1 /*
2 * Copyright (c) 2001, 2017, 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/concurrentG1Refine.hpp"
27 #include "gc/g1/dirtyCardQueue.hpp"
28 #include "gc/g1/g1BlockOffsetTable.inline.hpp"
29 #include "gc/g1/g1CollectedHeap.inline.hpp"
30 #include "gc/g1/g1FromCardCache.hpp"
31 #include "gc/g1/g1GCPhaseTimes.hpp"
32 #include "gc/g1/g1HotCardCache.hpp"
33 #include "gc/g1/g1OopClosures.inline.hpp"
34 #include "gc/g1/g1RemSet.inline.hpp"
35 #include "gc/g1/g1SATBCardTableModRefBS.inline.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 "memory/iterator.hpp"
41 #include "memory/resourceArea.hpp"
42 #include "oops/oop.inline.hpp"
43 #include "utilities/globalDefinitions.hpp"
44 #include "utilities/intHisto.hpp"
45 #include "utilities/stack.inline.hpp"
46
47 // Collects information about the overall remembered set scan progress during an evacuation.
48 class G1RemSetScanState : public CHeapObj<mtGC> {
49 private:
50 class G1ClearCardTableTask : public AbstractGangTask {
51 G1CollectedHeap* _g1h;
52 uint* _dirty_region_list;
53 size_t _num_dirty_regions;
54 size_t _chunk_length;
55
56 size_t volatile _cur_dirty_regions;
57 public:
58 G1ClearCardTableTask(G1CollectedHeap* g1h,
59 uint* dirty_region_list,
60 size_t num_dirty_regions,
61 size_t chunk_length) :
62 AbstractGangTask("G1 Clear Card Table Task"),
63 _g1h(g1h),
64 _dirty_region_list(dirty_region_list),
65 _num_dirty_regions(num_dirty_regions),
66 _chunk_length(chunk_length),
67 _cur_dirty_regions(0) {
68
69 assert(chunk_length > 0, "must be");
70 }
71
72 static size_t chunk_size() { return M; }
73
74 void work(uint worker_id) {
75 G1SATBCardTableModRefBS* ct_bs = _g1h->g1_barrier_set();
76
77 while (_cur_dirty_regions < _num_dirty_regions) {
78 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
79 size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
80
81 for (size_t i = next; i < max; i++) {
82 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
83 if (!r->is_survivor()) {
84 ct_bs->clear(MemRegion(r->bottom(), r->end()));
85 }
86 }
87 }
88 }
89 };
90
91 size_t _max_regions;
92
93 // Scan progress for the remembered set of a single region. Transitions from
94 // Unclaimed -> Claimed -> Complete.
95 // At each of the transitions the thread that does the transition needs to perform
96 // some special action once. This is the reason for the extra "Claimed" state.
97 typedef jint G1RemsetIterState;
98
99 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
100 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned.
101 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned.
102
103 G1RemsetIterState volatile* _iter_states;
104 // The current location where the next thread should continue scanning in a region's
105 // remembered set.
106 size_t volatile* _iter_claims;
107
108 // Temporary buffer holding the regions we used to store remembered set scan duplicate
109 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
110 uint* _dirty_region_buffer;
111
112 typedef jbyte IsDirtyRegionState;
113 static const IsDirtyRegionState Clean = 0;
114 static const IsDirtyRegionState Dirty = 1;
115 // Holds a flag for every region whether it is in the _dirty_region_buffer already
116 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools.
117 IsDirtyRegionState* _in_dirty_region_buffer;
118 size_t _cur_dirty_region;
119 public:
120 G1RemSetScanState() :
121 _max_regions(0),
122 _iter_states(NULL),
123 _iter_claims(NULL),
124 _dirty_region_buffer(NULL),
125 _in_dirty_region_buffer(NULL),
126 _cur_dirty_region(0) {
127
128 }
129
130 ~G1RemSetScanState() {
131 if (_iter_states != NULL) {
132 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
133 }
134 if (_iter_claims != NULL) {
135 FREE_C_HEAP_ARRAY(size_t, _iter_claims);
136 }
137 if (_dirty_region_buffer != NULL) {
138 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
139 }
140 if (_in_dirty_region_buffer != NULL) {
141 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer);
142 }
143 }
144
145 void initialize(uint max_regions) {
146 assert(_iter_states == NULL, "Must not be initialized twice");
147 assert(_iter_claims == NULL, "Must not be initialized twice");
148 _max_regions = max_regions;
149 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
150 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
151 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
152 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC);
153 }
154
155 void reset() {
156 for (uint i = 0; i < _max_regions; i++) {
157 _iter_states[i] = Unclaimed;
158 }
159 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
160 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState));
161 _cur_dirty_region = 0;
162 }
163
164 // Attempt to claim the remembered set of the region for iteration. Returns true
165 // if this call caused the transition from Unclaimed to Claimed.
166 inline bool claim_iter(uint region) {
167 assert(region < _max_regions, "Tried to access invalid region %u", region);
168 if (_iter_states[region] != Unclaimed) {
169 return false;
170 }
171 jint res = Atomic::cmpxchg(Claimed, (jint*)(&_iter_states[region]), Unclaimed);
172 return (res == Unclaimed);
173 }
174
175 // Try to atomically sets the iteration state to "complete". Returns true for the
176 // thread that caused the transition.
177 inline bool set_iter_complete(uint region) {
178 if (iter_is_complete(region)) {
179 return false;
180 }
181 jint res = Atomic::cmpxchg(Complete, (jint*)(&_iter_states[region]), Claimed);
182 return (res == Claimed);
183 }
184
185 // Returns true if the region's iteration is complete.
186 inline bool iter_is_complete(uint region) const {
187 assert(region < _max_regions, "Tried to access invalid region %u", region);
188 return _iter_states[region] == Complete;
189 }
190
191 // The current position within the remembered set of the given region.
192 inline size_t iter_claimed(uint region) const {
193 assert(region < _max_regions, "Tried to access invalid region %u", region);
194 return _iter_claims[region];
195 }
196
197 // Claim the next block of cards within the remembered set of the region with
198 // step size.
199 inline size_t iter_claimed_next(uint region, size_t step) {
200 return Atomic::add(step, &_iter_claims[region]) - step;
201 }
202
203 void add_dirty_region(uint region) {
204 if (_in_dirty_region_buffer[region] == Dirty) {
205 return;
206 }
207
208 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean;
209 if (marked_as_dirty) {
210 size_t allocated = Atomic::add(1, &_cur_dirty_region) - 1;
211 _dirty_region_buffer[allocated] = region;
212 }
213 }
214
215 // Clear the card table of "dirty" regions.
216 void clear_card_table(WorkGang* workers) {
217 if (_cur_dirty_region == 0) {
218 return;
219 }
220
221 size_t const num_chunks = align_size_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
222 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
223 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
224
225 // Iterate over the dirty cards region list.
226 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
227
228 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
229 "units of work for " SIZE_FORMAT " regions.",
230 cl.name(), num_workers, num_chunks, _cur_dirty_region);
231 workers->run_task(&cl, num_workers);
232
233 #ifndef PRODUCT
234 // Need to synchronize with concurrent cleanup since it needs to
235 // finish its card table clearing before we can verify.
236 G1CollectedHeap::heap()->wait_while_free_regions_coming();
237 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
238 #endif
239 }
240 };
241
242 G1RemSet::G1RemSet(G1CollectedHeap* g1,
243 CardTableModRefBS* ct_bs,
244 G1HotCardCache* hot_card_cache) :
245 _g1(g1),
246 _scan_state(new G1RemSetScanState()),
247 _conc_refine_cards(0),
248 _ct_bs(ct_bs),
249 _g1p(_g1->g1_policy()),
250 _hot_card_cache(hot_card_cache),
251 _prev_period_summary(),
252 _into_cset_dirty_card_queue_set(false)
253 {
254 if (log_is_enabled(Trace, gc, remset)) {
255 _prev_period_summary.initialize(this);
256 }
257 // Initialize the card queue set used to hold cards containing
258 // references into the collection set.
259 _into_cset_dirty_card_queue_set.initialize(NULL, // Should never be called by the Java code
260 DirtyCardQ_CBL_mon,
261 DirtyCardQ_FL_lock,
262 -1, // never trigger processing
263 -1, // no limit on length
264 Shared_DirtyCardQ_lock,
265 &JavaThread::dirty_card_queue_set());
266 }
267
268 G1RemSet::~G1RemSet() {
269 if (_scan_state != NULL) {
270 delete _scan_state;
271 }
272 }
273
274 uint G1RemSet::num_par_rem_sets() {
275 return DirtyCardQueueSet::num_par_ids() + ConcurrentG1Refine::thread_num() + MAX2(ConcGCThreads, ParallelGCThreads);
276 }
277
278 void G1RemSet::initialize(size_t capacity, uint max_regions) {
279 G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
280 _scan_state->initialize(max_regions);
281 {
282 GCTraceTime(Debug, gc, marking)("Initialize Card Live Data");
283 _card_live_data.initialize(capacity, max_regions);
284 }
285 if (G1PretouchAuxiliaryMemory) {
286 GCTraceTime(Debug, gc, marking)("Pre-Touch Card Live Data");
287 _card_live_data.pretouch();
288 }
289 }
290
291 G1ScanRSClosure::G1ScanRSClosure(G1RemSetScanState* scan_state,
292 G1ParPushHeapRSClosure* push_heap_cl,
293 CodeBlobClosure* code_root_cl,
294 uint worker_i) :
295 _scan_state(scan_state),
296 _push_heap_cl(push_heap_cl),
297 _code_root_cl(code_root_cl),
298 _strong_code_root_scan_time_sec(0.0),
299 _cards(0),
300 _cards_done(0),
301 _worker_i(worker_i) {
302 _g1h = G1CollectedHeap::heap();
303 _bot = _g1h->bot();
304 _ct_bs = _g1h->g1_barrier_set();
305 _block_size = MAX2<size_t>(G1RSetScanBlockSize, 1);
306 }
307
308 void G1ScanRSClosure::scan_card(size_t index, HeapWord* card_start, HeapRegion *r) {
309 MemRegion card_region(card_start, BOTConstants::N_words);
310 MemRegion pre_gc_allocated(r->bottom(), r->scan_top());
311 MemRegion mr = pre_gc_allocated.intersection(card_region);
312 if (!mr.is_empty() && !_ct_bs->is_card_claimed(index)) {
313 // We make the card as "claimed" lazily (so races are possible
314 // but they're benign), which reduces the number of duplicate
315 // scans (the rsets of the regions in the cset can intersect).
316 _ct_bs->set_card_claimed(index);
317 _push_heap_cl->set_region(r);
318 r->oops_on_card_seq_iterate_careful<true>(mr, _push_heap_cl);
319 _cards_done++;
320 }
321 }
322
323 void G1ScanRSClosure::scan_strong_code_roots(HeapRegion* r) {
324 double scan_start = os::elapsedTime();
325 r->strong_code_roots_do(_code_root_cl);
326 _strong_code_root_scan_time_sec += (os::elapsedTime() - scan_start);
327 }
328
329 bool G1ScanRSClosure::doHeapRegion(HeapRegion* r) {
330 assert(r->in_collection_set(), "should only be called on elements of CS.");
331 uint region_idx = r->hrm_index();
332
333 if (_scan_state->iter_is_complete(region_idx)) {
334 return false;
335 }
336 if (_scan_state->claim_iter(region_idx)) {
337 // If we ever free the collection set concurrently, we should also
338 // clear the card table concurrently therefore we won't need to
339 // add regions of the collection set to the dirty cards region.
340 _scan_state->add_dirty_region(region_idx);
341 }
342
343 HeapRegionRemSetIterator iter(r->rem_set());
344 size_t card_index;
345
346 // We claim cards in block so as to reduce the contention. The block size is determined by
347 // the G1RSetScanBlockSize parameter.
348 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, _block_size);
349 for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
350 if (current_card >= claimed_card_block + _block_size) {
351 claimed_card_block = _scan_state->iter_claimed_next(region_idx, _block_size);
352 }
353 if (current_card < claimed_card_block) {
354 continue;
355 }
356 HeapWord* card_start = _g1h->bot()->address_for_index(card_index);
357
358 HeapRegion* card_region = _g1h->heap_region_containing(card_start);
359 _cards++;
360
361 _scan_state->add_dirty_region(card_region->hrm_index());
362
363 // If the card is dirty, then we will scan it during updateRS.
364 if (!card_region->in_collection_set() &&
365 !_ct_bs->is_card_dirty(card_index)) {
366 scan_card(card_index, card_start, card_region);
367 }
368 }
369 if (_scan_state->set_iter_complete(region_idx)) {
370 // Scan the strong code root list attached to the current region
371 scan_strong_code_roots(r);
372 }
373 return false;
374 }
375
376 size_t G1RemSet::scan_rem_set(G1ParPushHeapRSClosure* oops_in_heap_closure,
377 CodeBlobClosure* heap_region_codeblobs,
378 uint worker_i) {
379 double rs_time_start = os::elapsedTime();
380
381 G1ScanRSClosure cl(_scan_state, oops_in_heap_closure, heap_region_codeblobs, worker_i);
382 _g1->collection_set_iterate_from(&cl, worker_i);
383
384 double scan_rs_time_sec = (os::elapsedTime() - rs_time_start) -
385 cl.strong_code_root_scan_time_sec();
386
387 _g1p->phase_times()->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, scan_rs_time_sec);
388 _g1p->phase_times()->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time_sec());
389
390 return cl.cards_done();
391 }
392
393 // Closure used for updating RSets and recording references that
394 // point into the collection set. Only called during an
395 // evacuation pause.
396
397 class RefineRecordRefsIntoCSCardTableEntryClosure: public CardTableEntryClosure {
398 G1RemSet* _g1rs;
399 DirtyCardQueue* _into_cset_dcq;
400 G1ParPushHeapRSClosure* _cl;
401 public:
402 RefineRecordRefsIntoCSCardTableEntryClosure(G1CollectedHeap* g1h,
403 DirtyCardQueue* into_cset_dcq,
404 G1ParPushHeapRSClosure* cl) :
405 _g1rs(g1h->g1_rem_set()), _into_cset_dcq(into_cset_dcq), _cl(cl)
406 {}
407
408 bool do_card_ptr(jbyte* card_ptr, uint worker_i) {
409 // The only time we care about recording cards that
410 // contain references that point into the collection set
411 // is during RSet updating within an evacuation pause.
412 // In this case worker_i should be the id of a GC worker thread.
413 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
414 assert(worker_i < ParallelGCThreads, "should be a GC worker");
415
416 if (_g1rs->refine_card_during_gc(card_ptr, worker_i, _cl)) {
417 // 'card_ptr' contains references that point into the collection
418 // set. We need to record the card in the DCQS
419 // (_into_cset_dirty_card_queue_set)
420 // that's used for that purpose.
421 //
422 // Enqueue the card
423 _into_cset_dcq->enqueue(card_ptr);
424 }
425 return true;
426 }
427 };
428
429 void G1RemSet::update_rem_set(DirtyCardQueue* into_cset_dcq,
430 G1ParPushHeapRSClosure* oops_in_heap_closure,
431 uint worker_i) {
432 RefineRecordRefsIntoCSCardTableEntryClosure into_cset_update_rs_cl(_g1, into_cset_dcq, oops_in_heap_closure);
433
434 G1GCParPhaseTimesTracker x(_g1p->phase_times(), G1GCPhaseTimes::UpdateRS, worker_i);
435 if (G1HotCardCache::default_use_cache()) {
436 // Apply the closure to the entries of the hot card cache.
437 G1GCParPhaseTimesTracker y(_g1p->phase_times(), G1GCPhaseTimes::ScanHCC, worker_i);
438 _g1->iterate_hcc_closure(&into_cset_update_rs_cl, worker_i);
439 }
440 // Apply the closure to all remaining log entries.
441 _g1->iterate_dirty_card_closure(&into_cset_update_rs_cl, worker_i);
442 }
443
444 void G1RemSet::cleanupHRRS() {
445 HeapRegionRemSet::cleanup();
446 }
447
448 size_t G1RemSet::oops_into_collection_set_do(G1ParPushHeapRSClosure* cl,
449 CodeBlobClosure* heap_region_codeblobs,
450 uint worker_i) {
451 // A DirtyCardQueue that is used to hold cards containing references
452 // that point into the collection set. This DCQ is associated with a
453 // special DirtyCardQueueSet (see g1CollectedHeap.hpp). Under normal
454 // circumstances (i.e. the pause successfully completes), these cards
455 // are just discarded (there's no need to update the RSets of regions
456 // that were in the collection set - after the pause these regions
457 // are wholly 'free' of live objects. In the event of an evacuation
458 // failure the cards/buffers in this queue set are passed to the
459 // DirtyCardQueueSet that is used to manage RSet updates
460 DirtyCardQueue into_cset_dcq(&_into_cset_dirty_card_queue_set);
461
462 update_rem_set(&into_cset_dcq, cl, worker_i);
463 return scan_rem_set(cl, heap_region_codeblobs, worker_i);;
464 }
465
466 void G1RemSet::prepare_for_oops_into_collection_set_do() {
467 _g1->set_refine_cte_cl_concurrency(false);
468 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
469 dcqs.concatenate_logs();
470
471 _scan_state->reset();
472 }
473
474 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
475 G1GCPhaseTimes* phase_times = _g1->g1_policy()->phase_times();
476 // Cleanup after copy
477 _g1->set_refine_cte_cl_concurrency(true);
478
479 // Set all cards back to clean.
480 double start = os::elapsedTime();
481 _scan_state->clear_card_table(_g1->workers());
482 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
483
484 DirtyCardQueueSet& into_cset_dcqs = _into_cset_dirty_card_queue_set;
485
486 if (_g1->evacuation_failed()) {
487 double restore_remembered_set_start = os::elapsedTime();
488
489 // Restore remembered sets for the regions pointing into the collection set.
490 // We just need to transfer the completed buffers from the DirtyCardQueueSet
491 // used to hold cards that contain references that point into the collection set
492 // to the DCQS used to hold the deferred RS updates.
493 _g1->dirty_card_queue_set().merge_bufferlists(&into_cset_dcqs);
494 phase_times->record_evac_fail_restore_remsets((os::elapsedTime() - restore_remembered_set_start) * 1000.0);
495 }
496
497 // Free any completed buffers in the DirtyCardQueueSet used to hold cards
498 // which contain references that point into the collection.
499 _into_cset_dirty_card_queue_set.clear();
500 assert(_into_cset_dirty_card_queue_set.completed_buffers_num() == 0,
501 "all buffers should be freed");
502 _into_cset_dirty_card_queue_set.clear_n_completed_buffers();
503 }
504
505 inline void check_card_ptr(jbyte* card_ptr, CardTableModRefBS* ct_bs) {
506 #ifdef ASSERT
507 G1CollectedHeap* g1 = G1CollectedHeap::heap();
508 assert(g1->is_in_exact(ct_bs->addr_for(card_ptr)),
509 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
510 p2i(card_ptr),
511 ct_bs->index_for(ct_bs->addr_for(card_ptr)),
512 p2i(ct_bs->addr_for(card_ptr)),
513 g1->addr_to_region(ct_bs->addr_for(card_ptr)));
514 #endif
515 }
516
517 G1UpdateRSOrPushRefOopClosure::G1UpdateRSOrPushRefOopClosure(G1CollectedHeap* g1h,
518 G1ParPushHeapRSClosure* push_ref_cl,
519 bool record_refs_into_cset,
520 uint worker_i) :
521 _g1(g1h),
522 _from(NULL),
523 _record_refs_into_cset(record_refs_into_cset),
524 _has_refs_into_cset(false),
525 _push_ref_cl(push_ref_cl),
526 _worker_i(worker_i) { }
527
528 void G1RemSet::refine_card_concurrently(jbyte* card_ptr,
529 uint worker_i) {
530 assert(!_g1->is_gc_active(), "Only call concurrently");
531
532 check_card_ptr(card_ptr, _ct_bs);
533
534 // If the card is no longer dirty, nothing to do.
535 if (*card_ptr != CardTableModRefBS::dirty_card_val()) {
536 return;
537 }
538
539 // Construct the region representing the card.
540 HeapWord* start = _ct_bs->addr_for(card_ptr);
541 // And find the region containing it.
542 HeapRegion* r = _g1->heap_region_containing(start);
543
544 // This check is needed for some uncommon cases where we should
545 // ignore the card.
546 //
547 // The region could be young. Cards for young regions are
548 // distinctly marked (set to g1_young_gen), so the post-barrier will
549 // filter them out. However, that marking is performed
550 // concurrently. A write to a young object could occur before the
551 // card has been marked young, slipping past the filter.
552 //
553 // The card could be stale, because the region has been freed since
554 // the card was recorded. In this case the region type could be
555 // anything. If (still) free or (reallocated) young, just ignore
556 // it. If (reallocated) old or humongous, the later card trimming
557 // and additional checks in iteration may detect staleness. At
558 // worst, we end up processing a stale card unnecessarily.
559 //
560 // In the normal (non-stale) case, the synchronization between the
561 // enqueueing of the card and processing it here will have ensured
562 // we see the up-to-date region type here.
563 if (!r->is_old_or_humongous()) {
564 return;
565 }
566
567 // While we are processing RSet buffers during the collection, we
568 // actually don't want to scan any cards on the collection set,
569 // since we don't want to update remembered sets with entries that
570 // point into the collection set, given that live objects from the
571 // collection set are about to move and such entries will be stale
572 // very soon. This change also deals with a reliability issue which
573 // involves scanning a card in the collection set and coming across
574 // an array that was being chunked and looking malformed. Note,
575 // however, that if evacuation fails, we have to scan any objects
576 // that were not moved and create any missing entries.
577 if (r->in_collection_set()) {
578 return;
579 }
580
581 // The result from the hot card cache insert call is either:
582 // * pointer to the current card
583 // (implying that the current card is not 'hot'),
584 // * null
585 // (meaning we had inserted the card ptr into the "hot" card cache,
586 // which had some headroom),
587 // * a pointer to a "hot" card that was evicted from the "hot" cache.
588 //
589
590 if (_hot_card_cache->use_cache()) {
591 assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
592
593 const jbyte* orig_card_ptr = card_ptr;
594 card_ptr = _hot_card_cache->insert(card_ptr);
595 if (card_ptr == NULL) {
596 // There was no eviction. Nothing to do.
597 return;
598 } else if (card_ptr != orig_card_ptr) {
599 // Original card was inserted and an old card was evicted.
600 start = _ct_bs->addr_for(card_ptr);
601 r = _g1->heap_region_containing(start);
602
603 // Check whether the region formerly in the cache should be
604 // ignored, as discussed earlier for the original card. The
605 // region could have been freed while in the cache. The cset is
606 // not relevant here, since we're in concurrent phase.
607 if (!r->is_old_or_humongous()) {
608 return;
609 }
610 } // Else we still have the original card.
611 }
612
613 // Trim the region designated by the card to what's been allocated
614 // in the region. The card could be stale, or the card could cover
615 // (part of) an object at the end of the allocated space and extend
616 // beyond the end of allocation.
617
618 // Non-humongous objects are only allocated in the old-gen during
619 // GC, so if region is old then top is stable. Humongous object
620 // allocation sets top last; if top has not yet been set, this is
621 // a stale card and we'll end up with an empty intersection. If
622 // this is not a stale card, the synchronization between the
623 // enqueuing of the card and processing it here will have ensured
624 // we see the up-to-date top here.
625 HeapWord* scan_limit = r->top();
626
627 if (scan_limit <= start) {
628 // If the trimmed region is empty, the card must be stale.
629 return;
630 }
631
632 // Okay to clean and process the card now. There are still some
633 // stale card cases that may be detected by iteration and dealt with
634 // as iteration failure.
635 *const_cast<volatile jbyte*>(card_ptr) = CardTableModRefBS::clean_card_val();
636
637 // This fence serves two purposes. First, the card must be cleaned
638 // before processing the contents. Second, we can't proceed with
639 // processing until after the read of top, for synchronization with
640 // possibly concurrent humongous object allocation. It's okay that
641 // reading top and reading type were racy wrto each other. We need
642 // both set, in any order, to proceed.
643 OrderAccess::fence();
644
645 // Don't use addr_for(card_ptr + 1) which can ask for
646 // a card beyond the heap.
647 HeapWord* end = start + CardTableModRefBS::card_size_in_words;
648 MemRegion dirty_region(start, MIN2(scan_limit, end));
649 assert(!dirty_region.is_empty(), "sanity");
650
651 G1ConcurrentRefineOopClosure conc_refine_cl(_g1, worker_i);
652
653 bool card_processed =
654 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl);
655
656 // If unable to process the card then we encountered an unparsable
657 // part of the heap (e.g. a partially allocated object) while
658 // processing a stale card. Despite the card being stale, redirty
659 // and re-enqueue, because we've already cleaned the card. Without
660 // this we could incorrectly discard a non-stale card.
661 if (!card_processed) {
662 // The card might have gotten re-dirtied and re-enqueued while we
663 // worked. (In fact, it's pretty likely.)
664 if (*card_ptr != CardTableModRefBS::dirty_card_val()) {
665 *card_ptr = CardTableModRefBS::dirty_card_val();
666 MutexLockerEx x(Shared_DirtyCardQ_lock,
667 Mutex::_no_safepoint_check_flag);
668 DirtyCardQueue* sdcq =
669 JavaThread::dirty_card_queue_set().shared_dirty_card_queue();
670 sdcq->enqueue(card_ptr);
671 }
672 } else {
673 _conc_refine_cards++;
674 }
675 }
676
677 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr,
678 uint worker_i,
679 G1ParPushHeapRSClosure* oops_in_heap_closure) {
680 assert(_g1->is_gc_active(), "Only call during GC");
681
682 check_card_ptr(card_ptr, _ct_bs);
683
684 // If the card is no longer dirty, nothing to do. This covers cards that were already
685 // scanned as parts of the remembered sets.
686 if (*card_ptr != CardTableModRefBS::dirty_card_val()) {
687 // No need to return that this card contains refs that point
688 // into the collection set.
689 return false;
690 }
691
692 // Construct the region representing the card.
693 HeapWord* start = _ct_bs->addr_for(card_ptr);
694 // And find the region containing it.
695 HeapRegion* r = _g1->heap_region_containing(start);
696
697 // This check is needed for some uncommon cases where we should
698 // ignore the card.
699 //
700 // The region could be young. Cards for young regions are
701 // distinctly marked (set to g1_young_gen), so the post-barrier will
702 // filter them out. However, that marking is performed
703 // concurrently. A write to a young object could occur before the
704 // card has been marked young, slipping past the filter.
705 //
706 // The card could be stale, because the region has been freed since
707 // the card was recorded. In this case the region type could be
708 // anything. If (still) free or (reallocated) young, just ignore
709 // it. If (reallocated) old or humongous, the later card trimming
710 // and additional checks in iteration may detect staleness. At
711 // worst, we end up processing a stale card unnecessarily.
712 //
713 // In the normal (non-stale) case, the synchronization between the
714 // enqueueing of the card and processing it here will have ensured
715 // we see the up-to-date region type here.
716 if (!r->is_old_or_humongous()) {
717 return false;
718 }
719
720 // While we are processing RSet buffers during the collection, we
721 // actually don't want to scan any cards on the collection set,
722 // since we don't want to update remembered sets with entries that
723 // point into the collection set, given that live objects from the
724 // collection set are about to move and such entries will be stale
725 // very soon. This change also deals with a reliability issue which
726 // involves scanning a card in the collection set and coming across
727 // an array that was being chunked and looking malformed. Note,
728 // however, that if evacuation fails, we have to scan any objects
729 // that were not moved and create any missing entries.
730 if (r->in_collection_set()) {
731 return false;
732 }
733
734 // Trim the region designated by the card to what's been allocated
735 // in the region. The card could be stale, or the card could cover
736 // (part of) an object at the end of the allocated space and extend
737 // beyond the end of allocation.
738
739 // If we're in a STW GC, then a card might be in a GC alloc region
740 // and extend onto a GC LAB, which may not be parsable. Stop such
741 // at the "scan_top" of the region.
742 HeapWord* scan_limit = r->scan_top();
743
744 if (scan_limit <= start) {
745 // If the trimmed region is empty, the card must be stale.
746 return false;
747 }
748
749 // Okay to clean and process the card now. There are still some
750 // stale card cases that may be detected by iteration and dealt with
751 // as iteration failure.
752 *const_cast<volatile jbyte*>(card_ptr) = CardTableModRefBS::clean_card_val();
753
754 // Don't use addr_for(card_ptr + 1) which can ask for
755 // a card beyond the heap.
756 HeapWord* end = start + CardTableModRefBS::card_size_in_words;
757 MemRegion dirty_region(start, MIN2(scan_limit, end));
758 assert(!dirty_region.is_empty(), "sanity");
759
760 G1UpdateRSOrPushRefOopClosure update_rs_oop_cl(_g1,
761 oops_in_heap_closure,
762 true,
763 worker_i);
764 update_rs_oop_cl.set_from(r);
765
766 bool card_processed =
767 r->oops_on_card_seq_iterate_careful<true>(dirty_region, &update_rs_oop_cl);
768 assert(card_processed, "must be");
769 _conc_refine_cards++;
770
771 return update_rs_oop_cl.has_refs_into_cset();
772 }
773
774 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
775 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
776 (period_count % G1SummarizeRSetStatsPeriod == 0)) {
777
778 if (!_prev_period_summary.initialized()) {
779 _prev_period_summary.initialize(this);
780 }
781
782 G1RemSetSummary current;
783 current.initialize(this);
784 _prev_period_summary.subtract_from(¤t);
785
786 Log(gc, remset) log;
787 log.trace("%s", header);
788 ResourceMark rm;
789 _prev_period_summary.print_on(log.trace_stream());
790
791 _prev_period_summary.set(¤t);
792 }
793 }
794
795 void G1RemSet::print_summary_info() {
796 Log(gc, remset, exit) log;
797 if (log.is_trace()) {
798 log.trace(" Cumulative RS summary");
799 G1RemSetSummary current;
800 current.initialize(this);
801 ResourceMark rm;
802 current.print_on(log.trace_stream());
803 }
804 }
805
806 void G1RemSet::prepare_for_verify() {
807 if (G1HRRSFlushLogBuffersOnVerify &&
808 (VerifyBeforeGC || VerifyAfterGC)
809 && (!_g1->collector_state()->full_collection() || G1VerifyRSetsDuringFullGC)) {
810 cleanupHRRS();
811 _g1->set_refine_cte_cl_concurrency(false);
812 if (SafepointSynchronize::is_at_safepoint()) {
813 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
814 dcqs.concatenate_logs();
815 }
816
817 bool use_hot_card_cache = _hot_card_cache->use_cache();
818 _hot_card_cache->set_use_cache(false);
819
820 DirtyCardQueue into_cset_dcq(&_into_cset_dirty_card_queue_set);
821 update_rem_set(&into_cset_dcq, NULL, 0);
822 _into_cset_dirty_card_queue_set.clear();
823
824 _hot_card_cache->set_use_cache(use_hot_card_cache);
825 assert(JavaThread::dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
826 }
827 }
828
829 void G1RemSet::create_card_live_data(WorkGang* workers, G1CMBitMap* mark_bitmap) {
830 _card_live_data.create(workers, mark_bitmap);
831 }
832
833 void G1RemSet::finalize_card_live_data(WorkGang* workers, G1CMBitMap* mark_bitmap) {
834 _card_live_data.finalize(workers, mark_bitmap);
835 }
836
837 void G1RemSet::verify_card_live_data(WorkGang* workers, G1CMBitMap* bitmap) {
838 _card_live_data.verify(workers, bitmap);
839 }
840
841 void G1RemSet::clear_card_live_data(WorkGang* workers) {
842 _card_live_data.clear(workers);
843 }
844
845 #ifdef ASSERT
846 void G1RemSet::verify_card_live_data_is_clear() {
847 _card_live_data.verify_is_clear();
848 }
849 #endif