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 }
139 return false;
140 }
141 };
142
143 // For each region, contains the maximum top() value to be used during this garbage
144 // collection. Subsumes common checks like filtering out everything but old and
145 // humongous regions outside the collection set.
146 // This is valid because we are not interested in scanning stray remembered set
147 // entries from free or archive regions.
148 HeapWord** _scan_top;
149 public:
150 G1RemSetScanState() :
151 _max_regions(0),
152 _iter_states(NULL),
153 _iter_claims(NULL),
154 _dirty_region_buffer(NULL),
155 _in_dirty_region_buffer(NULL),
156 _cur_dirty_region(0),
157 _scan_top(NULL) {
158 }
159
160 ~G1RemSetScanState() {
161 if (_iter_states != NULL) {
162 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
163 }
164 if (_iter_claims != NULL) {
165 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
166 }
167 if (_dirty_region_buffer != NULL) {
168 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
169 }
170 if (_in_dirty_region_buffer != NULL) {
171 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
172 }
173 if (_scan_top != NULL) {
174 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
175 }
176 }
177
178 void initialize(uint max_regions) {
179 assert(_iter_states == NULL, "Must not be initialized twice");
180 assert(_iter_claims == NULL, "Must not be initialized twice");
181 _max_regions = max_regions;
182 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
183 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
184 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
185 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
186 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
187 }
188
189 void reset() {
190 for (uint i = 0; i < _max_regions; i++) {
191 _iter_states[i] = Unclaimed;
192 _scan_top[i] = NULL;
193 }
194
195 G1ResetScanTopClosure cl(_scan_top);
196 G1CollectedHeap::heap()->heap_region_iterate(&cl);
197
198 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
199 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState));
200 _cur_dirty_region = 0;
201 }
202
203 // Attempt to claim the remembered set of the region for iteration. Returns true
204 // if this call caused the transition from Unclaimed to Claimed.
205 inline bool claim_iter(uint region) {
206 assert(region < _max_regions, "Tried to access invalid region %u", region);
207 if (_iter_states[region] != Unclaimed) {
208 return false;
209 }
210 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
211 return (res == Unclaimed);
212 }
213
214 // Try to atomically sets the iteration state to "complete". Returns true for the
215 // thread that caused the transition.
216 inline bool set_iter_complete(uint region) {
217 if (iter_is_complete(region)) {
218 return false;
219 }
220 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
221 return (res == Claimed);
222 }
223
224 // Returns true if the region's iteration is complete.
225 inline bool iter_is_complete(uint region) const {
226 assert(region < _max_regions, "Tried to access invalid region %u", region);
227 return _iter_states[region] == Complete;
228 }
229
230 // The current position within the remembered set of the given region.
231 inline size_t iter_claimed(uint region) const {
232 assert(region < _max_regions, "Tried to access invalid region %u", region);
233 return _iter_claims[region];
234 }
235
236 // Claim the next block of cards within the remembered set of the region with
237 // step size.
238 inline size_t iter_claimed_next(uint region, size_t step) {
239 return Atomic::add(step, &_iter_claims[region]) - step;
240 }
241
242 void add_dirty_region(uint region) {
243 if (_in_dirty_region_buffer[region] == Dirty) {
244 return;
245 }
246
247 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean;
248 if (marked_as_dirty) {
249 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
250 _dirty_region_buffer[allocated] = region;
251 }
252 }
253
254 HeapWord* scan_top(uint region_idx) const {
255 return _scan_top[region_idx];
256 }
257
258 // Clear the card table of "dirty" regions.
259 void clear_card_table(WorkGang* workers) {
260 if (_cur_dirty_region == 0) {
261 return;
262 }
263
264 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
265 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
266 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
267
268 // Iterate over the dirty cards region list.
269 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
270
271 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
272 "units of work for " SIZE_FORMAT " regions.",
273 cl.name(), num_workers, num_chunks, _cur_dirty_region);
274 workers->run_task(&cl, num_workers);
275
276 #ifndef PRODUCT
277 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
278 #endif
279 }
280 };
281
282 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
283 G1CardTable* ct,
284 G1HotCardCache* hot_card_cache) :
285 _scan_state(new G1RemSetScanState()),
286 _prev_period_summary(),
287 _g1h(g1h),
288 _num_conc_refined_cards(0),
289 _ct(ct),
290 _g1p(_g1h->g1_policy()),
291 _hot_card_cache(hot_card_cache) {
292 }
293
294 G1RemSet::~G1RemSet() {
295 if (_scan_state != NULL) {
296 delete _scan_state;
297 }
298 }
299
300 uint G1RemSet::num_par_rem_sets() {
301 return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
302 }
303
304 void G1RemSet::initialize(size_t capacity, uint max_regions) {
305 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
306 _scan_state->initialize(max_regions);
307 }
308
309 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
310 G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
311 G1ParScanThreadState* pss,
312 uint worker_i) :
313 _g1h(G1CollectedHeap::heap()),
314 _ct(_g1h->card_table()),
315 _pss(pss),
316 _scan_objs_on_card_cl(scan_obj_on_card),
317 _scan_state(scan_state),
318 _worker_i(worker_i),
319 _cards_scanned(0),
320 _cards_claimed(0),
321 _cards_skipped(0),
322 _rem_set_root_scan_time(),
323 _rem_set_trim_partially_time(),
324 _strong_code_root_scan_time(),
325 _strong_code_trim_partially_time() {
326 }
327
328 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
329 _ct->set_card_claimed(card_index);
330 _scan_state->add_dirty_region(region_idx_for_card);
331 }
332
333 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
334 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
335 _scan_objs_on_card_cl->set_region(card_region);
336 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
337 _scan_objs_on_card_cl->trim_queue_partially();
338 _cards_scanned++;
339 }
340
341 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
342 EventGCPhaseParallel event;
343 uint const region_idx = r->hrm_index();
344
345 if (_scan_state->claim_iter(region_idx)) {
346 // If we ever free the collection set concurrently, we should also
347 // clear the card table concurrently therefore we won't need to
348 // add regions of the collection set to the dirty cards region.
349 _scan_state->add_dirty_region(region_idx);
350 }
351
352 if (r->rem_set()->cardset_is_empty()) {
353 return;
354 }
355
356 // We claim cards in blocks so as to reduce the contention.
357 size_t const block_size = G1RSetScanBlockSize;
358
359 HeapRegionRemSetIterator iter(r->rem_set());
360 size_t card_index;
361
362 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
363 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
364 if (current_card >= claimed_card_block + block_size) {
365 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
366 }
367 if (current_card < claimed_card_block) {
368 _cards_skipped++;
369 continue;
370 }
371 _cards_claimed++;
372
373 HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
374 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
375
376 #ifdef ASSERT
377 HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
378 assert(hr == NULL || hr->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 #endif
381 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
382 if (card_start >= top) {
383 continue;
384 }
385
386 // If the card is dirty, then G1 will scan it during Update RS.
387 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
388 continue;
389 }
390
391 // We claim lazily (so races are possible but they're benign), which reduces the
392 // number of duplicate scans (the rsets of the regions in the cset can intersect).
393 // Claim the card after checking bounds above: the remembered set may contain
394 // random cards into current survivor, and we would then have an incorrectly
395 // claimed card in survivor space. Card table clear does not reset the card table
396 // of survivor space regions.
397 claim_card(card_index, region_idx_for_card);
398
399 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
400
401 scan_card(mr, region_idx_for_card);
402 }
403 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::ScanRS));
404 }
405
406 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
407 EventGCPhaseParallel event;
408 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
409 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
410 }
411
412 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
413 assert(r->in_collection_set(),
414 "Should only be called on elements of the collection set but region %u is not.",
415 r->hrm_index());
416 uint const region_idx = r->hrm_index();
417
418 // Do an early out if we know we are complete.
419 if (_scan_state->iter_is_complete(region_idx)) {
420 return false;
421 }
422
423 {
424 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
425 scan_rem_set_roots(r);
426 }
427
428 if (_scan_state->set_iter_complete(region_idx)) {
429 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
430 // Scan the strong code root list attached to the current region
431 scan_strong_code_roots(r);
432 }
433 return false;
434 }
435
436 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
437 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
438 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i);
439 _g1h->collection_set_iterate_from(&cl, worker_i);
440
441 G1GCPhaseTimes* p = _g1p->phase_times();
442
443 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
444 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
445
446 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
447 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
448 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
449
450 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
451 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
452 }
453
454 // Closure used for updating rem sets. Only called during an evacuation pause.
455 class G1RefineCardClosure: public CardTableEntryClosure {
456 G1RemSet* _g1rs;
457 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
458
459 size_t _cards_scanned;
460 size_t _cards_skipped;
461 public:
462 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
463 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
464 {}
465
466 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
467 // The only time we care about recording cards that
468 // contain references that point into the collection set
469 // is during RSet updating within an evacuation pause.
470 // In this case worker_i should be the id of a GC worker thread.
471 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
472
473 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
474
475 if (card_scanned) {
476 _update_rs_cl->trim_queue_partially();
477 _cards_scanned++;
478 } else {
479 _cards_skipped++;
480 }
481 return true;
482 }
483
484 size_t cards_scanned() const { return _cards_scanned; }
485 size_t cards_skipped() const { return _cards_skipped; }
486 };
487
488 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
489 G1GCPhaseTimes* p = _g1p->phase_times();
490
491 // Apply closure to log entries in the HCC.
492 if (G1HotCardCache::default_use_cache()) {
493 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
494
495 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss, worker_i);
496 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
497 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
498 }
499
500 // Now apply the closure to all remaining log entries.
501 {
502 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
503
504 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i);
505 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
506 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
507
508 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
509 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
510 }
511 }
512
513 void G1RemSet::cleanupHRRS() {
514 HeapRegionRemSet::cleanup();
515 }
516
517 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
518 update_rem_set(pss, worker_i);
519 scan_rem_set(pss, worker_i);;
520 }
521
522 void G1RemSet::prepare_for_oops_into_collection_set_do() {
523 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
524 dcqs.concatenate_logs();
525
526 _scan_state->reset();
527 }
528
529 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
530 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times();
531
532 // Set all cards back to clean.
533 double start = os::elapsedTime();
534 _scan_state->clear_card_table(_g1h->workers());
535 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
536 }
537
538 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
539 #ifdef ASSERT
540 G1CollectedHeap* g1h = G1CollectedHeap::heap();
541 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
542 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
543 p2i(card_ptr),
544 ct->index_for(ct->addr_for(card_ptr)),
545 p2i(ct->addr_for(card_ptr)),
546 g1h->addr_to_region(ct->addr_for(card_ptr)));
547 #endif
548 }
549
550 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
551 uint worker_i) {
552 assert(!_g1h->is_gc_active(), "Only call concurrently");
553
554 // Construct the region representing the card.
555 HeapWord* start = _ct->addr_for(card_ptr);
556 // And find the region containing it.
557 HeapRegion* r = _g1h->heap_region_containing_or_null(start);
558
559 // If this is a (stale) card into an uncommitted region, exit.
560 if (r == NULL) {
561 return;
562 }
563
564 check_card_ptr(card_ptr, _ct);
565
566 // If the card is no longer dirty, nothing to do.
567 if (*card_ptr != G1CardTable::dirty_card_val()) {
568 return;
569 }
570
571 // This check is needed for some uncommon cases where we should
572 // ignore the card.
573 //
574 // The region could be young. Cards for young regions are
575 // distinctly marked (set to g1_young_gen), so the post-barrier will
576 // filter them out. However, that marking is performed
577 // concurrently. A write to a young object could occur before the
578 // card has been marked young, slipping past the filter.
579 //
580 // The card could be stale, because the region has been freed since
581 // the card was recorded. In this case the region type could be
582 // anything. If (still) free or (reallocated) young, just ignore
583 // it. If (reallocated) old or humongous, the later card trimming
584 // and additional checks in iteration may detect staleness. At
585 // worst, we end up processing a stale card unnecessarily.
586 //
587 // In the normal (non-stale) case, the synchronization between the
588 // enqueueing of the card and processing it here will have ensured
589 // we see the up-to-date region type here.
590 if (!r->is_old_or_humongous_or_archive()) {
591 return;
592 }
593
594 // The result from the hot card cache insert call is either:
595 // * pointer to the current card
596 // (implying that the current card is not 'hot'),
597 // * null
598 // (meaning we had inserted the card ptr into the "hot" card cache,
599 // which had some headroom),
600 // * a pointer to a "hot" card that was evicted from the "hot" cache.
601 //
602
603 if (_hot_card_cache->use_cache()) {
604 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
605
606 const jbyte* orig_card_ptr = card_ptr;
607 card_ptr = _hot_card_cache->insert(card_ptr);
608 if (card_ptr == NULL) {
609 // There was no eviction. Nothing to do.
610 return;
611 } else if (card_ptr != orig_card_ptr) {
612 // Original card was inserted and an old card was evicted.
613 start = _ct->addr_for(card_ptr);
614 r = _g1h->heap_region_containing(start);
615
616 // Check whether the region formerly in the cache should be
617 // ignored, as discussed earlier for the original card. The
618 // region could have been freed while in the cache.
619 if (!r->is_old_or_humongous_or_archive()) {
620 return;
621 }
622 } // Else we still have the original card.
623 }
624
625 // Trim the region designated by the card to what's been allocated
626 // in the region. The card could be stale, or the card could cover
627 // (part of) an object at the end of the allocated space and extend
628 // beyond the end of allocation.
629
630 // Non-humongous objects are only allocated in the old-gen during
631 // GC, so if region is old then top is stable. Humongous object
632 // allocation sets top last; if top has not yet been set, this is
633 // a stale card and we'll end up with an empty intersection. If
634 // this is not a stale card, the synchronization between the
635 // enqueuing of the card and processing it here will have ensured
636 // we see the up-to-date top here.
637 HeapWord* scan_limit = r->top();
638
639 if (scan_limit <= start) {
640 // If the trimmed region is empty, the card must be stale.
641 return;
642 }
643
644 // Okay to clean and process the card now. There are still some
645 // stale card cases that may be detected by iteration and dealt with
646 // as iteration failure.
647 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
648
649 // This fence serves two purposes. First, the card must be cleaned
650 // before processing the contents. Second, we can't proceed with
651 // processing until after the read of top, for synchronization with
652 // possibly concurrent humongous object allocation. It's okay that
653 // reading top and reading type were racy wrto each other. We need
654 // both set, in any order, to proceed.
655 OrderAccess::fence();
656
657 // Don't use addr_for(card_ptr + 1) which can ask for
658 // a card beyond the heap.
659 HeapWord* end = start + G1CardTable::card_size_in_words;
660 MemRegion dirty_region(start, MIN2(scan_limit, end));
661 assert(!dirty_region.is_empty(), "sanity");
662
663 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
664
665 bool card_processed =
666 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
667
668 // If unable to process the card then we encountered an unparsable
669 // part of the heap (e.g. a partially allocated object) while
670 // processing a stale card. Despite the card being stale, redirty
671 // and re-enqueue, because we've already cleaned the card. Without
672 // this we could incorrectly discard a non-stale card.
673 if (!card_processed) {
674 // The card might have gotten re-dirtied and re-enqueued while we
675 // worked. (In fact, it's pretty likely.)
676 if (*card_ptr != G1CardTable::dirty_card_val()) {
677 *card_ptr = G1CardTable::dirty_card_val();
678 MutexLockerEx x(Shared_DirtyCardQ_lock,
679 Mutex::_no_safepoint_check_flag);
680 DirtyCardQueue* sdcq =
681 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
682 sdcq->enqueue(card_ptr);
683 }
684 } else {
685 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
686 }
687 }
688
689 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
690 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
691 assert(_g1h->is_gc_active(), "Only call during GC");
692
693 // Construct the region representing the card.
694 HeapWord* card_start = _ct->addr_for(card_ptr);
695 // And find the region containing it.
696 uint const card_region_idx = _g1h->addr_to_region(card_start);
697
698 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
699 if (scan_limit == NULL) {
700 // This is a card into an uncommitted region. We need to bail out early as we
701 // should not access the corresponding card table entry.
702 return false;
703 }
704
705 check_card_ptr(card_ptr, _ct);
706
707 // If the card is no longer dirty, nothing to do. This covers cards that were already
708 // scanned as parts of the remembered sets.
709 if (*card_ptr != G1CardTable::dirty_card_val()) {
710 return false;
711 }
712
713 // We claim lazily (so races are possible but they're benign), which reduces the
714 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
715 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
716
717 _scan_state->add_dirty_region(card_region_idx);
718 if (scan_limit <= card_start) {
719 // If the card starts above the area in the region containing objects to scan, skip it.
720 return false;
721 }
722
723 // Don't use addr_for(card_ptr + 1) which can ask for
724 // a card beyond the heap.
725 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
726 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
727 assert(!dirty_region.is_empty(), "sanity");
728
729 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
730 update_rs_cl->set_region(card_region);
731 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
732 assert(card_processed, "must be");
733 return true;
734 }
735
736 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
737 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
738 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
739
740 G1RemSetSummary current(this);
741 _prev_period_summary.subtract_from(¤t);
742
743 Log(gc, remset) log;
744 log.trace("%s", header);
745 ResourceMark rm;
746 LogStream ls(log.trace());
747 _prev_period_summary.print_on(&ls);
748
749 _prev_period_summary.set(¤t);
750 }
751 }
752
753 void G1RemSet::print_summary_info() {
754 Log(gc, remset, exit) log;
755 if (log.is_trace()) {
756 log.trace(" Cumulative RS summary");
757 G1RemSetSummary current(this);
758 ResourceMark rm;
759 LogStream ls(log.trace());
760 current.print_on(&ls);
761 }
762 }
763
764 class G1RebuildRemSetTask: public AbstractGangTask {
765 // Aggregate the counting data that was constructed concurrently
766 // with marking.
767 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
768 G1ConcurrentMark* _cm;
769 G1RebuildRemSetClosure _update_cl;
770
771 // Applies _update_cl to the references of the given object, limiting objArrays
772 // to the given MemRegion. Returns the amount of words actually scanned.
773 size_t scan_for_references(oop const obj, MemRegion mr) {
774 size_t const obj_size = obj->size();
775 // All non-objArrays and objArrays completely within the mr
776 // can be scanned without passing the mr.
777 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
778 obj->oop_iterate(&_update_cl);
779 return obj_size;
780 }
781 // This path is for objArrays crossing the given MemRegion. Only scan the
782 // area within the MemRegion.
783 obj->oop_iterate(&_update_cl, mr);
784 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
785 }
786
787 // A humongous object is live (with respect to the scanning) either
788 // a) it is marked on the bitmap as such
789 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
790 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
791 return bitmap->is_marked(humongous_obj) || (tars > tams);
792 }
793
794 // Iterator over the live objects within the given MemRegion.
795 class LiveObjIterator : public StackObj {
796 const G1CMBitMap* const _bitmap;
797 const HeapWord* _tams;
798 const MemRegion _mr;
799 HeapWord* _current;
800
801 bool is_below_tams() const {
802 return _current < _tams;
803 }
804
805 bool is_live(HeapWord* obj) const {
806 return !is_below_tams() || _bitmap->is_marked(obj);
807 }
808
809 HeapWord* bitmap_limit() const {
810 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
811 }
812
813 void move_if_below_tams() {
814 if (is_below_tams() && has_next()) {
815 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
816 }
817 }
818 public:
819 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
820 _bitmap(bitmap),
821 _tams(tams),
822 _mr(mr),
823 _current(first_oop_into_mr) {
824
825 assert(_current <= _mr.start(),
826 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
827 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
828
829 // Step to the next live object within the MemRegion if needed.
830 if (is_live(_current)) {
831 // Non-objArrays were scanned by the previous part of that region.
832 if (_current < mr.start() && !oop(_current)->is_objArray()) {
833 _current += oop(_current)->size();
834 // We might have positioned _current on a non-live object. Reposition to the next
835 // live one if needed.
836 move_if_below_tams();
837 }
838 } else {
839 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
840 // immediately.
841 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
842 assert(_current == _mr.end() || is_live(_current),
843 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
844 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
845 }
846 }
847
848 void move_to_next() {
849 _current += next()->size();
850 move_if_below_tams();
851 }
852
853 oop next() const {
854 oop result = oop(_current);
855 assert(is_live(_current),
856 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
857 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
858 return result;
859 }
860
861 bool has_next() const {
862 return _current < _mr.end();
863 }
864 };
865
866 // Rebuild remembered sets in the part of the region specified by mr and hr.
867 // Objects between the bottom of the region and the TAMS are checked for liveness
868 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
869 // Returns the number of live words between bottom and TAMS.
870 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
871 HeapWord* const top_at_mark_start,
872 HeapWord* const top_at_rebuild_start,
873 HeapRegion* hr,
874 MemRegion mr) {
875 size_t marked_words = 0;
876
877 if (hr->is_humongous()) {
878 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
879 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
880 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
881 // however in case of humongous objects it is sufficient to scan the encompassing
882 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
883 // two areas will be zero sized. I.e. TAMS is either
884 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
885 // value: this would mean that TAMS points somewhere into the object.
886 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
887 "More than one object in the humongous region?");
888 humongous_obj->oop_iterate(&_update_cl, mr);
889 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
890 } else {
891 return 0;
892 }
893 }
894
895 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
896 oop obj = it.next();
897 size_t scanned_size = scan_for_references(obj, mr);
898 if ((HeapWord*)obj < top_at_mark_start) {
899 marked_words += scanned_size;
900 }
901 }
902
903 return marked_words * HeapWordSize;
904 }
905 public:
906 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
907 G1ConcurrentMark* cm,
908 uint worker_id) :
909 HeapRegionClosure(),
910 _cm(cm),
911 _update_cl(g1h, worker_id) { }
912
913 bool do_heap_region(HeapRegion* hr) {
914 if (_cm->has_aborted()) {
915 return true;
916 }
917
918 uint const region_idx = hr->hrm_index();
919 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
920 assert(top_at_rebuild_start_check == NULL ||
921 top_at_rebuild_start_check > hr->bottom(),
922 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
923 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
924
925 size_t total_marked_bytes = 0;
926 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
927
928 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
929
930 HeapWord* cur = hr->bottom();
931 while (cur < hr->end()) {
932 // After every iteration (yield point) we need to check whether the region's
933 // TARS changed due to e.g. eager reclaim.
934 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
935 if (top_at_rebuild_start == NULL) {
936 return false;
937 }
938
939 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
940 if (next_chunk.is_empty()) {
941 break;
942 }
943
944 const Ticks start = Ticks::now();
945 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
946 top_at_mark_start,
947 top_at_rebuild_start,
948 hr,
949 next_chunk);
950 Tickspan time = Ticks::now() - start;
951
952 log_trace(gc, remset, tracking)("Rebuilt region %u "
953 "live " SIZE_FORMAT " "
954 "time %.3fms "
955 "marked bytes " SIZE_FORMAT " "
956 "bot " PTR_FORMAT " "
957 "TAMS " PTR_FORMAT " "
958 "TARS " PTR_FORMAT,
959 region_idx,
960 _cm->liveness(region_idx) * HeapWordSize,
961 time.seconds() * 1000.0,
962 marked_bytes,
963 p2i(hr->bottom()),
964 p2i(top_at_mark_start),
965 p2i(top_at_rebuild_start));
966
967 if (marked_bytes > 0) {
968 total_marked_bytes += marked_bytes;
969 }
970 cur += chunk_size_in_words;
971
972 _cm->do_yield_check();
973 if (_cm->has_aborted()) {
974 return true;
975 }
976 }
977 // In the final iteration of the loop the region might have been eagerly reclaimed.
978 // Simply filter out those regions. We can not just use region type because there
979 // might have already been new allocations into these regions.
980 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
981 assert(top_at_rebuild_start == NULL ||
982 total_marked_bytes == hr->marked_bytes(),
983 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
984 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
985 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
986 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
987 // Abort state may have changed after the yield check.
988 return _cm->has_aborted();
989 }
990 };
991
992 HeapRegionClaimer _hr_claimer;
993 G1ConcurrentMark* _cm;
994
995 uint _worker_id_offset;
996 public:
997 G1RebuildRemSetTask(G1ConcurrentMark* cm,
998 uint n_workers,
999 uint worker_id_offset) :
1000 AbstractGangTask("G1 Rebuild Remembered Set"),
1001 _hr_claimer(n_workers),
1002 _cm(cm),
1003 _worker_id_offset(worker_id_offset) {
1004 }
1005
1006 void work(uint worker_id) {
1007 SuspendibleThreadSetJoiner sts_join;
1008
1009 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1010
1011 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1012 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1013 }
1014 };
1015
1016 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1017 WorkGang* workers,
1018 uint worker_id_offset) {
1019 uint num_workers = workers->active_workers();
1020
1021 G1RebuildRemSetTask cl(cm,
1022 num_workers,
1023 worker_id_offset);
1024 workers->run_task(&cl, num_workers);
1025 }