1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/dirtyCardQueue.hpp"
27 #include "gc/g1/g1BarrierSet.hpp"
28 #include "gc/g1/g1BlockOffsetTable.inline.hpp"
29 #include "gc/g1/g1CardTable.inline.hpp"
30 #include "gc/g1/g1CollectedHeap.inline.hpp"
31 #include "gc/g1/g1ConcurrentRefine.hpp"
32 #include "gc/g1/g1FromCardCache.hpp"
33 #include "gc/g1/g1GCPhaseTimes.hpp"
34 #include "gc/g1/g1HotCardCache.hpp"
35 #include "gc/g1/g1OopClosures.inline.hpp"
36 #include "gc/g1/g1RootClosures.hpp"
37 #include "gc/g1/g1RemSet.hpp"
38 #include "gc/g1/heapRegion.inline.hpp"
39 #include "gc/g1/heapRegionManager.inline.hpp"
40 #include "gc/g1/heapRegionRemSet.hpp"
41 #include "gc/shared/gcTraceTime.inline.hpp"
42 #include "gc/shared/suspendibleThreadSet.hpp"
43 #include "jfr/jfrEvents.hpp"
44 #include "memory/iterator.hpp"
45 #include "memory/resourceArea.hpp"
46 #include "oops/access.inline.hpp"
47 #include "oops/oop.inline.hpp"
48 #include "runtime/os.hpp"
49 #include "utilities/align.hpp"
50 #include "utilities/globalDefinitions.hpp"
51 #include "utilities/intHisto.hpp"
52 #include "utilities/stack.inline.hpp"
53 #include "utilities/ticks.hpp"
54
55 // Collects information about the overall remembered set scan progress during an evacuation.
56 class G1RemSetScanState : public CHeapObj<mtGC> {
57 private:
58 class G1ClearCardTableTask : public AbstractGangTask {
59 G1CollectedHeap* _g1h;
60 uint* _dirty_region_list;
61 size_t _num_dirty_regions;
62 size_t _chunk_length;
63
64 size_t volatile _cur_dirty_regions;
65 public:
66 G1ClearCardTableTask(G1CollectedHeap* g1h,
67 uint* dirty_region_list,
68 size_t num_dirty_regions,
69 size_t chunk_length) :
70 AbstractGangTask("G1 Clear Card Table Task"),
71 _g1h(g1h),
72 _dirty_region_list(dirty_region_list),
73 _num_dirty_regions(num_dirty_regions),
74 _chunk_length(chunk_length),
75 _cur_dirty_regions(0) {
76
77 assert(chunk_length > 0, "must be");
78 }
79
80 static size_t chunk_size() { return M; }
81
82 void work(uint worker_id) {
83 while (_cur_dirty_regions < _num_dirty_regions) {
84 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
85 size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
86
87 for (size_t i = next; i < max; i++) {
88 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
89 if (!r->is_survivor()) {
90 r->clear_cardtable();
91 }
92 }
93 }
94 }
95 };
96
97 size_t _max_regions;
98
99 // Scan progress for the remembered set of a single region. Transitions from
100 // Unclaimed -> Claimed -> Complete.
101 // At each of the transitions the thread that does the transition needs to perform
102 // some special action once. This is the reason for the extra "Claimed" state.
103 typedef jint G1RemsetIterState;
104
105 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
106 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned.
107 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned.
108
109 G1RemsetIterState volatile* _iter_states;
110 // The current location where the next thread should continue scanning in a region's
111 // remembered set.
112 size_t volatile* _iter_claims;
113
114 // Temporary buffer holding the regions we used to store remembered set scan duplicate
115 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
116 uint* _dirty_region_buffer;
117
118 typedef jbyte IsDirtyRegionState;
119 static const IsDirtyRegionState Clean = 0;
120 static const IsDirtyRegionState Dirty = 1;
121 // Holds a flag for every region whether it is in the _dirty_region_buffer already
122 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools.
123 IsDirtyRegionState* _in_dirty_region_buffer;
124 size_t _cur_dirty_region;
125
126 // Creates a snapshot of the current _top values at the start of collection to
127 // filter out card marks that we do not want to scan.
128 class G1ResetScanTopClosure : public HeapRegionClosure {
129 private:
130 HeapWord** _scan_top;
131 public:
132 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
133
134 virtual bool do_heap_region(HeapRegion* r) {
135 uint hrm_index = r->hrm_index();
136 if (!r->in_collection_set() && r->is_old_or_humongous_or_archive()) {
137 _scan_top[hrm_index] = r->top();
138 } else {
139 _scan_top[hrm_index] = r->bottom();
140 }
141 return false;
142 }
143 };
144
145 // For each region, contains the maximum top() value to be used during this garbage
146 // collection. Subsumes common checks like filtering out everything but old and
147 // humongous regions outside the collection set.
148 // This is valid because we are not interested in scanning stray remembered set
149 // entries from free or archive regions.
150 HeapWord** _scan_top;
151 public:
152 G1RemSetScanState() :
153 _max_regions(0),
154 _iter_states(NULL),
155 _iter_claims(NULL),
156 _dirty_region_buffer(NULL),
157 _in_dirty_region_buffer(NULL),
158 _cur_dirty_region(0),
159 _scan_top(NULL) {
160 }
161
162 ~G1RemSetScanState() {
163 if (_iter_states != NULL) {
164 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
165 }
166 if (_iter_claims != NULL) {
167 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
168 }
169 if (_dirty_region_buffer != NULL) {
170 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
171 }
172 if (_in_dirty_region_buffer != NULL) {
173 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
174 }
175 if (_scan_top != NULL) {
176 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
177 }
178 }
179
180 void initialize(uint max_regions) {
181 assert(_iter_states == NULL, "Must not be initialized twice");
182 assert(_iter_claims == NULL, "Must not be initialized twice");
183 _max_regions = max_regions;
184 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
185 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
186 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
187 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
188 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
189 }
190
191 void reset() {
192 for (uint i = 0; i < _max_regions; i++) {
193 _iter_states[i] = Unclaimed;
194 }
195
196 G1ResetScanTopClosure cl(_scan_top);
197 G1CollectedHeap::heap()->heap_region_iterate(&cl);
198
199 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
200 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState));
201 _cur_dirty_region = 0;
202 }
203
204 // Attempt to claim the remembered set of the region for iteration. Returns true
205 // if this call caused the transition from Unclaimed to Claimed.
206 inline bool claim_iter(uint region) {
207 assert(region < _max_regions, "Tried to access invalid region %u", region);
208 if (_iter_states[region] != Unclaimed) {
209 return false;
210 }
211 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
212 return (res == Unclaimed);
213 }
214
215 // Try to atomically sets the iteration state to "complete". Returns true for the
216 // thread that caused the transition.
217 inline bool set_iter_complete(uint region) {
218 if (iter_is_complete(region)) {
219 return false;
220 }
221 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
222 return (res == Claimed);
223 }
224
225 // Returns true if the region's iteration is complete.
226 inline bool iter_is_complete(uint region) const {
227 assert(region < _max_regions, "Tried to access invalid region %u", region);
228 return _iter_states[region] == Complete;
229 }
230
231 // The current position within the remembered set of the given region.
232 inline size_t iter_claimed(uint region) const {
233 assert(region < _max_regions, "Tried to access invalid region %u", region);
234 return _iter_claims[region];
235 }
236
237 // Claim the next block of cards within the remembered set of the region with
238 // step size.
239 inline size_t iter_claimed_next(uint region, size_t step) {
240 return Atomic::add(step, &_iter_claims[region]) - step;
241 }
242
243 void add_dirty_region(uint region) {
244 if (_in_dirty_region_buffer[region] == Dirty) {
245 return;
246 }
247
248 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean;
249 if (marked_as_dirty) {
250 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
251 _dirty_region_buffer[allocated] = region;
252 }
253 }
254
255 HeapWord* scan_top(uint region_idx) const {
256 return _scan_top[region_idx];
257 }
258
259 // Clear the card table of "dirty" regions.
260 void clear_card_table(WorkGang* workers) {
261 if (_cur_dirty_region == 0) {
262 return;
263 }
264
265 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
266 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
267 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
268
269 // Iterate over the dirty cards region list.
270 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
271
272 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
273 "units of work for " SIZE_FORMAT " regions.",
274 cl.name(), num_workers, num_chunks, _cur_dirty_region);
275 workers->run_task(&cl, num_workers);
276
277 #ifndef PRODUCT
278 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
279 #endif
280 }
281 };
282
283 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
284 G1CardTable* ct,
285 G1HotCardCache* hot_card_cache) :
286 _scan_state(new G1RemSetScanState()),
287 _prev_period_summary(),
288 _g1h(g1h),
289 _num_conc_refined_cards(0),
290 _ct(ct),
291 _g1p(_g1h->g1_policy()),
292 _hot_card_cache(hot_card_cache) {
293 }
294
295 G1RemSet::~G1RemSet() {
296 if (_scan_state != NULL) {
297 delete _scan_state;
298 }
299 }
300
301 uint G1RemSet::num_par_rem_sets() {
302 return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
303 }
304
305 void G1RemSet::initialize(size_t capacity, uint max_regions) {
306 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
307 _scan_state->initialize(max_regions);
308 }
309
310 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
311 G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
312 G1ParScanThreadState* pss,
313 uint worker_i) :
314 _g1h(G1CollectedHeap::heap()),
315 _ct(_g1h->card_table()),
316 _pss(pss),
317 _scan_objs_on_card_cl(scan_obj_on_card),
318 _scan_state(scan_state),
319 _worker_i(worker_i),
320 _cards_scanned(0),
321 _cards_claimed(0),
322 _cards_skipped(0),
323 _rem_set_root_scan_time(),
324 _rem_set_trim_partially_time(),
325 _strong_code_root_scan_time(),
326 _strong_code_trim_partially_time() {
327 }
328
329 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
330 _ct->set_card_claimed(card_index);
331 _scan_state->add_dirty_region(region_idx_for_card);
332 }
333
334 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
335 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
336 _scan_objs_on_card_cl->set_region(card_region);
337 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
338 _scan_objs_on_card_cl->trim_queue_partially();
339 _cards_scanned++;
340 }
341
342 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
343 EventGCPhaseParallel event;
344 uint const region_idx = r->hrm_index();
345
346 if (_scan_state->claim_iter(region_idx)) {
347 // If we ever free the collection set concurrently, we should also
348 // clear the card table concurrently therefore we won't need to
349 // add regions of the collection set to the dirty cards region.
350 _scan_state->add_dirty_region(region_idx);
351 }
352
353 // We claim cards in blocks so as to reduce the contention.
354 size_t const block_size = G1RSetScanBlockSize;
355
356 HeapRegionRemSetIterator iter(r->rem_set());
357 size_t card_index;
358
359 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
360 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
361 if (current_card >= claimed_card_block + block_size) {
362 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
363 }
364 if (current_card < claimed_card_block) {
365 _cards_skipped++;
366 continue;
367 }
368 _cards_claimed++;
369
370 // If the card is dirty, then G1 will scan it during Update RS.
371 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
372 continue;
373 }
374
375 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index);
376 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
377
378 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start),
379 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
380 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
381 if (card_start >= top) {
382 continue;
383 }
384
385 // We claim lazily (so races are possible but they're benign), which reduces the
386 // number of duplicate scans (the rsets of the regions in the cset can intersect).
387 // Claim the card after checking bounds above: the remembered set may contain
388 // random cards into current survivor, and we would then have an incorrectly
389 // claimed card in survivor space. Card table clear does not reset the card table
390 // of survivor space regions.
391 claim_card(card_index, region_idx_for_card);
392
393 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
394
395 scan_card(mr, region_idx_for_card);
396 }
397 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::ScanRS));
398 }
399
400 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
401 EventGCPhaseParallel event;
402 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
403 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
404 }
405
406 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
407 assert(r->in_collection_set(),
408 "Should only be called on elements of the collection set but region %u is not.",
409 r->hrm_index());
410 uint const region_idx = r->hrm_index();
411
412 // Do an early out if we know we are complete.
413 if (_scan_state->iter_is_complete(region_idx)) {
414 return false;
415 }
416
417 {
418 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
419 scan_rem_set_roots(r);
420 }
421
422 if (_scan_state->set_iter_complete(region_idx)) {
423 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
424 // Scan the strong code root list attached to the current region
425 scan_strong_code_roots(r);
426 }
427 return false;
428 }
429
430 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
431 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
432 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i);
433 _g1h->collection_set_iterate_from(&cl, worker_i);
434
435 G1GCPhaseTimes* p = _g1p->phase_times();
436
437 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
438 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
439
440 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
441 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
442 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
443
444 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
445 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
446 }
447
448 // Closure used for updating rem sets. Only called during an evacuation pause.
449 class G1RefineCardClosure: public CardTableEntryClosure {
450 G1RemSet* _g1rs;
451 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
452
453 size_t _cards_scanned;
454 size_t _cards_skipped;
455 public:
456 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
457 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
458 {}
459
460 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
461 // The only time we care about recording cards that
462 // contain references that point into the collection set
463 // is during RSet updating within an evacuation pause.
464 // In this case worker_i should be the id of a GC worker thread.
465 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
466
467 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
468
469 if (card_scanned) {
470 _update_rs_cl->trim_queue_partially();
471 _cards_scanned++;
472 } else {
473 _cards_skipped++;
474 }
475 return true;
476 }
477
478 size_t cards_scanned() const { return _cards_scanned; }
479 size_t cards_skipped() const { return _cards_skipped; }
480 };
481
482 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
483 G1GCPhaseTimes* p = _g1p->phase_times();
484
485 // Apply closure to log entries in the HCC.
486 if (G1HotCardCache::default_use_cache()) {
487 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
488
489 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss, worker_i);
490 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
491 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
492 }
493
494 // Now apply the closure to all remaining log entries.
495 {
496 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
497
498 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i);
499 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
500 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
501
502 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
503 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
504 }
505 }
506
507 void G1RemSet::cleanupHRRS() {
508 HeapRegionRemSet::cleanup();
509 }
510
511 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
512 update_rem_set(pss, worker_i);
513 scan_rem_set(pss, worker_i);;
514 }
515
516 void G1RemSet::prepare_for_oops_into_collection_set_do() {
517 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
518 dcqs.concatenate_logs();
519
520 _scan_state->reset();
521 }
522
523 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
524 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times();
525
526 // Set all cards back to clean.
527 double start = os::elapsedTime();
528 _scan_state->clear_card_table(_g1h->workers());
529 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
530 }
531
532 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
533 #ifdef ASSERT
534 G1CollectedHeap* g1h = G1CollectedHeap::heap();
535 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
536 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
537 p2i(card_ptr),
538 ct->index_for(ct->addr_for(card_ptr)),
539 p2i(ct->addr_for(card_ptr)),
540 g1h->addr_to_region(ct->addr_for(card_ptr)));
541 #endif
542 }
543
544 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
545 uint worker_i) {
546 assert(!_g1h->is_gc_active(), "Only call concurrently");
547
548 check_card_ptr(card_ptr, _ct);
549
550 // If the card is no longer dirty, nothing to do.
551 if (*card_ptr != G1CardTable::dirty_card_val()) {
552 return;
553 }
554
555 // Construct the region representing the card.
556 HeapWord* start = _ct->addr_for(card_ptr);
557 // And find the region containing it.
558 HeapRegion* r = _g1h->heap_region_containing(start);
559
560 // This check is needed for some uncommon cases where we should
561 // ignore the card.
562 //
563 // The region could be young. Cards for young regions are
564 // distinctly marked (set to g1_young_gen), so the post-barrier will
565 // filter them out. However, that marking is performed
566 // concurrently. A write to a young object could occur before the
567 // card has been marked young, slipping past the filter.
568 //
569 // The card could be stale, because the region has been freed since
570 // the card was recorded. In this case the region type could be
571 // anything. If (still) free or (reallocated) young, just ignore
572 // it. If (reallocated) old or humongous, the later card trimming
573 // and additional checks in iteration may detect staleness. At
574 // worst, we end up processing a stale card unnecessarily.
575 //
576 // In the normal (non-stale) case, the synchronization between the
577 // enqueueing of the card and processing it here will have ensured
578 // we see the up-to-date region type here.
579 if (!r->is_old_or_humongous_or_archive()) {
580 return;
581 }
582
583 // The result from the hot card cache insert call is either:
584 // * pointer to the current card
585 // (implying that the current card is not 'hot'),
586 // * null
587 // (meaning we had inserted the card ptr into the "hot" card cache,
588 // which had some headroom),
589 // * a pointer to a "hot" card that was evicted from the "hot" cache.
590 //
591
592 if (_hot_card_cache->use_cache()) {
593 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
594
595 const jbyte* orig_card_ptr = card_ptr;
596 card_ptr = _hot_card_cache->insert(card_ptr);
597 if (card_ptr == NULL) {
598 // There was no eviction. Nothing to do.
599 return;
600 } else if (card_ptr != orig_card_ptr) {
601 // Original card was inserted and an old card was evicted.
602 start = _ct->addr_for(card_ptr);
603 r = _g1h->heap_region_containing(start);
604
605 // Check whether the region formerly in the cache should be
606 // ignored, as discussed earlier for the original card. The
607 // region could have been freed while in the cache.
608 if (!r->is_old_or_humongous_or_archive()) {
609 return;
610 }
611 } // Else we still have the original card.
612 }
613
614 // Trim the region designated by the card to what's been allocated
615 // in the region. The card could be stale, or the card could cover
616 // (part of) an object at the end of the allocated space and extend
617 // beyond the end of allocation.
618
619 // Non-humongous objects are only allocated in the old-gen during
620 // GC, so if region is old then top is stable. Humongous object
621 // allocation sets top last; if top has not yet been set, this is
622 // a stale card and we'll end up with an empty intersection. If
623 // this is not a stale card, the synchronization between the
624 // enqueuing of the card and processing it here will have ensured
625 // we see the up-to-date top here.
626 HeapWord* scan_limit = r->top();
627
628 if (scan_limit <= start) {
629 // If the trimmed region is empty, the card must be stale.
630 return;
631 }
632
633 // Okay to clean and process the card now. There are still some
634 // stale card cases that may be detected by iteration and dealt with
635 // as iteration failure.
636 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
637
638 // This fence serves two purposes. First, the card must be cleaned
639 // before processing the contents. Second, we can't proceed with
640 // processing until after the read of top, for synchronization with
641 // possibly concurrent humongous object allocation. It's okay that
642 // reading top and reading type were racy wrto each other. We need
643 // both set, in any order, to proceed.
644 OrderAccess::fence();
645
646 // Don't use addr_for(card_ptr + 1) which can ask for
647 // a card beyond the heap.
648 HeapWord* end = start + G1CardTable::card_size_in_words;
649 MemRegion dirty_region(start, MIN2(scan_limit, end));
650 assert(!dirty_region.is_empty(), "sanity");
651
652 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
653
654 bool card_processed =
655 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
656
657 // If unable to process the card then we encountered an unparsable
658 // part of the heap (e.g. a partially allocated object) while
659 // processing a stale card. Despite the card being stale, redirty
660 // and re-enqueue, because we've already cleaned the card. Without
661 // this we could incorrectly discard a non-stale card.
662 if (!card_processed) {
663 // The card might have gotten re-dirtied and re-enqueued while we
664 // worked. (In fact, it's pretty likely.)
665 if (*card_ptr != G1CardTable::dirty_card_val()) {
666 *card_ptr = G1CardTable::dirty_card_val();
667 MutexLockerEx x(Shared_DirtyCardQ_lock,
668 Mutex::_no_safepoint_check_flag);
669 DirtyCardQueue* sdcq =
670 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
671 sdcq->enqueue(card_ptr);
672 }
673 } else {
674 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
675 }
676 }
677
678 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
679 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
680 assert(_g1h->is_gc_active(), "Only call during GC");
681
682 check_card_ptr(card_ptr, _ct);
683
684 // If the card is no longer dirty, nothing to do. This covers cards that were already
685 // scanned as parts of the remembered sets.
686 if (*card_ptr != G1CardTable::dirty_card_val()) {
687 return false;
688 }
689
690 // We claim lazily (so races are possible but they're benign), which reduces the
691 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
692 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
693
694 // Construct the region representing the card.
695 HeapWord* card_start = _ct->addr_for(card_ptr);
696 // And find the region containing it.
697 uint const card_region_idx = _g1h->addr_to_region(card_start);
698
699 _scan_state->add_dirty_region(card_region_idx);
700 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
701 if (scan_limit <= card_start) {
702 // If the card starts above the area in the region containing objects to scan, skip it.
703 return false;
704 }
705
706 // Don't use addr_for(card_ptr + 1) which can ask for
707 // a card beyond the heap.
708 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
709 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
710 assert(!dirty_region.is_empty(), "sanity");
711
712 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
713 update_rs_cl->set_region(card_region);
714 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
715 assert(card_processed, "must be");
716 return true;
717 }
718
719 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
720 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
721 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
722
723 G1RemSetSummary current(this);
724 _prev_period_summary.subtract_from(¤t);
725
726 Log(gc, remset) log;
727 log.trace("%s", header);
728 ResourceMark rm;
729 LogStream ls(log.trace());
730 _prev_period_summary.print_on(&ls);
731
732 _prev_period_summary.set(¤t);
733 }
734 }
735
736 void G1RemSet::print_summary_info() {
737 Log(gc, remset, exit) log;
738 if (log.is_trace()) {
739 log.trace(" Cumulative RS summary");
740 G1RemSetSummary current(this);
741 ResourceMark rm;
742 LogStream ls(log.trace());
743 current.print_on(&ls);
744 }
745 }
746
747 class G1RebuildRemSetTask: public AbstractGangTask {
748 // Aggregate the counting data that was constructed concurrently
749 // with marking.
750 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
751 G1ConcurrentMark* _cm;
752 G1RebuildRemSetClosure _update_cl;
753
754 // Applies _update_cl to the references of the given object, limiting objArrays
755 // to the given MemRegion. Returns the amount of words actually scanned.
756 size_t scan_for_references(oop const obj, MemRegion mr) {
757 size_t const obj_size = obj->size();
758 // All non-objArrays and objArrays completely within the mr
759 // can be scanned without passing the mr.
760 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
761 obj->oop_iterate(&_update_cl);
762 return obj_size;
763 }
764 // This path is for objArrays crossing the given MemRegion. Only scan the
765 // area within the MemRegion.
766 obj->oop_iterate(&_update_cl, mr);
767 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
768 }
769
770 // A humongous object is live (with respect to the scanning) either
771 // a) it is marked on the bitmap as such
772 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
773 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
774 return bitmap->is_marked(humongous_obj) || (tars > tams);
775 }
776
777 // Iterator over the live objects within the given MemRegion.
778 class LiveObjIterator : public StackObj {
779 const G1CMBitMap* const _bitmap;
780 const HeapWord* _tams;
781 const MemRegion _mr;
782 HeapWord* _current;
783
784 bool is_below_tams() const {
785 return _current < _tams;
786 }
787
788 bool is_live(HeapWord* obj) const {
789 return !is_below_tams() || _bitmap->is_marked(obj);
790 }
791
792 HeapWord* bitmap_limit() const {
793 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
794 }
795
796 void move_if_below_tams() {
797 if (is_below_tams() && has_next()) {
798 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
799 }
800 }
801 public:
802 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
803 _bitmap(bitmap),
804 _tams(tams),
805 _mr(mr),
806 _current(first_oop_into_mr) {
807
808 assert(_current <= _mr.start(),
809 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
810 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
811
812 // Step to the next live object within the MemRegion if needed.
813 if (is_live(_current)) {
814 // Non-objArrays were scanned by the previous part of that region.
815 if (_current < mr.start() && !oop(_current)->is_objArray()) {
816 _current += oop(_current)->size();
817 // We might have positioned _current on a non-live object. Reposition to the next
818 // live one if needed.
819 move_if_below_tams();
820 }
821 } else {
822 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
823 // immediately.
824 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
825 assert(_current == _mr.end() || is_live(_current),
826 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
827 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
828 }
829 }
830
831 void move_to_next() {
832 _current += next()->size();
833 move_if_below_tams();
834 }
835
836 oop next() const {
837 oop result = oop(_current);
838 assert(is_live(_current),
839 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
840 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
841 return result;
842 }
843
844 bool has_next() const {
845 return _current < _mr.end();
846 }
847 };
848
849 // Rebuild remembered sets in the part of the region specified by mr and hr.
850 // Objects between the bottom of the region and the TAMS are checked for liveness
851 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
852 // Returns the number of live words between bottom and TAMS.
853 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
854 HeapWord* const top_at_mark_start,
855 HeapWord* const top_at_rebuild_start,
856 HeapRegion* hr,
857 MemRegion mr) {
858 size_t marked_words = 0;
859
860 if (hr->is_humongous()) {
861 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
862 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
863 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
864 // however in case of humongous objects it is sufficient to scan the encompassing
865 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
866 // two areas will be zero sized. I.e. TAMS is either
867 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
868 // value: this would mean that TAMS points somewhere into the object.
869 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
870 "More than one object in the humongous region?");
871 humongous_obj->oop_iterate(&_update_cl, mr);
872 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
873 } else {
874 return 0;
875 }
876 }
877
878 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
879 oop obj = it.next();
880 size_t scanned_size = scan_for_references(obj, mr);
881 if ((HeapWord*)obj < top_at_mark_start) {
882 marked_words += scanned_size;
883 }
884 }
885
886 return marked_words * HeapWordSize;
887 }
888 public:
889 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
890 G1ConcurrentMark* cm,
891 uint worker_id) :
892 HeapRegionClosure(),
893 _cm(cm),
894 _update_cl(g1h, worker_id) { }
895
896 bool do_heap_region(HeapRegion* hr) {
897 if (_cm->has_aborted()) {
898 return true;
899 }
900
901 uint const region_idx = hr->hrm_index();
902 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
903 assert(top_at_rebuild_start_check == NULL ||
904 top_at_rebuild_start_check > hr->bottom(),
905 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
906 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
907
908 size_t total_marked_bytes = 0;
909 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
910
911 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
912
913 HeapWord* cur = hr->bottom();
914 while (cur < hr->end()) {
915 // After every iteration (yield point) we need to check whether the region's
916 // TARS changed due to e.g. eager reclaim.
917 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
918 if (top_at_rebuild_start == NULL) {
919 return false;
920 }
921
922 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
923 if (next_chunk.is_empty()) {
924 break;
925 }
926
927 const Ticks start = Ticks::now();
928 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
929 top_at_mark_start,
930 top_at_rebuild_start,
931 hr,
932 next_chunk);
933 Tickspan time = Ticks::now() - start;
934
935 log_trace(gc, remset, tracking)("Rebuilt region %u "
936 "live " SIZE_FORMAT " "
937 "time %.3fms "
938 "marked bytes " SIZE_FORMAT " "
939 "bot " PTR_FORMAT " "
940 "TAMS " PTR_FORMAT " "
941 "TARS " PTR_FORMAT,
942 region_idx,
943 _cm->liveness(region_idx) * HeapWordSize,
944 time.seconds() * 1000.0,
945 marked_bytes,
946 p2i(hr->bottom()),
947 p2i(top_at_mark_start),
948 p2i(top_at_rebuild_start));
949
950 if (marked_bytes > 0) {
951 total_marked_bytes += marked_bytes;
952 }
953 cur += chunk_size_in_words;
954
955 _cm->do_yield_check();
956 if (_cm->has_aborted()) {
957 return true;
958 }
959 }
960 // In the final iteration of the loop the region might have been eagerly reclaimed.
961 // Simply filter out those regions. We can not just use region type because there
962 // might have already been new allocations into these regions.
963 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
964 assert(top_at_rebuild_start == NULL ||
965 total_marked_bytes == hr->marked_bytes(),
966 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
967 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
968 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
969 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
970 // Abort state may have changed after the yield check.
971 return _cm->has_aborted();
972 }
973 };
974
975 HeapRegionClaimer _hr_claimer;
976 G1ConcurrentMark* _cm;
977
978 uint _worker_id_offset;
979 public:
980 G1RebuildRemSetTask(G1ConcurrentMark* cm,
981 uint n_workers,
982 uint worker_id_offset) :
983 AbstractGangTask("G1 Rebuild Remembered Set"),
984 _hr_claimer(n_workers),
985 _cm(cm),
986 _worker_id_offset(worker_id_offset) {
987 }
988
989 void work(uint worker_id) {
990 SuspendibleThreadSetJoiner sts_join;
991
992 G1CollectedHeap* g1h = G1CollectedHeap::heap();
993
994 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
995 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
996 }
997 };
998
999 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1000 WorkGang* workers,
1001 uint worker_id_offset) {
1002 uint num_workers = workers->active_workers();
1003
1004 G1RebuildRemSetTask cl(cm,
1005 num_workers,
1006 worker_id_offset);
1007 workers->run_task(&cl, num_workers);
1008 }