1/*
2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25#include "precompiled.hpp"
26#include "gc/g1/g1BarrierSet.hpp"
27#include "gc/g1/g1BlockOffsetTable.inline.hpp"
28#include "gc/g1/g1CardTable.inline.hpp"
29#include "gc/g1/g1CollectedHeap.inline.hpp"
30#include "gc/g1/g1ConcurrentRefine.hpp"
31#include "gc/g1/g1DirtyCardQueue.hpp"
32#include "gc/g1/g1FromCardCache.hpp"
33#include "gc/g1/g1GCPhaseTimes.hpp"
34#include "gc/g1/g1HotCardCache.hpp"
35#include "gc/g1/g1OopClosures.inline.hpp"
36#include "gc/g1/g1RootClosures.hpp"
37#include "gc/g1/g1RemSet.hpp"
38#include "gc/g1/g1SharedDirtyCardQueue.hpp"
39#include "gc/g1/heapRegion.inline.hpp"
40#include "gc/g1/heapRegionManager.inline.hpp"
41#include "gc/g1/heapRegionRemSet.hpp"
42#include "gc/shared/gcTraceTime.inline.hpp"
43#include "gc/shared/suspendibleThreadSet.hpp"
44#include "jfr/jfrEvents.hpp"
45#include "memory/iterator.hpp"
46#include "memory/resourceArea.hpp"
47#include "oops/access.inline.hpp"
48#include "oops/oop.inline.hpp"
49#include "runtime/os.hpp"
50#include "utilities/align.hpp"
51#include "utilities/globalDefinitions.hpp"
52#include "utilities/stack.inline.hpp"
53#include "utilities/ticks.hpp"
54
55// Collects information about the overall remembered set scan progress during an evacuation.
56class G1RemSetScanState : public CHeapObj<mtGC> {
57private:
58 class G1ClearCardTableTask : public AbstractGangTask {
59 G1CollectedHeap* _g1h;
60 uint* _dirty_region_list;
61 size_t _num_dirty_regions;
62 size_t _chunk_length;
63
64 size_t volatile _cur_dirty_regions;
65 public:
66 G1ClearCardTableTask(G1CollectedHeap* g1h,
67 uint* dirty_region_list,
68 size_t num_dirty_regions,
69 size_t chunk_length) :
70 AbstractGangTask("G1 Clear Card Table Task"),
71 _g1h(g1h),
72 _dirty_region_list(dirty_region_list),
73 _num_dirty_regions(num_dirty_regions),
74 _chunk_length(chunk_length),
75 _cur_dirty_regions(0) {
76
77 assert(chunk_length > 0, "must be");
78 }
79
80 static size_t chunk_size() { return M; }
81
82 void work(uint worker_id) {
83 while (_cur_dirty_regions < _num_dirty_regions) {
84 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
85 size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
86
87 for (size_t i = next; i < max; i++) {
88 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
89 if (!r->is_survivor()) {
90 r->clear_cardtable();
91 }
92 }
93 }
94 }
95 };
96
97 size_t _max_regions;
98
99 // Scan progress for the remembered set of a single region. Transitions from
100 // Unclaimed -> Claimed -> Complete.
101 // At each of the transitions the thread that does the transition needs to perform
102 // some special action once. This is the reason for the extra "Claimed" state.
103 typedef jint G1RemsetIterState;
104
105 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
106 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned.
107 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned.
108
109 G1RemsetIterState volatile* _iter_states;
110 // The current location where the next thread should continue scanning in a region's
111 // remembered set.
112 size_t volatile* _iter_claims;
113
114 // Temporary buffer holding the regions we used to store remembered set scan duplicate
115 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
116 uint* _dirty_region_buffer;
117
118 // Flag for every region whether it is in the _dirty_region_buffer already
119 // to avoid duplicates.
120 bool volatile* _in_dirty_region_buffer;
121 size_t _cur_dirty_region;
122
123 // Creates a snapshot of the current _top values at the start of collection to
124 // filter out card marks that we do not want to scan.
125 class G1ResetScanTopClosure : public HeapRegionClosure {
126 private:
127 HeapWord** _scan_top;
128 public:
129 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
130
131 virtual bool do_heap_region(HeapRegion* r) {
132 uint hrm_index = r->hrm_index();
133 if (!r->in_collection_set() && r->is_old_or_humongous_or_archive() && !r->is_empty()) {
134 _scan_top[hrm_index] = r->top();
135 } else {
136 _scan_top[hrm_index] = NULL;
137 }
138 return false;
139 }
140 };
141
142 // For each region, contains the maximum top() value to be used during this garbage
143 // collection. Subsumes common checks like filtering out everything but old and
144 // humongous regions outside the collection set.
145 // This is valid because we are not interested in scanning stray remembered set
146 // entries from free or archive regions.
147 HeapWord** _scan_top;
148public:
149 G1RemSetScanState() :
150 _max_regions(0),
151 _iter_states(NULL),
152 _iter_claims(NULL),
153 _dirty_region_buffer(NULL),
154 _in_dirty_region_buffer(NULL),
155 _cur_dirty_region(0),
156 _scan_top(NULL) {
157 }
158
159 ~G1RemSetScanState() {
160 if (_iter_states != NULL) {
161 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
162 }
163 if (_iter_claims != NULL) {
164 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
165 }
166 if (_dirty_region_buffer != NULL) {
167 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
168 }
169 if (_in_dirty_region_buffer != NULL) {
170 FREE_C_HEAP_ARRAY(bool, _in_dirty_region_buffer);
171 }
172 if (_scan_top != NULL) {
173 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
174 }
175 }
176
177 void initialize(uint max_regions) {
178 assert(_iter_states == NULL, "Must not be initialized twice");
179 assert(_iter_claims == NULL, "Must not be initialized twice");
180 _max_regions = max_regions;
181 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
182 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
183 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
184 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(bool, max_regions, mtGC);
185 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
186 }
187
188 void reset() {
189 for (uint i = 0; i < _max_regions; i++) {
190 _iter_states[i] = Unclaimed;
191 clear_scan_top(i);
192 }
193
194 G1ResetScanTopClosure cl(_scan_top);
195 G1CollectedHeap::heap()->heap_region_iterate(&cl);
196
197 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
198 memset((void*)_in_dirty_region_buffer, false, _max_regions * sizeof(bool));
199 _cur_dirty_region = 0;
200 }
201
202 // Attempt to claim the remembered set of the region for iteration. Returns true
203 // if this call caused the transition from Unclaimed to Claimed.
204 inline bool claim_iter(uint region) {
205 assert(region < _max_regions, "Tried to access invalid region %u", region);
206 if (_iter_states[region] != Unclaimed) {
207 return false;
208 }
209 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
210 return (res == Unclaimed);
211 }
212
213 // Try to atomically sets the iteration state to "complete". Returns true for the
214 // thread that caused the transition.
215 inline bool set_iter_complete(uint region) {
216 if (iter_is_complete(region)) {
217 return false;
218 }
219 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
220 return (res == Claimed);
221 }
222
223 // Returns true if the region's iteration is complete.
224 inline bool iter_is_complete(uint region) const {
225 assert(region < _max_regions, "Tried to access invalid region %u", region);
226 return _iter_states[region] == Complete;
227 }
228
229 // The current position within the remembered set of the given region.
230 inline size_t iter_claimed(uint region) const {
231 assert(region < _max_regions, "Tried to access invalid region %u", region);
232 return _iter_claims[region];
233 }
234
235 // Claim the next block of cards within the remembered set of the region with
236 // step size.
237 inline size_t iter_claimed_next(uint region, size_t step) {
238 return Atomic::add(step, &_iter_claims[region]) - step;
239 }
240
241 void add_dirty_region(uint region) {
242 if (_in_dirty_region_buffer[region]) {
243 return;
244 }
245
246 if (!Atomic::cmpxchg(true, &_in_dirty_region_buffer[region], false)) {
247 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
248 _dirty_region_buffer[allocated] = region;
249 }
250 }
251
252 HeapWord* scan_top(uint region_idx) const {
253 return _scan_top[region_idx];
254 }
255
256 void clear_scan_top(uint region_idx) {
257 _scan_top[region_idx] = NULL;
258 }
259
260 // Clear the card table of "dirty" regions.
261 void clear_card_table(WorkGang* workers) {
262 if (_cur_dirty_region == 0) {
263 return;
264 }
265
266 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
267 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
268 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
269
270 // Iterate over the dirty cards region list.
271 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
272
273 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
274 "units of work for " SIZE_FORMAT " regions.",
275 cl.name(), num_workers, num_chunks, _cur_dirty_region);
276 workers->run_task(&cl, num_workers);
277
278#ifndef PRODUCT
279 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
280#endif
281 }
282};
283
284G1RemSet::G1RemSet(G1CollectedHeap* g1h,
285 G1CardTable* ct,
286 G1HotCardCache* hot_card_cache) :
287 _scan_state(new G1RemSetScanState()),
288 _prev_period_summary(),
289 _g1h(g1h),
290 _num_conc_refined_cards(0),
291 _ct(ct),
292 _g1p(_g1h->policy()),
293 _hot_card_cache(hot_card_cache) {
294}
295
296G1RemSet::~G1RemSet() {
297 if (_scan_state != NULL) {
298 delete _scan_state;
299 }
300}
301
302uint G1RemSet::num_par_rem_sets() {
303 return G1DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
304}
305
306void G1RemSet::initialize(size_t capacity, uint max_regions) {
307 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
308 _scan_state->initialize(max_regions);
309}
310
311class G1ScanRSForRegionClosure : public HeapRegionClosure {
312 G1CollectedHeap* _g1h;
313 G1CardTable *_ct;
314
315 G1ParScanThreadState* _pss;
316 G1ScanCardClosure* _scan_objs_on_card_cl;
317
318 G1RemSetScanState* _scan_state;
319
320 G1GCPhaseTimes::GCParPhases _phase;
321
322 uint _worker_i;
323
324 size_t _opt_refs_scanned;
325 size_t _opt_refs_memory_used;
326
327 size_t _cards_scanned;
328 size_t _cards_claimed;
329 size_t _cards_skipped;
330
331 Tickspan _rem_set_root_scan_time;
332 Tickspan _rem_set_trim_partially_time;
333
334 Tickspan _strong_code_root_scan_time;
335 Tickspan _strong_code_trim_partially_time;
336
337 void claim_card(size_t card_index, const uint region_idx_for_card) {
338 _ct->set_card_claimed(card_index);
339 _scan_state->add_dirty_region(region_idx_for_card);
340 }
341
342 void scan_card(MemRegion mr, uint region_idx_for_card) {
343 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
344 assert(!card_region->is_young(), "Should not scan card in young region %u", region_idx_for_card);
345 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
346 _scan_objs_on_card_cl->trim_queue_partially();
347 _cards_scanned++;
348 }
349
350 void scan_opt_rem_set_roots(HeapRegion* r) {
351 EventGCPhaseParallel event;
352
353 G1OopStarChunkedList* opt_rem_set_list = _pss->oops_into_optional_region(r);
354
355 G1ScanCardClosure scan_cl(_g1h, _pss);
356 G1ScanRSForOptionalClosure cl(_g1h, &scan_cl);
357 _opt_refs_scanned += opt_rem_set_list->oops_do(&cl, _pss->closures()->raw_strong_oops());
358 _opt_refs_memory_used += opt_rem_set_list->used_memory();
359
360 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
361 }
362
363 void scan_rem_set_roots(HeapRegion* r) {
364 EventGCPhaseParallel event;
365 uint const region_idx = r->hrm_index();
366
367 if (_scan_state->claim_iter(region_idx)) {
368 // If we ever free the collection set concurrently, we should also
369 // clear the card table concurrently therefore we won't need to
370 // add regions of the collection set to the dirty cards region.
371 _scan_state->add_dirty_region(region_idx);
372 }
373
374 if (r->rem_set()->cardset_is_empty()) {
375 return;
376 }
377
378 // We claim cards in blocks so as to reduce the contention.
379 size_t const block_size = G1RSetScanBlockSize;
380
381 HeapRegionRemSetIterator iter(r->rem_set());
382 size_t card_index;
383
384 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
385 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
386 if (current_card >= claimed_card_block + block_size) {
387 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
388 }
389 if (current_card < claimed_card_block) {
390 _cards_skipped++;
391 continue;
392 }
393 _cards_claimed++;
394
395 HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
396 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
397
398#ifdef ASSERT
399 HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
400 assert(hr == NULL || hr->is_in_reserved(card_start),
401 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
402#endif
403 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
404 if (card_start >= top) {
405 continue;
406 }
407
408 // If the card is dirty, then G1 will scan it during Update RS.
409 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
410 continue;
411 }
412
413 // We claim lazily (so races are possible but they're benign), which reduces the
414 // number of duplicate scans (the rsets of the regions in the cset can intersect).
415 // Claim the card after checking bounds above: the remembered set may contain
416 // random cards into current survivor, and we would then have an incorrectly
417 // claimed card in survivor space. Card table clear does not reset the card table
418 // of survivor space regions.
419 claim_card(card_index, region_idx_for_card);
420
421 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
422
423 scan_card(mr, region_idx_for_card);
424 }
425 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
426 }
427
428 void scan_strong_code_roots(HeapRegion* r) {
429 EventGCPhaseParallel event;
430 // We pass a weak code blobs closure to the remembered set scanning because we want to avoid
431 // treating the nmethods visited to act as roots for concurrent marking.
432 // We only want to make sure that the oops in the nmethods are adjusted with regard to the
433 // objects copied by the current evacuation.
434 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
435 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
436 }
437
438public:
439 G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
440 G1ScanCardClosure* scan_obj_on_card,
441 G1ParScanThreadState* pss,
442 G1GCPhaseTimes::GCParPhases phase,
443 uint worker_i) :
444 _g1h(G1CollectedHeap::heap()),
445 _ct(_g1h->card_table()),
446 _pss(pss),
447 _scan_objs_on_card_cl(scan_obj_on_card),
448 _scan_state(scan_state),
449 _phase(phase),
450 _worker_i(worker_i),
451 _opt_refs_scanned(0),
452 _opt_refs_memory_used(0),
453 _cards_scanned(0),
454 _cards_claimed(0),
455 _cards_skipped(0),
456 _rem_set_root_scan_time(),
457 _rem_set_trim_partially_time(),
458 _strong_code_root_scan_time(),
459 _strong_code_trim_partially_time() { }
460
461 bool do_heap_region(HeapRegion* r) {
462 assert(r->in_collection_set(), "Region %u is not in the collection set.", r->hrm_index());
463 uint const region_idx = r->hrm_index();
464
465 // The individual references for the optional remembered set are per-worker, so we
466 // always need to scan them.
467 if (r->has_index_in_opt_cset()) {
468 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
469 scan_opt_rem_set_roots(r);
470 }
471
472 // Do an early out if we know we are complete.
473 if (_scan_state->iter_is_complete(region_idx)) {
474 return false;
475 }
476
477 {
478 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
479 scan_rem_set_roots(r);
480 }
481
482 if (_scan_state->set_iter_complete(region_idx)) {
483 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
484 // Scan the strong code root list attached to the current region
485 scan_strong_code_roots(r);
486 }
487 return false;
488 }
489
490 Tickspan rem_set_root_scan_time() const { return _rem_set_root_scan_time; }
491 Tickspan rem_set_trim_partially_time() const { return _rem_set_trim_partially_time; }
492
493 Tickspan strong_code_root_scan_time() const { return _strong_code_root_scan_time; }
494 Tickspan strong_code_root_trim_partially_time() const { return _strong_code_trim_partially_time; }
495
496 size_t cards_scanned() const { return _cards_scanned; }
497 size_t cards_claimed() const { return _cards_claimed; }
498 size_t cards_skipped() const { return _cards_skipped; }
499
500 size_t opt_refs_scanned() const { return _opt_refs_scanned; }
501 size_t opt_refs_memory_used() const { return _opt_refs_memory_used; }
502};
503
504void G1RemSet::scan_rem_set(G1ParScanThreadState* pss,
505 uint worker_i,
506 G1GCPhaseTimes::GCParPhases scan_phase,
507 G1GCPhaseTimes::GCParPhases objcopy_phase,
508 G1GCPhaseTimes::GCParPhases coderoots_phase) {
509 assert(pss->trim_ticks().value() == 0, "Queues must have been trimmed before entering.");
510
511 G1ScanCardClosure scan_cl(_g1h, pss);
512 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, scan_phase, worker_i);
513 _g1h->collection_set_iterate_increment_from(&cl, worker_i);
514
515 G1GCPhaseTimes* p = _g1p->phase_times();
516
517 p->record_or_add_time_secs(objcopy_phase, worker_i, cl.rem_set_trim_partially_time().seconds());
518
519 p->record_or_add_time_secs(scan_phase, worker_i, cl.rem_set_root_scan_time().seconds());
520 p->record_or_add_thread_work_item(scan_phase, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
521 p->record_or_add_thread_work_item(scan_phase, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
522 p->record_or_add_thread_work_item(scan_phase, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
523 // At this time we only record some metrics for the optional remembered set.
524 if (scan_phase == G1GCPhaseTimes::OptScanRS) {
525 p->record_or_add_thread_work_item(scan_phase, worker_i, cl.opt_refs_scanned(), G1GCPhaseTimes::ScanRSScannedOptRefs);
526 p->record_or_add_thread_work_item(scan_phase, worker_i, cl.opt_refs_memory_used(), G1GCPhaseTimes::ScanRSUsedMemory);
527 }
528
529 p->record_or_add_time_secs(coderoots_phase, worker_i, cl.strong_code_root_scan_time().seconds());
530 p->add_time_secs(objcopy_phase, worker_i, cl.strong_code_root_trim_partially_time().seconds());
531}
532
533// Closure used for updating rem sets. Only called during an evacuation pause.
534class G1RefineCardClosure: public G1CardTableEntryClosure {
535 G1RemSet* _g1rs;
536 G1ScanCardClosure* _update_rs_cl;
537
538 size_t _cards_scanned;
539 size_t _cards_skipped;
540public:
541 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanCardClosure* update_rs_cl) :
542 _g1rs(g1h->rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
543 {}
544
545 bool do_card_ptr(CardValue* card_ptr, uint worker_i) {
546 // The only time we care about recording cards that
547 // contain references that point into the collection set
548 // is during RSet updating within an evacuation pause.
549 // In this case worker_i should be the id of a GC worker thread.
550 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
551
552 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
553
554 if (card_scanned) {
555 _update_rs_cl->trim_queue_partially();
556 _cards_scanned++;
557 } else {
558 _cards_skipped++;
559 }
560 return true;
561 }
562
563 size_t cards_scanned() const { return _cards_scanned; }
564 size_t cards_skipped() const { return _cards_skipped; }
565};
566
567void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
568 G1GCPhaseTimes* p = _g1p->phase_times();
569
570 // Apply closure to log entries in the HCC.
571 if (G1HotCardCache::default_use_cache()) {
572 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
573
574 G1ScanCardClosure scan_hcc_cl(_g1h, pss);
575 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
576 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
577 }
578
579 // Now apply the closure to all remaining log entries.
580 {
581 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
582
583 G1ScanCardClosure update_rs_cl(_g1h, pss);
584 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
585 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
586
587 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
588 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
589 }
590}
591
592void G1RemSet::prepare_for_scan_rem_set() {
593 G1BarrierSet::dirty_card_queue_set().concatenate_logs();
594 _scan_state->reset();
595}
596
597void G1RemSet::prepare_for_scan_rem_set(uint region_idx) {
598 _scan_state->clear_scan_top(region_idx);
599}
600
601void G1RemSet::cleanup_after_scan_rem_set() {
602 G1GCPhaseTimes* phase_times = _g1h->phase_times();
603
604 // Set all cards back to clean.
605 double start = os::elapsedTime();
606 _scan_state->clear_card_table(_g1h->workers());
607 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
608}
609
610inline void check_card_ptr(CardTable::CardValue* card_ptr, G1CardTable* ct) {
611#ifdef ASSERT
612 G1CollectedHeap* g1h = G1CollectedHeap::heap();
613 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
614 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
615 p2i(card_ptr),
616 ct->index_for(ct->addr_for(card_ptr)),
617 p2i(ct->addr_for(card_ptr)),
618 g1h->addr_to_region(ct->addr_for(card_ptr)));
619#endif
620}
621
622void G1RemSet::refine_card_concurrently(CardValue* card_ptr,
623 uint worker_i) {
624 assert(!_g1h->is_gc_active(), "Only call concurrently");
625
626 // Construct the region representing the card.
627 HeapWord* start = _ct->addr_for(card_ptr);
628 // And find the region containing it.
629 HeapRegion* r = _g1h->heap_region_containing_or_null(start);
630
631 // If this is a (stale) card into an uncommitted region, exit.
632 if (r == NULL) {
633 return;
634 }
635
636 check_card_ptr(card_ptr, _ct);
637
638 // If the card is no longer dirty, nothing to do.
639 if (*card_ptr != G1CardTable::dirty_card_val()) {
640 return;
641 }
642
643 // This check is needed for some uncommon cases where we should
644 // ignore the card.
645 //
646 // The region could be young. Cards for young regions are
647 // distinctly marked (set to g1_young_gen), so the post-barrier will
648 // filter them out. However, that marking is performed
649 // concurrently. A write to a young object could occur before the
650 // card has been marked young, slipping past the filter.
651 //
652 // The card could be stale, because the region has been freed since
653 // the card was recorded. In this case the region type could be
654 // anything. If (still) free or (reallocated) young, just ignore
655 // it. If (reallocated) old or humongous, the later card trimming
656 // and additional checks in iteration may detect staleness. At
657 // worst, we end up processing a stale card unnecessarily.
658 //
659 // In the normal (non-stale) case, the synchronization between the
660 // enqueueing of the card and processing it here will have ensured
661 // we see the up-to-date region type here.
662 if (!r->is_old_or_humongous_or_archive()) {
663 return;
664 }
665
666 // The result from the hot card cache insert call is either:
667 // * pointer to the current card
668 // (implying that the current card is not 'hot'),
669 // * null
670 // (meaning we had inserted the card ptr into the "hot" card cache,
671 // which had some headroom),
672 // * a pointer to a "hot" card that was evicted from the "hot" cache.
673 //
674
675 if (_hot_card_cache->use_cache()) {
676 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
677
678 const CardValue* orig_card_ptr = card_ptr;
679 card_ptr = _hot_card_cache->insert(card_ptr);
680 if (card_ptr == NULL) {
681 // There was no eviction. Nothing to do.
682 return;
683 } else if (card_ptr != orig_card_ptr) {
684 // Original card was inserted and an old card was evicted.
685 start = _ct->addr_for(card_ptr);
686 r = _g1h->heap_region_containing(start);
687
688 // Check whether the region formerly in the cache should be
689 // ignored, as discussed earlier for the original card. The
690 // region could have been freed while in the cache.
691 if (!r->is_old_or_humongous_or_archive()) {
692 return;
693 }
694 } // Else we still have the original card.
695 }
696
697 // Trim the region designated by the card to what's been allocated
698 // in the region. The card could be stale, or the card could cover
699 // (part of) an object at the end of the allocated space and extend
700 // beyond the end of allocation.
701
702 // Non-humongous objects are only allocated in the old-gen during
703 // GC, so if region is old then top is stable. Humongous object
704 // allocation sets top last; if top has not yet been set, this is
705 // a stale card and we'll end up with an empty intersection. If
706 // this is not a stale card, the synchronization between the
707 // enqueuing of the card and processing it here will have ensured
708 // we see the up-to-date top here.
709 HeapWord* scan_limit = r->top();
710
711 if (scan_limit <= start) {
712 // If the trimmed region is empty, the card must be stale.
713 return;
714 }
715
716 // Okay to clean and process the card now. There are still some
717 // stale card cases that may be detected by iteration and dealt with
718 // as iteration failure.
719 *const_cast<volatile CardValue*>(card_ptr) = G1CardTable::clean_card_val();
720
721 // This fence serves two purposes. First, the card must be cleaned
722 // before processing the contents. Second, we can't proceed with
723 // processing until after the read of top, for synchronization with
724 // possibly concurrent humongous object allocation. It's okay that
725 // reading top and reading type were racy wrto each other. We need
726 // both set, in any order, to proceed.
727 OrderAccess::fence();
728
729 // Don't use addr_for(card_ptr + 1) which can ask for
730 // a card beyond the heap.
731 HeapWord* end = start + G1CardTable::card_size_in_words;
732 MemRegion dirty_region(start, MIN2(scan_limit, end));
733 assert(!dirty_region.is_empty(), "sanity");
734
735 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
736 if (r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl)) {
737 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
738 return;
739 }
740
741 // If unable to process the card then we encountered an unparsable
742 // part of the heap (e.g. a partially allocated object, so only
743 // temporarily a problem) while processing a stale card. Despite
744 // the card being stale, we can't simply ignore it, because we've
745 // already marked the card cleaned, so taken responsibility for
746 // ensuring the card gets scanned.
747 //
748 // However, the card might have gotten re-dirtied and re-enqueued
749 // while we worked. (In fact, it's pretty likely.)
750 if (*card_ptr == G1CardTable::dirty_card_val()) {
751 return;
752 }
753
754 // Re-dirty the card and enqueue in the *shared* queue. Can't use
755 // the thread-local queue, because that might be the queue that is
756 // being processed by us; we could be a Java thread conscripted to
757 // perform refinement on our queue's current buffer.
758 *card_ptr = G1CardTable::dirty_card_val();
759 G1BarrierSet::shared_dirty_card_queue().enqueue(card_ptr);
760}
761
762bool G1RemSet::refine_card_during_gc(CardValue* card_ptr,
763 G1ScanCardClosure* update_rs_cl) {
764 assert(_g1h->is_gc_active(), "Only call during GC");
765
766 // Construct the region representing the card.
767 HeapWord* card_start = _ct->addr_for(card_ptr);
768 // And find the region containing it.
769 uint const card_region_idx = _g1h->addr_to_region(card_start);
770
771 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
772 if (scan_limit == NULL) {
773 // This is a card into an uncommitted region. We need to bail out early as we
774 // should not access the corresponding card table entry.
775 return false;
776 }
777
778 check_card_ptr(card_ptr, _ct);
779
780 // If the card is no longer dirty, nothing to do. This covers cards that were already
781 // scanned as parts of the remembered sets.
782 if (*card_ptr != G1CardTable::dirty_card_val()) {
783 return false;
784 }
785
786 // We claim lazily (so races are possible but they're benign), which reduces the
787 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
788 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
789
790 _scan_state->add_dirty_region(card_region_idx);
791 if (scan_limit <= card_start) {
792 // If the card starts above the area in the region containing objects to scan, skip it.
793 return false;
794 }
795
796 // Don't use addr_for(card_ptr + 1) which can ask for
797 // a card beyond the heap.
798 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
799 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
800 assert(!dirty_region.is_empty(), "sanity");
801
802 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
803 assert(!card_region->is_young(), "Should not scan card in young region %u", card_region_idx);
804 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
805 assert(card_processed, "must be");
806 return true;
807}
808
809void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
810 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
811 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
812
813 G1RemSetSummary current(this);
814 _prev_period_summary.subtract_from(&current);
815
816 Log(gc, remset) log;
817 log.trace("%s", header);
818 ResourceMark rm;
819 LogStream ls(log.trace());
820 _prev_period_summary.print_on(&ls);
821
822 _prev_period_summary.set(&current);
823 }
824}
825
826void G1RemSet::print_summary_info() {
827 Log(gc, remset, exit) log;
828 if (log.is_trace()) {
829 log.trace(" Cumulative RS summary");
830 G1RemSetSummary current(this);
831 ResourceMark rm;
832 LogStream ls(log.trace());
833 current.print_on(&ls);
834 }
835}
836
837class G1RebuildRemSetTask: public AbstractGangTask {
838 // Aggregate the counting data that was constructed concurrently
839 // with marking.
840 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
841 G1ConcurrentMark* _cm;
842 G1RebuildRemSetClosure _update_cl;
843
844 // Applies _update_cl to the references of the given object, limiting objArrays
845 // to the given MemRegion. Returns the amount of words actually scanned.
846 size_t scan_for_references(oop const obj, MemRegion mr) {
847 size_t const obj_size = obj->size();
848 // All non-objArrays and objArrays completely within the mr
849 // can be scanned without passing the mr.
850 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
851 obj->oop_iterate(&_update_cl);
852 return obj_size;
853 }
854 // This path is for objArrays crossing the given MemRegion. Only scan the
855 // area within the MemRegion.
856 obj->oop_iterate(&_update_cl, mr);
857 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
858 }
859
860 // A humongous object is live (with respect to the scanning) either
861 // a) it is marked on the bitmap as such
862 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
863 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
864 return bitmap->is_marked(humongous_obj) || (tars > tams);
865 }
866
867 // Iterator over the live objects within the given MemRegion.
868 class LiveObjIterator : public StackObj {
869 const G1CMBitMap* const _bitmap;
870 const HeapWord* _tams;
871 const MemRegion _mr;
872 HeapWord* _current;
873
874 bool is_below_tams() const {
875 return _current < _tams;
876 }
877
878 bool is_live(HeapWord* obj) const {
879 return !is_below_tams() || _bitmap->is_marked(obj);
880 }
881
882 HeapWord* bitmap_limit() const {
883 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
884 }
885
886 void move_if_below_tams() {
887 if (is_below_tams() && has_next()) {
888 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
889 }
890 }
891 public:
892 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
893 _bitmap(bitmap),
894 _tams(tams),
895 _mr(mr),
896 _current(first_oop_into_mr) {
897
898 assert(_current <= _mr.start(),
899 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
900 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
901
902 // Step to the next live object within the MemRegion if needed.
903 if (is_live(_current)) {
904 // Non-objArrays were scanned by the previous part of that region.
905 if (_current < mr.start() && !oop(_current)->is_objArray()) {
906 _current += oop(_current)->size();
907 // We might have positioned _current on a non-live object. Reposition to the next
908 // live one if needed.
909 move_if_below_tams();
910 }
911 } else {
912 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
913 // immediately.
914 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
915 assert(_current == _mr.end() || is_live(_current),
916 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
917 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
918 }
919 }
920
921 void move_to_next() {
922 _current += next()->size();
923 move_if_below_tams();
924 }
925
926 oop next() const {
927 oop result = oop(_current);
928 assert(is_live(_current),
929 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
930 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
931 return result;
932 }
933
934 bool has_next() const {
935 return _current < _mr.end();
936 }
937 };
938
939 // Rebuild remembered sets in the part of the region specified by mr and hr.
940 // Objects between the bottom of the region and the TAMS are checked for liveness
941 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
942 // Returns the number of live words between bottom and TAMS.
943 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
944 HeapWord* const top_at_mark_start,
945 HeapWord* const top_at_rebuild_start,
946 HeapRegion* hr,
947 MemRegion mr) {
948 size_t marked_words = 0;
949
950 if (hr->is_humongous()) {
951 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
952 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
953 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
954 // however in case of humongous objects it is sufficient to scan the encompassing
955 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
956 // two areas will be zero sized. I.e. TAMS is either
957 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
958 // value: this would mean that TAMS points somewhere into the object.
959 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
960 "More than one object in the humongous region?");
961 humongous_obj->oop_iterate(&_update_cl, mr);
962 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
963 } else {
964 return 0;
965 }
966 }
967
968 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
969 oop obj = it.next();
970 size_t scanned_size = scan_for_references(obj, mr);
971 if ((HeapWord*)obj < top_at_mark_start) {
972 marked_words += scanned_size;
973 }
974 }
975
976 return marked_words * HeapWordSize;
977 }
978public:
979 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
980 G1ConcurrentMark* cm,
981 uint worker_id) :
982 HeapRegionClosure(),
983 _cm(cm),
984 _update_cl(g1h, worker_id) { }
985
986 bool do_heap_region(HeapRegion* hr) {
987 if (_cm->has_aborted()) {
988 return true;
989 }
990
991 uint const region_idx = hr->hrm_index();
992 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
993 assert(top_at_rebuild_start_check == NULL ||
994 top_at_rebuild_start_check > hr->bottom(),
995 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
996 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
997
998 size_t total_marked_bytes = 0;
999 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
1000
1001 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
1002
1003 HeapWord* cur = hr->bottom();
1004 while (cur < hr->end()) {
1005 // After every iteration (yield point) we need to check whether the region's
1006 // TARS changed due to e.g. eager reclaim.
1007 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
1008 if (top_at_rebuild_start == NULL) {
1009 return false;
1010 }
1011
1012 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
1013 if (next_chunk.is_empty()) {
1014 break;
1015 }
1016
1017 const Ticks start = Ticks::now();
1018 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
1019 top_at_mark_start,
1020 top_at_rebuild_start,
1021 hr,
1022 next_chunk);
1023 Tickspan time = Ticks::now() - start;
1024
1025 log_trace(gc, remset, tracking)("Rebuilt region %u "
1026 "live " SIZE_FORMAT " "
1027 "time %.3fms "
1028 "marked bytes " SIZE_FORMAT " "
1029 "bot " PTR_FORMAT " "
1030 "TAMS " PTR_FORMAT " "
1031 "TARS " PTR_FORMAT,
1032 region_idx,
1033 _cm->liveness(region_idx) * HeapWordSize,
1034 time.seconds() * 1000.0,
1035 marked_bytes,
1036 p2i(hr->bottom()),
1037 p2i(top_at_mark_start),
1038 p2i(top_at_rebuild_start));
1039
1040 if (marked_bytes > 0) {
1041 total_marked_bytes += marked_bytes;
1042 }
1043 cur += chunk_size_in_words;
1044
1045 _cm->do_yield_check();
1046 if (_cm->has_aborted()) {
1047 return true;
1048 }
1049 }
1050 // In the final iteration of the loop the region might have been eagerly reclaimed.
1051 // Simply filter out those regions. We can not just use region type because there
1052 // might have already been new allocations into these regions.
1053 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
1054 assert(top_at_rebuild_start == NULL ||
1055 total_marked_bytes == hr->marked_bytes(),
1056 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
1057 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
1058 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
1059 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
1060 // Abort state may have changed after the yield check.
1061 return _cm->has_aborted();
1062 }
1063 };
1064
1065 HeapRegionClaimer _hr_claimer;
1066 G1ConcurrentMark* _cm;
1067
1068 uint _worker_id_offset;
1069public:
1070 G1RebuildRemSetTask(G1ConcurrentMark* cm,
1071 uint n_workers,
1072 uint worker_id_offset) :
1073 AbstractGangTask("G1 Rebuild Remembered Set"),
1074 _hr_claimer(n_workers),
1075 _cm(cm),
1076 _worker_id_offset(worker_id_offset) {
1077 }
1078
1079 void work(uint worker_id) {
1080 SuspendibleThreadSetJoiner sts_join;
1081
1082 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1083
1084 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1085 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1086 }
1087};
1088
1089void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1090 WorkGang* workers,
1091 uint worker_id_offset) {
1092 uint num_workers = workers->active_workers();
1093
1094 G1RebuildRemSetTask cl(cm,
1095 num_workers,
1096 worker_id_offset);
1097 workers->run_task(&cl, num_workers);
1098}
1099