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