardTable::clean_card_val();
1353
1354 // This fence serves two purposes. First, the card must be cleaned
1355 // before processing the contents. Second, we can't proceed with
1356 // processing until after the read of top, for synchronization with
1357 // possibly concurrent humongous object allocation. It's okay that
1358 // reading top and reading type were racy wrto each other. We need
1359 // both set, in any order, to proceed.
1360 OrderAccess::fence();
1361
1362 // Don't use addr_for(card_ptr + 1) which can ask for
1363 // a card beyond the heap.
1364 HeapWord* end = start + G1CardTable::card_size_in_words;
1365 MemRegion dirty_region(start, MIN2(scan_limit, end));
1366 assert(!dirty_region.is_empty(), "sanity");
1367
1368 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_id);
1369 if (r->oops_on_memregion_seq_iterate_careful<false>(dirty_region, &conc_refine_cl) != NULL) {
1370 return;
1371 }
1372
1373 // If unable to process the card then we encountered an unparsable
1374 // part of the heap (e.g. a partially allocated object, so only
1375 // temporarily a problem) while processing a stale card. Despite
1376 // the card being stale, we can't simply ignore it, because we've
1377 // already marked the card cleaned, so taken responsibility for
1378 // ensuring the card gets scanned.
1379 //
1380 // However, the card might have gotten re-dirtied and re-enqueued
1381 // while we worked. (In fact, it's pretty likely.)
1382 if (*card_ptr == G1CardTable::dirty_card_val()) {
1383 return;
1384 }
1385
1386 // Re-dirty the card and enqueue in the *shared* queue. Can't use
1387 // the thread-local queue, because that might be the queue that is
1388 // being processed by us; we could be a Java thread conscripted to
1389 // perform refinement on our queue's current buffer.
1390 *card_ptr = G1CardTable::dirty_card_val();
1391 G1BarrierSet::shared_dirty_card_queue().enqueue(card_ptr);
1392 }
1393
1394 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
1395 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
1396 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
1397
1398 G1RemSetSummary current;
1399 _prev_period_summary.subtract_from(¤t);
1400
1401 Log(gc, remset) log;
1402 log.trace("%s", header);
1403 ResourceMark rm;
1404 LogStream ls(log.trace());
1405 _prev_period_summary.print_on(&ls);
1406
1407 _prev_period_summary.set(¤t);
1408 }
1409 }
1410
1411 void G1RemSet::print_summary_info() {
1412 Log(gc, remset, exit) log;
1413 if (log.is_trace()) {
1414 log.trace(" Cumulative RS summary");
1415 G1RemSetSummary current;
1416 ResourceMark rm;
1417 LogStream ls(log.trace());
1418 current.print_on(&ls);
1419 }
1420 }
1421
1422 class G1RebuildRemSetTask: public AbstractGangTask {
1423 // Aggregate the counting data that was constructed concurrently
1424 // with marking.
1425 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
1426 G1ConcurrentMark* _cm;
1427 G1RebuildRemSetClosure _update_cl;
1428
1429 // Applies _update_cl to the references of the given object, limiting objArrays
1430 // to the given MemRegion. Returns the amount of words actually scanned.
1431 size_t scan_for_references(oop const obj, MemRegion mr) {
1432 size_t const obj_size = obj->size();
1433 // All non-objArrays and objArrays completely within the mr
1434 // can be scanned without passing the mr.
1435 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
1436 obj->oop_iterate(&_update_cl);
1437 return obj_size;
1438 }
1439 // This path is for objArrays crossing the given MemRegion. Only scan the
1440 // area within the MemRegion.
1441 obj->oop_iterate(&_update_cl, mr);
1442 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
1443 }
1444
1445 // A humongous object is live (with respect to the scanning) either
1446 // a) it is marked on the bitmap as such
1447 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
1448 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
1449 return bitmap->is_marked(humongous_obj) || (tars > tams);
1450 }
1451
1452 // Iterator over the live objects within the given MemRegion.
1453 class LiveObjIterator : public StackObj {
1454 const G1CMBitMap* const _bitmap;
1455 const HeapWord* _tams;
1456 const MemRegion _mr;
1457 HeapWord* _current;
1458
1459 bool is_below_tams() const {
1460 return _current < _tams;
1461 }
1462
1463 bool is_live(HeapWord* obj) const {
1464 return !is_below_tams() || _bitmap->is_marked(obj);
1465 }
1466
1467 HeapWord* bitmap_limit() const {
1468 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
1469 }
1470
1471 void move_if_below_tams() {
1472 if (is_below_tams() && has_next()) {
1473 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
1474 }
1475 }
1476 public:
1477 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
1478 _bitmap(bitmap),
1479 _tams(tams),
1480 _mr(mr),
1481 _current(first_oop_into_mr) {
1482
1483 assert(_current <= _mr.start(),
1484 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
1485 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
1486
1487 // Step to the next live object within the MemRegion if needed.
1488 if (is_live(_current)) {
1489 // Non-objArrays were scanned by the previous part of that region.
1490 if (_current < mr.start() && !oop(_current)->is_objArray()) {
1491 _current += oop(_current)->size();
1492 // We might have positioned _current on a non-live object. Reposition to the next
1493 // live one if needed.
1494 move_if_below_tams();
1495 }
1496 } else {
1497 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
1498 // immediately.
1499 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
1500 assert(_current == _mr.end() || is_live(_current),
1501 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
1502 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
1503 }
1504 }
1505
1506 void move_to_next() {
1507 _current += next()->size();
1508 move_if_below_tams();
1509 }
1510
1511 oop next() const {
1512 oop result = oop(_current);
1513 assert(is_live(_current),
1514 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
1515 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
1516 return result;
1517 }
1518
1519 bool has_next() const {
1520 return _current < _mr.end();
1521 }
1522 };
1523
1524 // Rebuild remembered sets in the part of the region specified by mr and hr.
1525 // Objects between the bottom of the region and the TAMS are checked for liveness
1526 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
1527 // Returns the number of live words between bottom and TAMS.
1528 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
1529 HeapWord* const top_at_mark_start,
1530 HeapWord* const top_at_rebuild_start,
1531 HeapRegion* hr,
1532 MemRegion mr) {
1533 size_t marked_words = 0;
1534
1535 if (hr->is_humongous()) {
1536 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
1537 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
1538 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
1539 // however in case of humongous objects it is sufficient to scan the encompassing
1540 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
1541 // two areas will be zero sized. I.e. TAMS is either
1542 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
1543 // value: this would mean that TAMS points somewhere into the object.
1544 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
1545 "More than one object in the humongous region?");
1546 humongous_obj->oop_iterate(&_update_cl, mr);
1547 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
1548 } else {
1549 return 0;
1550 }
1551 }
1552
1553 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
1554 oop obj = it.next();
1555 size_t scanned_size = scan_for_references(obj, mr);
1556 if ((HeapWord*)obj < top_at_mark_start) {
1557 marked_words += scanned_size;
1558 }
1559 }
1560
1561 return marked_words * HeapWordSize;
1562 }
1563 public:
1564 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
1565 G1ConcurrentMark* cm,
1566 uint worker_id) :
1567 HeapRegionClosure(),
1568 _cm(cm),
1569 _update_cl(g1h, worker_id) { }
1570
1571 bool do_heap_region(HeapRegion* hr) {
1572 if (_cm->has_aborted()) {
1573 return true;
1574 }
1575
1576 uint const region_idx = hr->hrm_index();
1577 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
1578 assert(top_at_rebuild_start_check == NULL ||
1579 top_at_rebuild_start_check > hr->bottom(),
1580 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
1581 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
1582
1583 size_t total_marked_bytes = 0;
1584 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
1585
1586 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
1587
1588 HeapWord* cur = hr->bottom();
1589 while (cur < hr->end()) {
1590 // After every iteration (yield point) we need to check whether the region's
1591 // TARS changed due to e.g. eager reclaim.
1592 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
1593 if (top_at_rebuild_start == NULL) {
1594 return false;
1595 }
1596
1597 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
1598 if (next_chunk.is_empty()) {
1599 break;
1600 }
1601
1602 const Ticks start = Ticks::now();
1603 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
1604 top_at_mark_start,
1605 top_at_rebuild_start,
1606 hr,
1607 next_chunk);
1608 Tickspan time = Ticks::now() - start;
1609
1610 log_trace(gc, remset, tracking)("Rebuilt region %u "
1611 "live " SIZE_FORMAT " "
1612 "time %.3fms "
1613 "marked bytes " SIZE_FORMAT " "
1614 "bot " PTR_FORMAT " "
1615 "TAMS " PTR_FORMAT " "
1616 "TARS " PTR_FORMAT,
1617 region_idx,
1618 _cm->liveness(region_idx) * HeapWordSize,
1619 time.seconds() * 1000.0,
1620 marked_bytes,
1621 p2i(hr->bottom()),
1622 p2i(top_at_mark_start),
1623 p2i(top_at_rebuild_start));
1624
1625 if (marked_bytes > 0) {
1626 total_marked_bytes += marked_bytes;
1627 }
1628 cur += chunk_size_in_words;
1629
1630 _cm->do_yield_check();
1631 if (_cm->has_aborted()) {
1632 return true;
1633 }
1634 }
1635 // In the final iteration of the loop the region might have been eagerly reclaimed.
1636 // Simply filter out those regions. We can not just use region type because there
1637 // might have already been new allocations into these regions.
1638 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
1639 assert(top_at_rebuild_start == NULL ||
1640 total_marked_bytes == hr->marked_bytes(),
1641 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
1642 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
1643 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
1644 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
1645 // Abort state may have changed after the yield check.
1646 return _cm->has_aborted();
1647 }
1648 };
1649
1650 HeapRegionClaimer _hr_claimer;
1651 G1ConcurrentMark* _cm;
1652
1653 uint _worker_id_offset;
1654 public:
1655 G1RebuildRemSetTask(G1ConcurrentMark* cm,
1656 uint n_workers,
1657 uint worker_id_offset) :
1658 AbstractGangTask("G1 Rebuild Remembered Set"),
1659 _hr_claimer(n_workers),
1660 _cm(cm),
1661 _worker_id_offset(worker_id_offset) {
1662 }
1663
1664 void work(uint worker_id) {
1665 SuspendibleThreadSetJoiner sts_join;
1666
1667 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1668
1669 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1670 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1671 }
1672 };
1673
1674 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1675 WorkGang* workers,
1676 uint worker_id_offset) {
1677 uint num_workers = workers->active_workers();
1678
1679 G1RebuildRemSetTask cl(cm,
1680 num_workers,
1681 worker_id_offset);
1682 workers->run_task(&cl, num_workers);
1683 }