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