rev 49539 : imported patch 8178105-switch-at-remark
rev 49542 : imported patch 8178105-8200371-assert-problem
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/g1RemSet.hpp"
36 #include "gc/g1/heapRegion.inline.hpp"
37 #include "gc/g1/heapRegionManager.inline.hpp"
38 #include "gc/g1/heapRegionRemSet.hpp"
39 #include "gc/shared/gcTraceTime.inline.hpp"
40 #include "gc/shared/suspendibleThreadSet.hpp"
41 #include "memory/iterator.hpp"
42 #include "memory/resourceArea.hpp"
43 #include "oops/access.inline.hpp"
44 #include "oops/oop.inline.hpp"
45 #include "utilities/align.hpp"
46 #include "utilities/globalDefinitions.hpp"
47 #include "utilities/intHisto.hpp"
48 #include "utilities/stack.inline.hpp"
49 #include "utilities/ticks.inline.hpp"
50
51 // Collects information about the overall remembered set scan progress during an evacuation.
52 class G1RemSetScanState : public CHeapObj<mtGC> {
53 private:
54 class G1ClearCardTableTask : public AbstractGangTask {
55 G1CollectedHeap* _g1h;
56 uint* _dirty_region_list;
57 size_t _num_dirty_regions;
58 size_t _chunk_length;
59
60 size_t volatile _cur_dirty_regions;
61 public:
62 G1ClearCardTableTask(G1CollectedHeap* g1h,
63 uint* dirty_region_list,
64 size_t num_dirty_regions,
65 size_t chunk_length) :
66 AbstractGangTask("G1 Clear Card Table Task"),
67 _g1h(g1h),
68 _dirty_region_list(dirty_region_list),
69 _num_dirty_regions(num_dirty_regions),
70 _chunk_length(chunk_length),
71 _cur_dirty_regions(0) {
72
73 assert(chunk_length > 0, "must be");
74 }
75
76 static size_t chunk_size() { return M; }
77
78 void work(uint worker_id) {
79 while (_cur_dirty_regions < _num_dirty_regions) {
80 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
81 size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
82
83 for (size_t i = next; i < max; i++) {
84 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
85 if (!r->is_survivor()) {
86 r->clear_cardtable();
87 }
88 }
89 }
90 }
91 };
92
93 size_t _max_regions;
94
95 // Scan progress for the remembered set of a single region. Transitions from
96 // Unclaimed -> Claimed -> Complete.
97 // At each of the transitions the thread that does the transition needs to perform
98 // some special action once. This is the reason for the extra "Claimed" state.
99 typedef jint G1RemsetIterState;
100
101 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
102 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned.
103 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned.
104
105 G1RemsetIterState volatile* _iter_states;
106 // The current location where the next thread should continue scanning in a region's
107 // remembered set.
108 size_t volatile* _iter_claims;
109
110 // Temporary buffer holding the regions we used to store remembered set scan duplicate
111 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
112 uint* _dirty_region_buffer;
113
114 typedef jbyte IsDirtyRegionState;
115 static const IsDirtyRegionState Clean = 0;
116 static const IsDirtyRegionState Dirty = 1;
117 // Holds a flag for every region whether it is in the _dirty_region_buffer already
118 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools.
119 IsDirtyRegionState* _in_dirty_region_buffer;
120 size_t _cur_dirty_region;
121
122 // Creates a snapshot of the current _top values at the start of collection to
123 // filter out card marks that we do not want to scan.
124 class G1ResetScanTopClosure : public HeapRegionClosure {
125 private:
126 HeapWord** _scan_top;
127 public:
128 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
129
130 virtual bool do_heap_region(HeapRegion* r) {
131 uint hrm_index = r->hrm_index();
132 if (!r->in_collection_set() && r->is_old_or_humongous()) {
133 _scan_top[hrm_index] = r->top();
134 } else {
135 _scan_top[hrm_index] = r->bottom();
136 }
137 return false;
138 }
139 };
140
141 // For each region, contains the maximum top() value to be used during this garbage
142 // collection. Subsumes common checks like filtering out everything but old and
143 // humongous regions outside the collection set.
144 // This is valid because we are not interested in scanning stray remembered set
145 // entries from free or archive regions.
146 HeapWord** _scan_top;
147 public:
148 G1RemSetScanState() :
149 _max_regions(0),
150 _iter_states(NULL),
151 _iter_claims(NULL),
152 _dirty_region_buffer(NULL),
153 _in_dirty_region_buffer(NULL),
154 _cur_dirty_region(0),
155 _scan_top(NULL) {
156 }
157
158 ~G1RemSetScanState() {
159 if (_iter_states != NULL) {
160 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
161 }
162 if (_iter_claims != NULL) {
163 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
164 }
165 if (_dirty_region_buffer != NULL) {
166 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
167 }
168 if (_in_dirty_region_buffer != NULL) {
169 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
170 }
171 if (_scan_top != NULL) {
172 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
173 }
174 }
175
176 void initialize(uint max_regions) {
177 assert(_iter_states == NULL, "Must not be initialized twice");
178 assert(_iter_claims == NULL, "Must not be initialized twice");
179 _max_regions = max_regions;
180 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
181 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
182 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
183 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
184 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
185 }
186
187 void reset() {
188 for (uint i = 0; i < _max_regions; i++) {
189 _iter_states[i] = Unclaimed;
190 }
191
192 G1ResetScanTopClosure cl(_scan_top);
193 G1CollectedHeap::heap()->heap_region_iterate(&cl);
194
195 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
196 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState));
197 _cur_dirty_region = 0;
198 }
199
200 // Attempt to claim the remembered set of the region for iteration. Returns true
201 // if this call caused the transition from Unclaimed to Claimed.
202 inline bool claim_iter(uint region) {
203 assert(region < _max_regions, "Tried to access invalid region %u", region);
204 if (_iter_states[region] != Unclaimed) {
205 return false;
206 }
207 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
208 return (res == Unclaimed);
209 }
210
211 // Try to atomically sets the iteration state to "complete". Returns true for the
212 // thread that caused the transition.
213 inline bool set_iter_complete(uint region) {
214 if (iter_is_complete(region)) {
215 return false;
216 }
217 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
218 return (res == Claimed);
219 }
220
221 // Returns true if the region's iteration is complete.
222 inline bool iter_is_complete(uint region) const {
223 assert(region < _max_regions, "Tried to access invalid region %u", region);
224 return _iter_states[region] == Complete;
225 }
226
227 // The current position within the remembered set of the given region.
228 inline size_t iter_claimed(uint region) const {
229 assert(region < _max_regions, "Tried to access invalid region %u", region);
230 return _iter_claims[region];
231 }
232
233 // Claim the next block of cards within the remembered set of the region with
234 // step size.
235 inline size_t iter_claimed_next(uint region, size_t step) {
236 return Atomic::add(step, &_iter_claims[region]) - step;
237 }
238
239 void add_dirty_region(uint region) {
240 if (_in_dirty_region_buffer[region] == Dirty) {
241 return;
242 }
243
244 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean;
245 if (marked_as_dirty) {
246 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
247 _dirty_region_buffer[allocated] = region;
248 }
249 }
250
251 HeapWord* scan_top(uint region_idx) const {
252 return _scan_top[region_idx];
253 }
254
255 // Clear the card table of "dirty" regions.
256 void clear_card_table(WorkGang* workers) {
257 if (_cur_dirty_region == 0) {
258 return;
259 }
260
261 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
262 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
263 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
264
265 // Iterate over the dirty cards region list.
266 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
267
268 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
269 "units of work for " SIZE_FORMAT " regions.",
270 cl.name(), num_workers, num_chunks, _cur_dirty_region);
271 workers->run_task(&cl, num_workers);
272
273 #ifndef PRODUCT
274 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
275 #endif
276 }
277 };
278
279 G1RemSet::G1RemSet(G1CollectedHeap* g1,
280 G1CardTable* ct,
281 G1HotCardCache* hot_card_cache) :
282 _g1(g1),
283 _scan_state(new G1RemSetScanState()),
284 _num_conc_refined_cards(0),
285 _ct(ct),
286 _g1p(_g1->g1_policy()),
287 _hot_card_cache(hot_card_cache),
288 _prev_period_summary() {
289 }
290
291 G1RemSet::~G1RemSet() {
292 if (_scan_state != NULL) {
293 delete _scan_state;
294 }
295 }
296
297 uint G1RemSet::num_par_rem_sets() {
298 return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
299 }
300
301 void G1RemSet::initialize(size_t capacity, uint max_regions) {
302 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
303 _scan_state->initialize(max_regions);
304 }
305
306 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
307 G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
308 CodeBlobClosure* code_root_cl,
309 uint worker_i) :
310 _scan_state(scan_state),
311 _scan_objs_on_card_cl(scan_obj_on_card),
312 _code_root_cl(code_root_cl),
313 _strong_code_root_scan_time_sec(0.0),
314 _cards_claimed(0),
315 _cards_scanned(0),
316 _cards_skipped(0),
317 _worker_i(worker_i) {
318 _g1h = G1CollectedHeap::heap();
319 _bot = _g1h->bot();
320 _ct = _g1h->card_table();
321 }
322
323 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
324 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
325 _scan_objs_on_card_cl->set_region(card_region);
326 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
327 _cards_scanned++;
328 }
329
330 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
331 double scan_start = os::elapsedTime();
332 r->strong_code_roots_do(_code_root_cl);
333 _strong_code_root_scan_time_sec += (os::elapsedTime() - scan_start);
334 }
335
336 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
337 _ct->set_card_claimed(card_index);
338 _scan_state->add_dirty_region(region_idx_for_card);
339 }
340
341 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
342 assert(r->in_collection_set(), "should only be called on elements of CS.");
343 uint region_idx = r->hrm_index();
344
345 if (_scan_state->iter_is_complete(region_idx)) {
346 return false;
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 // We claim cards in blocks so as to reduce the contention.
356 size_t const block_size = G1RSetScanBlockSize;
357
358 HeapRegionRemSetIterator iter(r->rem_set());
359 size_t card_index;
360
361 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
362 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
363 if (current_card >= claimed_card_block + block_size) {
364 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
365 }
366 if (current_card < claimed_card_block) {
367 _cards_skipped++;
368 continue;
369 }
370 _cards_claimed++;
371
372 // If the card is dirty, then G1 will scan it during Update RS.
373 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
374 continue;
375 }
376
377 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index);
378 uint const region_idx_for_card = _g1h->addr_to_region(card_start);
379
380 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start),
381 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
382 HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
383 if (card_start >= top) {
384 continue;
385 }
386
387 // We claim lazily (so races are possible but they're benign), which reduces the
388 // number of duplicate scans (the rsets of the regions in the cset can intersect).
389 // Claim the card after checking bounds above: the remembered set may contain
390 // random cards into current survivor, and we would then have an incorrectly
391 // claimed card in survivor space. Card table clear does not reset the card table
392 // of survivor space regions.
393 claim_card(card_index, region_idx_for_card);
394
395 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
396
397 scan_card(mr, region_idx_for_card);
398 }
399 if (_scan_state->set_iter_complete(region_idx)) {
400 // Scan the strong code root list attached to the current region
401 scan_strong_code_roots(r);
402 }
403 return false;
404 }
405
406 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss,
407 CodeBlobClosure* heap_region_codeblobs,
408 uint worker_i) {
409 double rs_time_start = os::elapsedTime();
410
411 G1ScanObjsDuringScanRSClosure scan_cl(_g1, pss);
412 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, heap_region_codeblobs, worker_i);
413 _g1->collection_set_iterate_from(&cl, worker_i);
414
415 double scan_rs_time_sec = (os::elapsedTime() - rs_time_start) -
416 cl.strong_code_root_scan_time_sec();
417
418 G1GCPhaseTimes* p = _g1p->phase_times();
419
420 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, scan_rs_time_sec);
421 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
422 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
423 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
424
425 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time_sec());
426 }
427
428 // Closure used for updating rem sets. Only called during an evacuation pause.
429 class G1RefineCardClosure: public CardTableEntryClosure {
430 G1RemSet* _g1rs;
431 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
432
433 size_t _cards_scanned;
434 size_t _cards_skipped;
435 public:
436 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
437 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
438 {}
439
440 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
441 // The only time we care about recording cards that
442 // contain references that point into the collection set
443 // is during RSet updating within an evacuation pause.
444 // In this case worker_i should be the id of a GC worker thread.
445 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
446
447 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
448
449 if (card_scanned) {
450 _cards_scanned++;
451 } else {
452 _cards_skipped++;
453 }
454 return true;
455 }
456
457 size_t cards_scanned() const { return _cards_scanned; }
458 size_t cards_skipped() const { return _cards_skipped; }
459 };
460
461 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
462 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1, pss, worker_i);
463 G1RefineCardClosure refine_card_cl(_g1, &update_rs_cl);
464
465 G1GCParPhaseTimesTracker x(_g1p->phase_times(), G1GCPhaseTimes::UpdateRS, worker_i);
466 if (G1HotCardCache::default_use_cache()) {
467 // Apply the closure to the entries of the hot card cache.
468 G1GCParPhaseTimesTracker y(_g1p->phase_times(), G1GCPhaseTimes::ScanHCC, worker_i);
469 _g1->iterate_hcc_closure(&refine_card_cl, worker_i);
470 }
471 // Apply the closure to all remaining log entries.
472 _g1->iterate_dirty_card_closure(&refine_card_cl, worker_i);
473
474 G1GCPhaseTimes* p = _g1p->phase_times();
475 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
476 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
477 }
478
479 void G1RemSet::cleanupHRRS() {
480 HeapRegionRemSet::cleanup();
481 }
482
483 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss,
484 CodeBlobClosure* heap_region_codeblobs,
485 uint worker_i) {
486 update_rem_set(pss, worker_i);
487 scan_rem_set(pss, heap_region_codeblobs, worker_i);;
488 }
489
490 void G1RemSet::prepare_for_oops_into_collection_set_do() {
491 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
492 dcqs.concatenate_logs();
493
494 _scan_state->reset();
495 }
496
497 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
498 G1GCPhaseTimes* phase_times = _g1->g1_policy()->phase_times();
499
500 // Set all cards back to clean.
501 double start = os::elapsedTime();
502 _scan_state->clear_card_table(_g1->workers());
503 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
504 }
505
506 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) {
507 #ifdef ASSERT
508 G1CollectedHeap* g1 = G1CollectedHeap::heap();
509 assert(g1->is_in_exact(ct->addr_for(card_ptr)),
510 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
511 p2i(card_ptr),
512 ct->index_for(ct->addr_for(card_ptr)),
513 p2i(ct->addr_for(card_ptr)),
514 g1->addr_to_region(ct->addr_for(card_ptr)));
515 #endif
516 }
517
518 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
519 uint worker_i) {
520 assert(!_g1->is_gc_active(), "Only call concurrently");
521
522 check_card_ptr(card_ptr, _ct);
523
524 // If the card is no longer dirty, nothing to do.
525 if (*card_ptr != G1CardTable::dirty_card_val()) {
526 return;
527 }
528
529 // Construct the region representing the card.
530 HeapWord* start = _ct->addr_for(card_ptr);
531 // And find the region containing it.
532 HeapRegion* r = _g1->heap_region_containing(start);
533
534 // This check is needed for some uncommon cases where we should
535 // ignore the card.
536 //
537 // The region could be young. Cards for young regions are
538 // distinctly marked (set to g1_young_gen), so the post-barrier will
539 // filter them out. However, that marking is performed
540 // concurrently. A write to a young object could occur before the
541 // card has been marked young, slipping past the filter.
542 //
543 // The card could be stale, because the region has been freed since
544 // the card was recorded. In this case the region type could be
545 // anything. If (still) free or (reallocated) young, just ignore
546 // it. If (reallocated) old or humongous, the later card trimming
547 // and additional checks in iteration may detect staleness. At
548 // worst, we end up processing a stale card unnecessarily.
549 //
550 // In the normal (non-stale) case, the synchronization between the
551 // enqueueing of the card and processing it here will have ensured
552 // we see the up-to-date region type here.
553 if (!r->is_old_or_humongous()) {
554 return;
555 }
556
557 // The result from the hot card cache insert call is either:
558 // * pointer to the current card
559 // (implying that the current card is not 'hot'),
560 // * null
561 // (meaning we had inserted the card ptr into the "hot" card cache,
562 // which had some headroom),
563 // * a pointer to a "hot" card that was evicted from the "hot" cache.
564 //
565
566 if (_hot_card_cache->use_cache()) {
567 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
568
569 const jbyte* orig_card_ptr = card_ptr;
570 card_ptr = _hot_card_cache->insert(card_ptr);
571 if (card_ptr == NULL) {
572 // There was no eviction. Nothing to do.
573 return;
574 } else if (card_ptr != orig_card_ptr) {
575 // Original card was inserted and an old card was evicted.
576 start = _ct->addr_for(card_ptr);
577 r = _g1->heap_region_containing(start);
578
579 // Check whether the region formerly in the cache should be
580 // ignored, as discussed earlier for the original card. The
581 // region could have been freed while in the cache.
582 if (!r->is_old_or_humongous()) {
583 return;
584 }
585 } // Else we still have the original card.
586 }
587
588 // Trim the region designated by the card to what's been allocated
589 // in the region. The card could be stale, or the card could cover
590 // (part of) an object at the end of the allocated space and extend
591 // beyond the end of allocation.
592
593 // Non-humongous objects are only allocated in the old-gen during
594 // GC, so if region is old then top is stable. Humongous object
595 // allocation sets top last; if top has not yet been set, this is
596 // a stale card and we'll end up with an empty intersection. If
597 // this is not a stale card, the synchronization between the
598 // enqueuing of the card and processing it here will have ensured
599 // we see the up-to-date top here.
600 HeapWord* scan_limit = r->top();
601
602 if (scan_limit <= start) {
603 // If the trimmed region is empty, the card must be stale.
604 return;
605 }
606
607 // Okay to clean and process the card now. There are still some
608 // stale card cases that may be detected by iteration and dealt with
609 // as iteration failure.
610 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val();
611
612 // This fence serves two purposes. First, the card must be cleaned
613 // before processing the contents. Second, we can't proceed with
614 // processing until after the read of top, for synchronization with
615 // possibly concurrent humongous object allocation. It's okay that
616 // reading top and reading type were racy wrto each other. We need
617 // both set, in any order, to proceed.
618 OrderAccess::fence();
619
620 // Don't use addr_for(card_ptr + 1) which can ask for
621 // a card beyond the heap.
622 HeapWord* end = start + G1CardTable::card_size_in_words;
623 MemRegion dirty_region(start, MIN2(scan_limit, end));
624 assert(!dirty_region.is_empty(), "sanity");
625
626 G1ConcurrentRefineOopClosure conc_refine_cl(_g1, worker_i);
627
628 bool card_processed =
629 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
630
631 // If unable to process the card then we encountered an unparsable
632 // part of the heap (e.g. a partially allocated object) while
633 // processing a stale card. Despite the card being stale, redirty
634 // and re-enqueue, because we've already cleaned the card. Without
635 // this we could incorrectly discard a non-stale card.
636 if (!card_processed) {
637 // The card might have gotten re-dirtied and re-enqueued while we
638 // worked. (In fact, it's pretty likely.)
639 if (*card_ptr != G1CardTable::dirty_card_val()) {
640 *card_ptr = G1CardTable::dirty_card_val();
641 MutexLockerEx x(Shared_DirtyCardQ_lock,
642 Mutex::_no_safepoint_check_flag);
643 DirtyCardQueue* sdcq =
644 JavaThread::dirty_card_queue_set().shared_dirty_card_queue();
645 sdcq->enqueue(card_ptr);
646 }
647 } else {
648 _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
649 }
650 }
651
652 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
653 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
654 assert(_g1->is_gc_active(), "Only call during GC");
655
656 check_card_ptr(card_ptr, _ct);
657
658 // If the card is no longer dirty, nothing to do. This covers cards that were already
659 // scanned as parts of the remembered sets.
660 if (*card_ptr != G1CardTable::dirty_card_val()) {
661 return false;
662 }
663
664 // We claim lazily (so races are possible but they're benign), which reduces the
665 // number of potential duplicate scans (multiple threads may enqueue the same card twice).
666 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
667
668 // Construct the region representing the card.
669 HeapWord* card_start = _ct->addr_for(card_ptr);
670 // And find the region containing it.
671 uint const card_region_idx = _g1->addr_to_region(card_start);
672
673 _scan_state->add_dirty_region(card_region_idx);
674 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
675 if (scan_limit <= card_start) {
676 // If the card starts above the area in the region containing objects to scan, skip it.
677 return false;
678 }
679
680 // Don't use addr_for(card_ptr + 1) which can ask for
681 // a card beyond the heap.
682 HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
683 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
684 assert(!dirty_region.is_empty(), "sanity");
685
686 HeapRegion* const card_region = _g1->region_at(card_region_idx);
687 update_rs_cl->set_region(card_region);
688 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
689 assert(card_processed, "must be");
690 return true;
691 }
692
693 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
694 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
695 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
696
697 G1RemSetSummary current(this);
698 _prev_period_summary.subtract_from(¤t);
699
700 Log(gc, remset) log;
701 log.trace("%s", header);
702 ResourceMark rm;
703 LogStream ls(log.trace());
704 _prev_period_summary.print_on(&ls);
705
706 _prev_period_summary.set(¤t);
707 }
708 }
709
710 void G1RemSet::print_summary_info() {
711 Log(gc, remset, exit) log;
712 if (log.is_trace()) {
713 log.trace(" Cumulative RS summary");
714 G1RemSetSummary current(this);
715 ResourceMark rm;
716 LogStream ls(log.trace());
717 current.print_on(&ls);
718 }
719 }
720
721 class G1RebuildRemSetTask: public AbstractGangTask {
722 // Aggregate the counting data that was constructed concurrently
723 // with marking.
724 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
725 G1ConcurrentMark* _cm;
726 G1RebuildRemSetClosure _update_cl;
727
728 // Applies _update_cl to the references of the given object, limiting objArrays
729 // to the given MemRegion. Returns the amount of words actually scanned.
730 size_t scan_for_references(oop const obj, MemRegion mr) {
731 size_t const obj_size = obj->size();
732 // All non-objArrays and objArrays completely within the mr
733 // can be scanned without passing the mr.
734 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
735 obj->oop_iterate(&_update_cl);
736 return obj_size;
737 }
738 // This path is for objArrays crossing the given MemRegion. Only scan the
739 // area within the MemRegion.
740 obj->oop_iterate(&_update_cl, mr);
741 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
742 }
743
744 // A humongous object is live (with respect to the scanning) either
745 // a) it is marked on the bitmap as such
746 // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
747 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
748 return bitmap->is_marked(humongous_obj) || (tars > tams);
749 }
750
751 // Iterator over the live objects within the given MemRegion.
752 class LiveObjIterator : public StackObj {
753 const G1CMBitMap* const _bitmap;
754 const HeapWord* _tams;
755 const MemRegion _mr;
756 HeapWord* _current;
757
758 bool is_below_tams() const {
759 return _current < _tams;
760 }
761
762 bool is_live(HeapWord* obj) const {
763 return !is_below_tams() || _bitmap->is_marked(obj);
764 }
765
766 HeapWord* bitmap_limit() const {
767 return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
768 }
769
770 void move_if_below_tams() {
771 if (is_below_tams() && has_next()) {
772 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
773 }
774 }
775 public:
776 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
777 _bitmap(bitmap),
778 _tams(tams),
779 _mr(mr),
780 _current(first_oop_into_mr) {
781
782 assert(_current <= _mr.start(),
783 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
784 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
785
786 // Step to the next live object within the MemRegion if needed.
787 if (is_live(_current)) {
788 // Non-objArrays were scanned by the previous part of that region.
789 if (_current < mr.start() && !oop(_current)->is_objArray()) {
790 _current += oop(_current)->size();
791 // We might have positioned _current on a non-live object. Reposition to the next
792 // live one if needed.
793 move_if_below_tams();
794 }
795 } else {
796 // The object at _current can only be dead if below TAMS, so we can use the bitmap.
797 // immediately.
798 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
799 assert(_current == _mr.end() || is_live(_current),
800 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
801 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
802 }
803 }
804
805 void move_to_next() {
806 _current += next()->size();
807 move_if_below_tams();
808 }
809
810 oop next() const {
811 oop result = oop(_current);
812 assert(is_live(_current),
813 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
814 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
815 return result;
816 }
817
818 bool has_next() const {
819 return _current < _mr.end();
820 }
821 };
822
823 // Rebuild remembered sets in the part of the region specified by mr and hr.
824 // Objects between the bottom of the region and the TAMS are checked for liveness
825 // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
826 // Returns the number of live words between bottom and TAMS.
827 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
828 HeapWord* const top_at_mark_start,
829 HeapWord* const top_at_rebuild_start,
830 HeapRegion* hr,
831 MemRegion mr) {
832 size_t marked_words = 0;
833
834 if (hr->is_humongous()) {
835 oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
836 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
837 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
838 // however in case of humongous objects it is sufficient to scan the encompassing
839 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
840 // two areas will be zero sized. I.e. TAMS is either
841 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
842 // value: this would mean that TAMS points somewhere into the object.
843 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
844 "More than one object in the humongous region?");
845 humongous_obj->oop_iterate(&_update_cl, mr);
846 return top_at_mark_start != hr->bottom() ? mr.byte_size() : 0;
847 } else {
848 return 0;
849 }
850 }
851
852 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
853 oop obj = it.next();
854 size_t scanned_size = scan_for_references(obj, mr);
855 if ((HeapWord*)obj < top_at_mark_start) {
856 marked_words += scanned_size;
857 }
858 }
859
860 return marked_words * HeapWordSize;
861 }
862 public:
863 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
864 G1ConcurrentMark* cm,
865 uint worker_id) :
866 HeapRegionClosure(),
867 _cm(cm),
868 _update_cl(g1h, worker_id) { }
869
870 bool do_heap_region(HeapRegion* hr) {
871 if (_cm->has_aborted()) {
872 return true;
873 }
874
875 uint const region_idx = hr->hrm_index();
876 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
877 assert(top_at_rebuild_start_check == NULL ||
878 top_at_rebuild_start_check > hr->bottom(),
879 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
880 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str());
881
882 size_t total_marked_bytes = 0;
883 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
884
885 HeapWord* const top_at_mark_start = hr->next_top_at_mark_start();
886
887 HeapWord* cur = hr->bottom();
888 while (cur < hr->end()) {
889 // After every iteration (yield point) we need to check whether the region's
890 // TARS changed due to e.g. eager reclaim.
891 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
892 if (top_at_rebuild_start == NULL) {
893 return false;
894 }
895
896 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
897 if (next_chunk.is_empty()) {
898 break;
899 }
900
901 const Ticks start = Ticks::now();
902 size_t marked_bytes = rebuild_rem_set_in_region(_cm->next_mark_bitmap(),
903 top_at_mark_start,
904 top_at_rebuild_start,
905 hr,
906 next_chunk);
907 Tickspan time = Ticks::now() - start;
908
909 log_trace(gc, remset, tracking)("Rebuilt region %u "
910 "live " SIZE_FORMAT " "
911 "time %.3fms "
912 "marked bytes " SIZE_FORMAT " "
913 "bot " PTR_FORMAT " "
914 "TAMS " PTR_FORMAT " "
915 "TARS " PTR_FORMAT,
916 region_idx,
917 _cm->liveness(region_idx) * HeapWordSize,
918 TicksToTimeHelper::seconds(time) * 1000.0,
919 marked_bytes,
920 p2i(hr->bottom()),
921 p2i(top_at_mark_start),
922 p2i(top_at_rebuild_start));
923
924 if (marked_bytes > 0) {
925 hr->add_to_marked_bytes(marked_bytes);
926 total_marked_bytes += marked_bytes;
927 }
928 cur += chunk_size_in_words;
929
930 _cm->do_yield_check();
931 if (_cm->has_aborted()) {
932 return true;
933 }
934 }
935 // In the final iteration of the loop the region might have been eagerly reclaimed.
936 // Simply filter out those regions. We can not just use region type because there
937 // might have already been new allocations into these regions.
938 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
939 assert(!hr->is_old() ||
940 top_at_rebuild_start == NULL ||
941 total_marked_bytes == _cm->liveness(region_idx) * HeapWordSize,
942 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match liveness during mark " SIZE_FORMAT " "
943 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
944 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), _cm->liveness(region_idx) * HeapWordSize,
945 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
946 // Abort state may have changed after the yield check.
947 return _cm->has_aborted();
948 }
949 };
950
951 HeapRegionClaimer _hr_claimer;
952 G1ConcurrentMark* _cm;
953
954 uint _worker_id_offset;
955 public:
956 G1RebuildRemSetTask(G1ConcurrentMark* cm,
957 uint n_workers,
958 uint worker_id_offset) :
959 AbstractGangTask("G1 Rebuild Remembered Set"),
960 _cm(cm),
961 _hr_claimer(n_workers),
962 _worker_id_offset(worker_id_offset) {
963 }
964
965 void work(uint worker_id) {
966 SuspendibleThreadSetJoiner sts_join;
967
968 G1CollectedHeap* g1h = G1CollectedHeap::heap();
969
970 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
971 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
972 }
973 };
974
975 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
976 WorkGang* workers,
977 uint worker_id_offset) {
978 uint num_workers = workers->active_workers();
979
980 G1RebuildRemSetTask cl(cm,
981 num_workers,
982 worker_id_offset);
983 workers->run_task(&cl, num_workers);
984 }
--- EOF ---