1 /*
2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/g1BarrierSet.hpp"
27 #include "gc/g1/g1BlockOffsetTable.inline.hpp"
28 #include "gc/g1/g1CardTable.inline.hpp"
29 #include "gc/g1/g1CollectedHeap.inline.hpp"
30 #include "gc/g1/g1ConcurrentRefine.hpp"
31 #include "gc/g1/g1DirtyCardQueue.hpp"
32 #include "gc/g1/g1FromCardCache.hpp"
33 #include "gc/g1/g1GCPhaseTimes.hpp"
34 #include "gc/g1/g1HotCardCache.hpp"
35 #include "gc/g1/g1OopClosures.inline.hpp"
36 #include "gc/g1/g1RootClosures.hpp"
37 #include "gc/g1/g1RemSet.hpp"
38 #include "gc/g1/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/stack.inline.hpp"
52 #include "utilities/ticks.hpp"
53
54 // Collects information about the overall remembered set scan progress during an evacuation.
55 class G1RemSetScanState : public CHeapObj<mtGC> {
56 private:
57 class G1ClearCardTableTask : public AbstractGangTask {
58 G1CollectedHeap* _g1h;
59 uint* _dirty_region_list;
60 size_t _num_dirty_regions;
61 size_t _chunk_length;
62
63 size_t volatile _cur_dirty_regions;
64 public:
65 G1ClearCardTableTask(G1CollectedHeap* g1h,
66 uint* dirty_region_list,
67 size_t num_dirty_regions,
68 size_t chunk_length) :
69 AbstractGangTask("G1 Clear Card Table Task"),
70 _g1h(g1h),
71 _dirty_region_list(dirty_region_list),
72 _num_dirty_regions(num_dirty_regions),
73 _chunk_length(chunk_length),
74 _cur_dirty_regions(0) {
75
76 assert(chunk_length > 0, "must be");
77 }
78
79 static size_t chunk_size() { return M; }
80
81 void work(uint worker_id) {
82 while (_cur_dirty_regions < _num_dirty_regions) {
83 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
84 size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
85
86 for (size_t i = next; i < max; i++) {
87 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
88 if (!r->is_survivor()) {
89 r->clear_cardtable();
90 }
91 }
92 }
93 }
94 };
95
96 size_t _max_regions;
97
98 // Scan progress for the remembered set of a single region. Transitions from
99 // Unclaimed -> Claimed -> Complete.
100 // At each of the transitions the thread that does the transition needs to perform
101 // some special action once. This is the reason for the extra "Claimed" state.
102 typedef jint G1RemsetIterState;
103
104 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
105 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned.
106 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned.
107
108 G1RemsetIterState volatile* _iter_states;
109 // The current location where the next thread should continue scanning in a region's
110 // remembered set.
111 size_t volatile* _iter_claims;
112
113 // Temporary buffer holding the regions we used to store remembered set scan duplicate
114 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
115 uint* _dirty_region_buffer;
116
117 typedef jbyte IsDirtyRegionState;
118 static const IsDirtyRegionState Clean = 0;
119 static const IsDirtyRegionState Dirty = 1;
120 // Holds a flag for every region whether it is in the _dirty_region_buffer already
121 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools.
122 IsDirtyRegionState* _in_dirty_region_buffer;
123 size_t _cur_dirty_region;
124
125 // Creates a snapshot of the current _top values at the start of collection to
126 // filter out card marks that we do not want to scan.
127 class G1ResetScanTopClosure : public HeapRegionClosure {
128 private:
129 HeapWord** _scan_top;
130 public:
131 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
132
133 virtual bool do_heap_region(HeapRegion* r) {
134 uint hrm_index = r->hrm_index();
135 if (!r->in_collection_set() && r->is_old_or_humongous_or_archive() && !r->is_empty()) {
136 _scan_top[hrm_index] = r->top();
137 } else {
138 _scan_top[hrm_index] = NULL;
139 }
140 return false;
141 }
142 };
143
144 // For each region, contains the maximum top() value to be used during this garbage
145 // collection. Subsumes common checks like filtering out everything but old and
146 // humongous regions outside the collection set.
147 // This is valid because we are not interested in scanning stray remembered set
148 // entries from free or archive regions.
149 HeapWord** _scan_top;
150 public:
151 G1RemSetScanState() :
152 _max_regions(0),
153 _iter_states(NULL),
154 _iter_claims(NULL),
155 _dirty_region_buffer(NULL),
156 _in_dirty_region_buffer(NULL),
157 _cur_dirty_region(0),
158 _scan_top(NULL) {
159 }
160
161 ~G1RemSetScanState() {
162 if (_iter_states != NULL) {
163 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
164 }
165 if (_iter_claims != NULL) {
166 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
167 }
168 if (_dirty_region_buffer != NULL) {
169 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
170 }
171 if (_in_dirty_region_buffer != NULL) {
172 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
173 }
174 if (_scan_top != NULL) {
175 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
176 }
177 }
178
179 void initialize(uint max_regions) {
180 assert(_iter_states == NULL, "Must not be initialized twice");
181 assert(_iter_claims == NULL, "Must not be initialized twice");
182 _max_regions = max_regions;
183 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
184 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
185 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
186 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
187 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
188 }
189
190 void reset() {
191 for (uint i = 0; i < _max_regions; i++) {
192 _iter_states[i] = Unclaimed;
193 _scan_top[i] = NULL;
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->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 G1DirtyCardQueueSet::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 G1GCPhaseTimes::GCParPhases phase,
314 uint worker_i) :
315 _g1h(G1CollectedHeap::heap()),
316 _ct(_g1h->card_table()),
317 _pss(pss),
318 _scan_objs_on_card_cl(scan_obj_on_card),
319 _scan_state(scan_state),
320 _phase(phase),
321 _worker_i(worker_i),
322 _cards_scanned(0),
323 _cards_claimed(0),
324 _cards_skipped(0),
325 _rem_set_root_scan_time(),
326 _rem_set_trim_partially_time(),
327 _strong_code_root_scan_time(),
328 _strong_code_trim_partially_time() {
329 }
330
331 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
332 _ct->set_card_claimed(card_index);
333 _scan_state->add_dirty_region(region_idx_for_card);
334 }
335
336 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
337 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
338 assert(!card_region->is_young(), "Should not scan card in young region %u", region_idx_for_card);
339 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
340 _scan_objs_on_card_cl->trim_queue_partially();
341 _cards_scanned++;
342 }
343
344 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
345 EventGCPhaseParallel event;
346 uint const region_idx = r->hrm_index();
347
348 if (_scan_state->claim_iter(region_idx)) {
349 // If we ever free the collection set concurrently, we should also
350 // clear the card table concurrently therefore we won't need to
351 // add regions of the collection set to the dirty cards region.
352 _scan_state->add_dirty_region(region_idx);
353 }
354
355 if (r->rem_set()->cardset_is_empty()) {
356 return;
357 }
358
359 // We claim cards in blocks so as to reduce the contention.
360 size_t const block_size = G1RSetScanBlockSize;
361
362 HeapRegionRemSetIterator iter(r->rem_set());
363 size_t card_index;
364
365 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
366 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
367 if (current_card >= claimed_card_block + block_size) {
368 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
369 }
370 if (current_card < claimed_card_block) {
371 _cards_skipped++;
372 continue;
373 }
374 _cards_claimed++;
375
376 HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
377 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
378
379 #ifdef ASSERT
380 HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
381 assert(hr == NULL || hr->is_in_reserved(card_start),
382 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
383 #endif
384 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
385 if (card_start >= top) {
386 continue;
387 }
388
389 // If the card is dirty, then G1 will scan it during Update RS.
390 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
391 continue;
392 }
393
394 // We claim lazily (so races are possible but they're benign), which reduces the
395 // number of duplicate scans (the rsets of the regions in the cset can intersect).
396 // Claim the card after checking bounds above: the remembered set may contain
397 // random cards into current survivor, and we would then have an incorrectly
398 // claimed card in survivor space. Card table clear does not reset the card table
399 // of survivor space regions.
400 claim_card(card_index, region_idx_for_card);
401
402 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
403
404 scan_card(mr, region_idx_for_card);
405 }
406 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
407 }
408
409 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
410 EventGCPhaseParallel event;
411 // We pass a weak code blobs closure to the remembered set scanning because we want to avoid
412 // treating the nmethods visited to act as roots for concurrent marking.
413 // We only want to make sure that the oops in the nmethods are adjusted with regard to the
414 // objects copied by the current evacuation.
415 r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
416 event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
417 }
418
419 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
420 assert(r->in_collection_set(),
421 "Should only be called on elements of the collection set but region %u is not.",
422 r->hrm_index());
423 uint const region_idx = r->hrm_index();
424
425 // Do an early out if we know we are complete.
426 if (_scan_state->iter_is_complete(region_idx)) {
427 return false;
428 }
429
430 {
431 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
432 scan_rem_set_roots(r);
433 }
434
435 if (_scan_state->set_iter_complete(region_idx)) {
436 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
437 // Scan the strong code root list attached to the current region
438 scan_strong_code_roots(r);
439 }
440 return false;
441 }
442
443 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
444 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
445 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, G1GCPhaseTimes::ScanRS, worker_i);
446 _g1h->collection_set_iterate_from(&cl, worker_i);
447
448 G1GCPhaseTimes* p = _g1p->phase_times();
449
450 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
451 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
452
453 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
454 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
455 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
456
457 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
458 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
459 }
460
461 // Closure used for updating rem sets. Only called during an evacuation pause.
462 class G1RefineCardClosure: public G1CardTableEntryClosure {
463 G1RemSet* _g1rs;
464 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
465
466 size_t _cards_scanned;
467 size_t _cards_skipped;
468 public:
469 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
470 _g1rs(g1h->rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
471 {}
472
473 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
474 // The only time we care about recording cards that
475 // contain references that point into the collection set
476 // is during RSet updating within an evacuation pause.
477 // In this case worker_i should be the id of a GC worker thread.
478 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
479
480 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
481
482 if (card_scanned) {
483 _update_rs_cl->trim_queue_partially();
484 _cards_scanned++;
485 } else {
486 _cards_skipped++;
487 }
488 return true;
489 }
490
491 size_t cards_scanned() const { return _cards_scanned; }
492 size_t cards_skipped() const { return _cards_skipped; }
493 };
494
495 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
496 G1GCPhaseTimes* p = _g1p->phase_times();
497
498 // Apply closure to log entries in the HCC.
499 if (G1HotCardCache::default_use_cache()) {
500 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
501
502 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss);
503 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
504 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
505 }
506
507 // Now apply the closure to all remaining log entries.
508 {
509 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
510
511 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss);
512 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
513 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
514
515 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
516 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
517 }
518 }
519
520 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
521 update_rem_set(pss, worker_i);
522 scan_rem_set(pss, worker_i);;
523 }
524
525 void G1RemSet::prepare_for_oops_into_collection_set_do() {
526 G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
527 dcqs.concatenate_logs();
528
529 _scan_state->reset();
530 }
531
532 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
533 G1GCPhaseTimes* phase_times = _g1h->phase_times();
534
535 // Set all cards back to clean.
536 double start = os::elapsedTime();
537 _scan_state->clear_card_table(_g1h->workers());
538 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
539 }
540
541 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
542 #ifdef ASSERT
543 G1CollectedHeap* g1h = G1CollectedHeap::heap();
544 assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
545 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
546 p2i(card_ptr),
547 ct->index_for(ct->addr_for(card_ptr)),
548 p2i(ct->addr_for(card_ptr)),
549 g1h->addr_to_region(ct->addr_for(card_ptr)));
550 #endif
551 }
552
553 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
554 uint worker_i) {
555 assert(!_g1h->is_gc_active(), "Only call concurrently");
556
557 // Construct the region representing the card.
558 HeapWord* start = _ct->addr_for(card_ptr);
559 // And find the region containing it.
560 HeapRegion* r = _g1h->heap_region_containing_or_null(start);
561
562 // If this is a (stale) card into an uncommitted region, exit.
563 if (r == NULL) {
564 return;
565 }
566
567 check_card_ptr(card_ptr, _ct);
568
569 // If the card is no longer dirty, nothing to do.
570 if (*card_ptr != G1CardTable::dirty_card_val()) {
571 return;
572 }
573
574 // This check is needed for some uncommon cases where we should
575 // ignore the card.
576 //
577 // The region could be young. Cards for young regions are
578 // distinctly marked (set to g1_young_gen), so the post-barrier will
579 // filter them out. However, that marking is performed
580 // concurrently. A write to a young object could occur before the
581 // card has been marked young, slipping past the filter.
582 //
583 // The card could be stale, because the region has been freed since
584 // the card was recorded. In this case the region type could be
585 // anything. If (still) free or (reallocated) young, just ignore
586 // it. If (reallocated) old or humongous, the later card trimming
587 // and additional checks in iteration may detect staleness. At
588 // worst, we end up processing a stale card unnecessarily.
589 //
590 // In the normal (non-stale) case, the synchronization between the
591 // enqueueing of the card and processing it here will have ensured
592 // we see the up-to-date region type here.
593 if (!r->is_old_or_humongous_or_archive()) {
594 return;
595 }
596
597 // The result from the hot card cache insert call is either:
598 // * pointer to the current card
599 // (implying that the current card is not 'hot'),
600 // * null
601 // (meaning we had inserted the card ptr into the "hot" card cache,
602 // which had some headroom),
603 // * a pointer to a "hot" card that was evicted from the "hot" cache.
604 //
605
606 if (_hot_card_cache->use_cache()) {
607 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
608
609 const jbyte* orig_card_ptr = card_ptr;
610 card_ptr = _hot_card_cache->insert(card_ptr);
611 if (card_ptr == NULL) {
612 // There was no eviction. Nothing to do.
613 return;
614 } else if (card_ptr != orig_card_ptr) {
615 // Original card was inserted and an old card was evicted.
616 start = _ct->addr_for(card_ptr);
617 r = _g1h->heap_region_containing(start);
618
619 // Check whether the region formerly in the cache should be
620 // ignored, as discussed earlier for the original card. The
621 // region could have been freed while in the cache.
622 if (!r->is_old_or_humongous_or_archive()) {
623 return;
624 }
625 } // Else we still have the original card.
626 }
627
628 // Trim the region designated by the card to what's been allocated
629 // in the region. The card could be stale, or the card could cover
630 // (part of) an object at the end of the allocated space and extend
631 // beyond the end of allocation.
632
633 // Non-humongous objects are only allocated in the old-gen during
634 // GC, so if region is old then top is stable. Humongous object
635 // allocation sets top last; if top has not yet been set, this is
636 // a stale card and we'll end up with an empty intersection. If
637 // this is not a stale card, the synchronization between the
638 // enqueuing of the card and processing it here will have ensured
639 // we see the up-to-date top here.
640 HeapWord* scan_limit = r->top();
641
642 if (scan_limit <= start) {
643 // If the trimmed region is empty, the card must be stale.
644 return;
645 }
646
647 // Okay to clean and process the card now. There are still some
648 // stale card cases that may be detected by iteration and dealt with
649 // as iteration failure.
650 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
651
652 // This fence serves two purposes. First, the card must be cleaned
653 // before processing the contents. Second, we can't proceed with
654 // processing until after the read of top, for synchronization with
655 // possibly concurrent humongous object allocation. It's okay that
656 // reading top and reading type were racy wrto each other. We need
657 // both set, in any order, to proceed.
658 OrderAccess::fence();
659
660 // Don't use addr_for(card_ptr + 1) which can ask for
661 // a card beyond the heap.
662 HeapWord* end = start + G1CardTable::card_size_in_words;
663 MemRegion dirty_region(start, MIN2(scan_limit, end));
664 assert(!dirty_region.is_empty(), "sanity");
665
666 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
667
668 bool card_processed =
669 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
670
671 // If unable to process the card then we encountered an unparsable
672 // part of the heap (e.g. a partially allocated object) while
673 // processing a stale card. Despite the card being stale, redirty
674 // and re-enqueue, because we've already cleaned the card. Without
675 // this we could incorrectly discard a non-stale card.
676 if (!card_processed) {
677 // The card might have gotten re-dirtied and re-enqueued while we
678 // worked. (In fact, it's pretty likely.)
679 if (*card_ptr != G1CardTable::dirty_card_val()) {
680 *card_ptr = G1CardTable::dirty_card_val();
681 MutexLockerEx x(Shared_DirtyCardQ_lock,
682 Mutex::_no_safepoint_check_flag);
683 G1DirtyCardQueue* sdcq =
684 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue();
685 sdcq->enqueue(card_ptr);
686 }
687 } else {
688 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
689 }
690 }
691
692 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
693 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
694 assert(_g1h->is_gc_active(), "Only call during GC");
695
696 // Construct the region representing the card.
697 HeapWord* card_start = _ct->addr_for(card_ptr);
698 // And find the region containing it.
699 uint const card_region_idx = _g1h->addr_to_region(card_start);
700
701 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
702 if (scan_limit == NULL) {
703 // This is a card into an uncommitted region. We need to bail out early as we
704 // should not access the corresponding card table entry.
705 return false;
706 }
707
708 check_card_ptr(card_ptr, _ct);
709
710 // If the card is no longer dirty, nothing to do. This covers cards that were already
711 // scanned as parts of the remembered sets.
712 if (*card_ptr != G1CardTable::dirty_card_val()) {
713 return false;
714 }
715
716 // We claim lazily (so races are possible but they're benign), which reduces the
717 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
718 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
719
720 _scan_state->add_dirty_region(card_region_idx);
721 if (scan_limit <= card_start) {
722 // If the card starts above the area in the region containing objects to scan, skip it.
723 return false;
724 }
725
726 // Don't use addr_for(card_ptr + 1) which can ask for
727 // a card beyond the heap.
728 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
729 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
730 assert(!dirty_region.is_empty(), "sanity");
731
732 HeapRegion* const card_region = _g1h->region_at(card_region_idx);
733 assert(!card_region->is_young(), "Should not scan card in young region %u", card_region_idx);
734 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
735 assert(card_processed, "must be");
736 return true;
737 }
738
739 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
740 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
741 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
742
743 G1RemSetSummary current(this);
744 _prev_period_summary.subtract_from(¤t);
745
746 Log(gc, remset) log;
747 log.trace("%s", header);
748 ResourceMark rm;
749 LogStream ls(log.trace());
750 _prev_period_summary.print_on(&ls);
751
752 _prev_period_summary.set(¤t);
753 }
754 }
755
756 void G1RemSet::print_summary_info() {
757 Log(gc, remset, exit) log;
758 if (log.is_trace()) {
759 log.trace(" Cumulative RS summary");
760 G1RemSetSummary current(this);
761 ResourceMark rm;
762 LogStream ls(log.trace());
763 current.print_on(&ls);
764 }
765 }
766
767 class G1RebuildRemSetTask: public AbstractGangTask {
768 // Aggregate the counting data that was constructed concurrently
769 // with marking.
770 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
771 G1ConcurrentMark* _cm;
772 G1RebuildRemSetClosure _update_cl;
773
774 // Applies _update_cl to the references of the given object, limiting objArrays
775 // to the given MemRegion. Returns the amount of words actually scanned.
776 size_t scan_for_references(oop const obj, MemRegion mr) {
777 size_t const obj_size = obj->size();
778 // All non-objArrays and objArrays completely within the mr
779 // can be scanned without passing the mr.
780 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
781 obj->oop_iterate(&_update_cl);
782 return obj_size;
783 }
784 // This path is for objArrays crossing the given MemRegion. Only scan the
785 // area within the MemRegion.
786 obj->oop_iterate(&_update_cl, mr);
787 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
788 }
789
790 // A humongous object is live (with respect to the scanning) either
791 // a) it is marked on the bitmap as such
792 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
793 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
794 return bitmap->is_marked(humongous_obj) || (tars > tams);
795 }
796
797 // Iterator over the live objects within the given MemRegion.
798 class LiveObjIterator : public StackObj {
799 const G1CMBitMap* const _bitmap;
800 const HeapWord* _tams;
801 const MemRegion _mr;
802 HeapWord* _current;
803
804 bool is_below_tams() const {
805 return _current < _tams;
806 }
807
808 bool is_live(HeapWord* obj) const {
809 return !is_below_tams() || _bitmap->is_marked(obj);
810 }
811
812 HeapWord* bitmap_limit() const {
813 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
814 }
815
816 void move_if_below_tams() {
817 if (is_below_tams() && has_next()) {
818 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
819 }
820 }
821 public:
822 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
823 _bitmap(bitmap),
824 _tams(tams),
825 _mr(mr),
826 _current(first_oop_into_mr) {
827
828 assert(_current <= _mr.start(),
829 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
830 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
831
832 // Step to the next live object within the MemRegion if needed.
833 if (is_live(_current)) {
834 // Non-objArrays were scanned by the previous part of that region.
835 if (_current < mr.start() && !oop(_current)->is_objArray()) {
836 _current += oop(_current)->size();
837 // We might have positioned _current on a non-live object. Reposition to the next
838 // live one if needed.
839 move_if_below_tams();
840 }
841 } else {
842 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
843 // immediately.
844 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
845 assert(_current == _mr.end() || is_live(_current),
846 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
847 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
848 }
849 }
850
851 void move_to_next() {
852 _current += next()->size();
853 move_if_below_tams();
854 }
855
856 oop next() const {
857 oop result = oop(_current);
858 assert(is_live(_current),
859 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
860 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
861 return result;
862 }
863
864 bool has_next() const {
865 return _current < _mr.end();
866 }
867 };
868
869 // Rebuild remembered sets in the part of the region specified by mr and hr.
870 // Objects between the bottom of the region and the TAMS are checked for liveness
871 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
872 // Returns the number of live words between bottom and TAMS.
873 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
874 HeapWord* const top_at_mark_start,
875 HeapWord* const top_at_rebuild_start,
876 HeapRegion* hr,
877 MemRegion mr) {
878 size_t marked_words = 0;
879
880 if (hr->is_humongous()) {
881 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
882 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
883 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
884 // however in case of humongous objects it is sufficient to scan the encompassing
885 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
886 // two areas will be zero sized. I.e. TAMS is either
887 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
888 // value: this would mean that TAMS points somewhere into the object.
889 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
890 "More than one object in the humongous region?");
891 humongous_obj->oop_iterate(&_update_cl, mr);
892 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
893 } else {
894 return 0;
895 }
896 }
897
898 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
899 oop obj = it.next();
900 size_t scanned_size = scan_for_references(obj, mr);
901 if ((HeapWord*)obj < top_at_mark_start) {
902 marked_words += scanned_size;
903 }
904 }
905
906 return marked_words * HeapWordSize;
907 }
908 public:
909 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
910 G1ConcurrentMark* cm,
911 uint worker_id) :
912 HeapRegionClosure(),
913 _cm(cm),
914 _update_cl(g1h, worker_id) { }
915
916 bool do_heap_region(HeapRegion* hr) {
917 if (_cm->has_aborted()) {
918 return true;
919 }
920
921 uint const region_idx = hr->hrm_index();
922 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
923 assert(top_at_rebuild_start_check == NULL ||
924 top_at_rebuild_start_check > hr->bottom(),
925 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
926 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
927
928 size_t total_marked_bytes = 0;
929 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
930
931 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
932
933 HeapWord* cur = hr->bottom();
934 while (cur < hr->end()) {
935 // After every iteration (yield point) we need to check whether the region's
936 // TARS changed due to e.g. eager reclaim.
937 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
938 if (top_at_rebuild_start == NULL) {
939 return false;
940 }
941
942 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
943 if (next_chunk.is_empty()) {
944 break;
945 }
946
947 const Ticks start = Ticks::now();
948 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
949 top_at_mark_start,
950 top_at_rebuild_start,
951 hr,
952 next_chunk);
953 Tickspan time = Ticks::now() - start;
954
955 log_trace(gc, remset, tracking)("Rebuilt region %u "
956 "live " SIZE_FORMAT " "
957 "time %.3fms "
958 "marked bytes " SIZE_FORMAT " "
959 "bot " PTR_FORMAT " "
960 "TAMS " PTR_FORMAT " "
961 "TARS " PTR_FORMAT,
962 region_idx,
963 _cm->liveness(region_idx) * HeapWordSize,
964 time.seconds() * 1000.0,
965 marked_bytes,
966 p2i(hr->bottom()),
967 p2i(top_at_mark_start),
968 p2i(top_at_rebuild_start));
969
970 if (marked_bytes > 0) {
971 total_marked_bytes += marked_bytes;
972 }
973 cur += chunk_size_in_words;
974
975 _cm->do_yield_check();
976 if (_cm->has_aborted()) {
977 return true;
978 }
979 }
980 // In the final iteration of the loop the region might have been eagerly reclaimed.
981 // Simply filter out those regions. We can not just use region type because there
982 // might have already been new allocations into these regions.
983 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
984 assert(top_at_rebuild_start == NULL ||
985 total_marked_bytes == hr->marked_bytes(),
986 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
987 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
988 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
989 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
990 // Abort state may have changed after the yield check.
991 return _cm->has_aborted();
992 }
993 };
994
995 HeapRegionClaimer _hr_claimer;
996 G1ConcurrentMark* _cm;
997
998 uint _worker_id_offset;
999 public:
1000 G1RebuildRemSetTask(G1ConcurrentMark* cm,
1001 uint n_workers,
1002 uint worker_id_offset) :
1003 AbstractGangTask("G1 Rebuild Remembered Set"),
1004 _hr_claimer(n_workers),
1005 _cm(cm),
1006 _worker_id_offset(worker_id_offset) {
1007 }
1008
1009 void work(uint worker_id) {
1010 SuspendibleThreadSetJoiner sts_join;
1011
1012 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1013
1014 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1015 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1016 }
1017 };
1018
1019 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1020 WorkGang* workers,
1021 uint worker_id_offset) {
1022 uint num_workers = workers->active_workers();
1023
1024 G1RebuildRemSetTask cl(cm,
1025 num_workers,
1026 worker_id_offset);
1027 workers->run_task(&cl, num_workers);
1028 }