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/g1SharedDirtyCardQueue.hpp"
  39 #include "gc/g1/heapRegion.inline.hpp"
  40 #include "gc/g1/heapRegionManager.inline.hpp"
  41 #include "gc/g1/heapRegionRemSet.hpp"
  42 #include "gc/shared/gcTraceTime.inline.hpp"
  43 #include "gc/shared/suspendibleThreadSet.hpp"
  44 #include "jfr/jfrEvents.hpp"
  45 #include "memory/iterator.hpp"
  46 #include "memory/resourceArea.hpp"
  47 #include "oops/access.inline.hpp"
  48 #include "oops/oop.inline.hpp"
  49 #include "runtime/os.hpp"
  50 #include "utilities/align.hpp"
  51 #include "utilities/globalDefinitions.hpp"
  52 #include "utilities/stack.inline.hpp"
  53 #include "utilities/ticks.hpp"
  54 
  55 // Collects information about the overall remembered set scan progress during an evacuation.
  56 class G1RemSetScanState : public CHeapObj<mtGC> {
  57 private:
  58   class G1ClearCardTableTask : public AbstractGangTask {
  59     G1CollectedHeap* _g1h;
  60     uint* _dirty_region_list;
  61     size_t _num_dirty_regions;
  62     size_t _chunk_length;
  63 
  64     size_t volatile _cur_dirty_regions;
  65   public:
  66     G1ClearCardTableTask(G1CollectedHeap* g1h,
  67                          uint* dirty_region_list,
  68                          size_t num_dirty_regions,
  69                          size_t chunk_length) :
  70       AbstractGangTask("G1 Clear Card Table Task"),
  71       _g1h(g1h),
  72       _dirty_region_list(dirty_region_list),
  73       _num_dirty_regions(num_dirty_regions),
  74       _chunk_length(chunk_length),
  75       _cur_dirty_regions(0) {
  76 
  77       assert(chunk_length > 0, "must be");
  78     }
  79 
  80     static size_t chunk_size() { return M; }
  81 
  82     void work(uint worker_id) {
  83       while (_cur_dirty_regions < _num_dirty_regions) {
  84         size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length;
  85         size_t max = MIN2(next + _chunk_length, _num_dirty_regions);
  86 
  87         for (size_t i = next; i < max; i++) {
  88           HeapRegion* r = _g1h->region_at(_dirty_region_list[i]);
  89           if (!r->is_survivor()) {
  90             r->clear_cardtable();
  91           }
  92         }
  93       }
  94     }
  95   };
  96 
  97   size_t _max_regions;
  98 
  99   // Scan progress for the remembered set of a single region. Transitions from
 100   // Unclaimed -> Claimed -> Complete.
 101   // At each of the transitions the thread that does the transition needs to perform
 102   // some special action once. This is the reason for the extra "Claimed" state.
 103   typedef jint G1RemsetIterState;
 104 
 105   static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet.
 106   static const G1RemsetIterState Claimed = 1;   // The remembered set is currently being scanned.
 107   static const G1RemsetIterState Complete = 2;  // The remembered set has been completely scanned.
 108 
 109   G1RemsetIterState volatile* _iter_states;
 110   // The current location where the next thread should continue scanning in a region's
 111   // remembered set.
 112   size_t volatile* _iter_claims;
 113 
 114   // Temporary buffer holding the regions we used to store remembered set scan duplicate
 115   // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region)
 116   uint* _dirty_region_buffer;
 117 
 118   // Flag for every region whether it is in the _dirty_region_buffer already
 119   // to avoid duplicates.
 120   bool volatile* _in_dirty_region_buffer;
 121   size_t _cur_dirty_region;
 122 
 123   // Creates a snapshot of the current _top values at the start of collection to
 124   // filter out card marks that we do not want to scan.
 125   class G1ResetScanTopClosure : public HeapRegionClosure {
 126   private:
 127     HeapWord** _scan_top;
 128   public:
 129     G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { }
 130 
 131     virtual bool do_heap_region(HeapRegion* r) {
 132       uint hrm_index = r->hrm_index();
 133       if (!r->in_collection_set() && r->is_old_or_humongous_or_archive() && !r->is_empty()) {
 134         _scan_top[hrm_index] = r->top();
 135       } else {
 136         _scan_top[hrm_index] = NULL;
 137       }
 138       return false;
 139     }
 140   };
 141 
 142   // For each region, contains the maximum top() value to be used during this garbage
 143   // collection. Subsumes common checks like filtering out everything but old and
 144   // humongous regions outside the collection set.
 145   // This is valid because we are not interested in scanning stray remembered set
 146   // entries from free or archive regions.
 147   HeapWord** _scan_top;
 148 public:
 149   G1RemSetScanState() :
 150     _max_regions(0),
 151     _iter_states(NULL),
 152     _iter_claims(NULL),
 153     _dirty_region_buffer(NULL),
 154     _in_dirty_region_buffer(NULL),
 155     _cur_dirty_region(0),
 156     _scan_top(NULL) {
 157   }
 158 
 159   ~G1RemSetScanState() {
 160     if (_iter_states != NULL) {
 161       FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states);
 162     }
 163     if (_iter_claims != NULL) {
 164       FREE_C_HEAP_ARRAY(size_t, _iter_claims);
 165     }
 166     if (_dirty_region_buffer != NULL) {
 167       FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer);
 168     }
 169     if (_in_dirty_region_buffer != NULL) {
 170       FREE_C_HEAP_ARRAY(bool, _in_dirty_region_buffer);
 171     }
 172     if (_scan_top != NULL) {
 173       FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
 174     }
 175   }
 176 
 177   void initialize(uint max_regions) {
 178     assert(_iter_states == NULL, "Must not be initialized twice");
 179     assert(_iter_claims == NULL, "Must not be initialized twice");
 180     _max_regions = max_regions;
 181     _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC);
 182     _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
 183     _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC);
 184     _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(bool, max_regions, mtGC);
 185     _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC);
 186   }
 187 
 188   void reset() {
 189     for (uint i = 0; i < _max_regions; i++) {
 190       _iter_states[i] = Unclaimed;
 191       _scan_top[i] = NULL;
 192     }
 193 
 194     G1ResetScanTopClosure cl(_scan_top);
 195     G1CollectedHeap::heap()->heap_region_iterate(&cl);
 196 
 197     memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t));
 198     memset((void*)_in_dirty_region_buffer, false, _max_regions * sizeof(bool));
 199     _cur_dirty_region = 0;
 200   }
 201 
 202   // Attempt to claim the remembered set of the region for iteration. Returns true
 203   // if this call caused the transition from Unclaimed to Claimed.
 204   inline bool claim_iter(uint region) {
 205     assert(region < _max_regions, "Tried to access invalid region %u", region);
 206     if (_iter_states[region] != Unclaimed) {
 207       return false;
 208     }
 209     G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed);
 210     return (res == Unclaimed);
 211   }
 212 
 213   // Try to atomically sets the iteration state to "complete". Returns true for the
 214   // thread that caused the transition.
 215   inline bool set_iter_complete(uint region) {
 216     if (iter_is_complete(region)) {
 217       return false;
 218     }
 219     G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed);
 220     return (res == Claimed);
 221   }
 222 
 223   // Returns true if the region's iteration is complete.
 224   inline bool iter_is_complete(uint region) const {
 225     assert(region < _max_regions, "Tried to access invalid region %u", region);
 226     return _iter_states[region] == Complete;
 227   }
 228 
 229   // The current position within the remembered set of the given region.
 230   inline size_t iter_claimed(uint region) const {
 231     assert(region < _max_regions, "Tried to access invalid region %u", region);
 232     return _iter_claims[region];
 233   }
 234 
 235   // Claim the next block of cards within the remembered set of the region with
 236   // step size.
 237   inline size_t iter_claimed_next(uint region, size_t step) {
 238     return Atomic::add(step, &_iter_claims[region]) - step;
 239   }
 240 
 241   void add_dirty_region(uint region) {
 242     if (_in_dirty_region_buffer[region]) {
 243       return;
 244     }
 245 
 246     if (!Atomic::cmpxchg(true, &_in_dirty_region_buffer[region], false)) {
 247       size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1;
 248       _dirty_region_buffer[allocated] = region;
 249     }
 250   }
 251 
 252   HeapWord* scan_top(uint region_idx) const {
 253     return _scan_top[region_idx];
 254   }
 255 
 256   // Clear the card table of "dirty" regions.
 257   void clear_card_table(WorkGang* workers) {
 258     if (_cur_dirty_region == 0) {
 259       return;
 260     }
 261 
 262     size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size();
 263     uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 264     size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion;
 265 
 266     // Iterate over the dirty cards region list.
 267     G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length);
 268 
 269     log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " "
 270                         "units of work for " SIZE_FORMAT " regions.",
 271                         cl.name(), num_workers, num_chunks, _cur_dirty_region);
 272     workers->run_task(&cl, num_workers);
 273 
 274 #ifndef PRODUCT
 275     G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup();
 276 #endif
 277   }
 278 };
 279 
 280 G1RemSet::G1RemSet(G1CollectedHeap* g1h,
 281                    G1CardTable* ct,
 282                    G1HotCardCache* hot_card_cache) :
 283   _scan_state(new G1RemSetScanState()),
 284   _prev_period_summary(),
 285   _g1h(g1h),
 286   _num_conc_refined_cards(0),
 287   _ct(ct),
 288   _g1p(_g1h->policy()),
 289   _hot_card_cache(hot_card_cache) {
 290 }
 291 
 292 G1RemSet::~G1RemSet() {
 293   if (_scan_state != NULL) {
 294     delete _scan_state;
 295   }
 296 }
 297 
 298 uint G1RemSet::num_par_rem_sets() {
 299   return G1DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads);
 300 }
 301 
 302 void G1RemSet::initialize(size_t capacity, uint max_regions) {
 303   G1FromCardCache::initialize(num_par_rem_sets(), max_regions);
 304   _scan_state->initialize(max_regions);
 305 }
 306 
 307 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state,
 308                                                    G1ScanObjsDuringScanRSClosure* scan_obj_on_card,
 309                                                    G1ParScanThreadState* pss,
 310                                                    G1GCPhaseTimes::GCParPhases phase,
 311                                                    uint worker_i) :
 312   _g1h(G1CollectedHeap::heap()),
 313   _ct(_g1h->card_table()),
 314   _pss(pss),
 315   _scan_objs_on_card_cl(scan_obj_on_card),
 316   _scan_state(scan_state),
 317   _phase(phase),
 318   _worker_i(worker_i),
 319   _opt_refs_scanned(0),
 320   _opt_refs_memory_used(0),
 321   _cards_scanned(0),
 322   _cards_claimed(0),
 323   _cards_skipped(0),
 324   _rem_set_root_scan_time(),
 325   _rem_set_trim_partially_time(),
 326   _strong_code_root_scan_time(),
 327   _strong_code_trim_partially_time() {
 328 }
 329 
 330 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
 331   _ct->set_card_claimed(card_index);
 332   _scan_state->add_dirty_region(region_idx_for_card);
 333 }
 334 
 335 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
 336   HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
 337   assert(!card_region->is_young(), "Should not scan card in young region %u", region_idx_for_card);
 338   card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
 339   _scan_objs_on_card_cl->trim_queue_partially();
 340   _cards_scanned++;
 341 }
 342 
 343 void G1ScanRSForRegionClosure::scan_opt_rem_set_roots(HeapRegion* r) {
 344   EventGCPhaseParallel event;
 345 
 346   G1OopStarChunkedList* opt_rem_set_list = _pss->oops_into_optional_region(r);
 347 
 348   G1ScanObjsDuringScanRSClosure scan_cl(_g1h, _pss);
 349   G1ScanRSForOptionalClosure cl(&scan_cl);
 350   _opt_refs_scanned += opt_rem_set_list->oops_do(&cl, _pss->closures()->raw_strong_oops());
 351   _opt_refs_memory_used += opt_rem_set_list->used_memory();
 352 
 353   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
 354 }
 355 
 356 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
 357   EventGCPhaseParallel event;
 358   uint const region_idx = r->hrm_index();
 359 
 360   if (_scan_state->claim_iter(region_idx)) {
 361     // If we ever free the collection set concurrently, we should also
 362     // clear the card table concurrently therefore we won't need to
 363     // add regions of the collection set to the dirty cards region.
 364     _scan_state->add_dirty_region(region_idx);
 365   }
 366 
 367   if (r->rem_set()->cardset_is_empty()) {
 368     return;
 369   }
 370 
 371   // We claim cards in blocks so as to reduce the contention.
 372   size_t const block_size = G1RSetScanBlockSize;
 373 
 374   HeapRegionRemSetIterator iter(r->rem_set());
 375   size_t card_index;
 376 
 377   size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
 378   for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
 379     if (current_card >= claimed_card_block + block_size) {
 380       claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
 381     }
 382     if (current_card < claimed_card_block) {
 383       _cards_skipped++;
 384       continue;
 385     }
 386     _cards_claimed++;
 387 
 388     HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
 389     uint const region_idx_for_card = _g1h->addr_to_region(card_start);
 390 
 391 #ifdef ASSERT
 392     HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
 393     assert(hr == NULL || hr->is_in_reserved(card_start),
 394            "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
 395 #endif
 396     HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
 397     if (card_start >= top) {
 398       continue;
 399     }
 400 
 401     // If the card is dirty, then G1 will scan it during Update RS.
 402     if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
 403       continue;
 404     }
 405 
 406     // We claim lazily (so races are possible but they're benign), which reduces the
 407     // number of duplicate scans (the rsets of the regions in the cset can intersect).
 408     // Claim the card after checking bounds above: the remembered set may contain
 409     // random cards into current survivor, and we would then have an incorrectly
 410     // claimed card in survivor space. Card table clear does not reset the card table
 411     // of survivor space regions.
 412     claim_card(card_index, region_idx_for_card);
 413 
 414     MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
 415 
 416     scan_card(mr, region_idx_for_card);
 417   }
 418   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
 419 }
 420 
 421 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
 422   EventGCPhaseParallel event;
 423   // We pass a weak code blobs closure to the remembered set scanning because we want to avoid
 424   // treating the nmethods visited to act as roots for concurrent marking.
 425   // We only want to make sure that the oops in the nmethods are adjusted with regard to the
 426   // objects copied by the current evacuation.
 427   r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
 428   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
 429 }
 430 
 431 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
 432   assert(r->in_collection_set(), "Region %u is not in the collection set.", r->hrm_index());
 433   uint const region_idx = r->hrm_index();
 434 
 435   // The individual references for the optional remembered set are per-worker, so we
 436   // always need to scan them.
 437   if (r->has_index_in_opt_cset()) {
 438     G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
 439     scan_opt_rem_set_roots(r);
 440   }
 441 
 442   // Do an early out if we know we are complete.
 443   if (_scan_state->iter_is_complete(region_idx)) {
 444     return false;
 445   }
 446 
 447   {
 448     G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
 449     scan_rem_set_roots(r);
 450   }
 451 
 452   if (_scan_state->set_iter_complete(region_idx)) {
 453     G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
 454     // Scan the strong code root list attached to the current region
 455     scan_strong_code_roots(r);
 456   }
 457   return false;
 458 }
 459 
 460 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss,
 461                             uint worker_i,
 462                             G1GCPhaseTimes::GCParPhases scan_phase,
 463                             G1GCPhaseTimes::GCParPhases objcopy_phase,
 464                             G1GCPhaseTimes::GCParPhases coderoots_phase) {
 465   assert(pss->trim_ticks().value() == 0, "Queues must have been trimmed before entering.");
 466 
 467   G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
 468   G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, scan_phase, worker_i);
 469   _g1h->collection_set_iterate_increment_from(&cl, worker_i);
 470 
 471   G1GCPhaseTimes* p = _g1p->phase_times();
 472 
 473   p->record_or_add_time_secs(objcopy_phase, worker_i, cl.rem_set_trim_partially_time().seconds());
 474 
 475   p->record_or_add_time_secs(scan_phase, worker_i, cl.rem_set_root_scan_time().seconds());
 476   p->record_or_add_thread_work_item(scan_phase, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
 477   p->record_or_add_thread_work_item(scan_phase, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
 478   p->record_or_add_thread_work_item(scan_phase, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
 479   // At this time we only record some metrics for the optional remembered set.
 480   if (scan_phase == G1GCPhaseTimes::OptScanRS) {
 481     p->record_or_add_thread_work_item(scan_phase, worker_i, cl.opt_refs_scanned(), G1GCPhaseTimes::ScanRSScannedOptRefs);
 482     p->record_or_add_thread_work_item(scan_phase, worker_i, cl.opt_refs_memory_used(), G1GCPhaseTimes::ScanRSUsedMemory);
 483   }
 484 
 485   p->record_or_add_time_secs(coderoots_phase, worker_i, cl.strong_code_root_scan_time().seconds());
 486   p->add_time_secs(objcopy_phase, worker_i, cl.strong_code_root_trim_partially_time().seconds());
 487 }
 488 
 489 // Closure used for updating rem sets. Only called during an evacuation pause.
 490 class G1RefineCardClosure: public G1CardTableEntryClosure {
 491   G1RemSet* _g1rs;
 492   G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
 493 
 494   size_t _cards_scanned;
 495   size_t _cards_skipped;
 496 public:
 497   G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
 498     _g1rs(g1h->rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
 499   {}
 500 
 501   bool do_card_ptr(CardValue* card_ptr, uint worker_i) {
 502     // The only time we care about recording cards that
 503     // contain references that point into the collection set
 504     // is during RSet updating within an evacuation pause.
 505     // In this case worker_i should be the id of a GC worker thread.
 506     assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
 507 
 508     bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
 509 
 510     if (card_scanned) {
 511       _update_rs_cl->trim_queue_partially();
 512       _cards_scanned++;
 513     } else {
 514       _cards_skipped++;
 515     }
 516     return true;
 517   }
 518 
 519   size_t cards_scanned() const { return _cards_scanned; }
 520   size_t cards_skipped() const { return _cards_skipped; }
 521 };
 522 
 523 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
 524   G1GCPhaseTimes* p = _g1p->phase_times();
 525 
 526   // Apply closure to log entries in the HCC.
 527   if (G1HotCardCache::default_use_cache()) {
 528     G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
 529 
 530     G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss);
 531     G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
 532     _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
 533   }
 534 
 535   // Now apply the closure to all remaining log entries.
 536   {
 537     G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
 538 
 539     G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss);
 540     G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
 541     _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
 542 
 543     p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
 544     p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
 545   }
 546 }
 547 
 548 void G1RemSet::prepare_for_oops_into_collection_set_do() {
 549   G1BarrierSet::dirty_card_queue_set().concatenate_logs();
 550   _scan_state->reset();
 551 }
 552 
 553 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
 554   G1GCPhaseTimes* phase_times = _g1h->phase_times();
 555 
 556   // Set all cards back to clean.
 557   double start = os::elapsedTime();
 558   _scan_state->clear_card_table(_g1h->workers());
 559   phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
 560 }
 561 
 562 inline void check_card_ptr(CardTable::CardValue* card_ptr, G1CardTable* ct) {
 563 #ifdef ASSERT
 564   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 565   assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
 566          "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
 567          p2i(card_ptr),
 568          ct->index_for(ct->addr_for(card_ptr)),
 569          p2i(ct->addr_for(card_ptr)),
 570          g1h->addr_to_region(ct->addr_for(card_ptr)));
 571 #endif
 572 }
 573 
 574 void G1RemSet::refine_card_concurrently(CardValue* card_ptr,
 575                                         uint worker_i) {
 576   assert(!_g1h->is_gc_active(), "Only call concurrently");
 577 
 578   // Construct the region representing the card.
 579   HeapWord* start = _ct->addr_for(card_ptr);
 580   // And find the region containing it.
 581   HeapRegion* r = _g1h->heap_region_containing_or_null(start);
 582 
 583   // If this is a (stale) card into an uncommitted region, exit.
 584   if (r == NULL) {
 585     return;
 586   }
 587 
 588   check_card_ptr(card_ptr, _ct);
 589 
 590   // If the card is no longer dirty, nothing to do.
 591   if (*card_ptr != G1CardTable::dirty_card_val()) {
 592     return;
 593   }
 594 
 595   // This check is needed for some uncommon cases where we should
 596   // ignore the card.
 597   //
 598   // The region could be young.  Cards for young regions are
 599   // distinctly marked (set to g1_young_gen), so the post-barrier will
 600   // filter them out.  However, that marking is performed
 601   // concurrently.  A write to a young object could occur before the
 602   // card has been marked young, slipping past the filter.
 603   //
 604   // The card could be stale, because the region has been freed since
 605   // the card was recorded. In this case the region type could be
 606   // anything.  If (still) free or (reallocated) young, just ignore
 607   // it.  If (reallocated) old or humongous, the later card trimming
 608   // and additional checks in iteration may detect staleness.  At
 609   // worst, we end up processing a stale card unnecessarily.
 610   //
 611   // In the normal (non-stale) case, the synchronization between the
 612   // enqueueing of the card and processing it here will have ensured
 613   // we see the up-to-date region type here.
 614   if (!r->is_old_or_humongous_or_archive()) {
 615     return;
 616   }
 617 
 618   // The result from the hot card cache insert call is either:
 619   //   * pointer to the current card
 620   //     (implying that the current card is not 'hot'),
 621   //   * null
 622   //     (meaning we had inserted the card ptr into the "hot" card cache,
 623   //     which had some headroom),
 624   //   * a pointer to a "hot" card that was evicted from the "hot" cache.
 625   //
 626 
 627   if (_hot_card_cache->use_cache()) {
 628     assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
 629 
 630     const CardValue* orig_card_ptr = card_ptr;
 631     card_ptr = _hot_card_cache->insert(card_ptr);
 632     if (card_ptr == NULL) {
 633       // There was no eviction. Nothing to do.
 634       return;
 635     } else if (card_ptr != orig_card_ptr) {
 636       // Original card was inserted and an old card was evicted.
 637       start = _ct->addr_for(card_ptr);
 638       r = _g1h->heap_region_containing(start);
 639 
 640       // Check whether the region formerly in the cache should be
 641       // ignored, as discussed earlier for the original card.  The
 642       // region could have been freed while in the cache.
 643       if (!r->is_old_or_humongous_or_archive()) {
 644         return;
 645       }
 646     } // Else we still have the original card.
 647   }
 648 
 649   // Trim the region designated by the card to what's been allocated
 650   // in the region.  The card could be stale, or the card could cover
 651   // (part of) an object at the end of the allocated space and extend
 652   // beyond the end of allocation.
 653 
 654   // Non-humongous objects are only allocated in the old-gen during
 655   // GC, so if region is old then top is stable.  Humongous object
 656   // allocation sets top last; if top has not yet been set, this is
 657   // a stale card and we'll end up with an empty intersection.  If
 658   // this is not a stale card, the synchronization between the
 659   // enqueuing of the card and processing it here will have ensured
 660   // we see the up-to-date top here.
 661   HeapWord* scan_limit = r->top();
 662 
 663   if (scan_limit <= start) {
 664     // If the trimmed region is empty, the card must be stale.
 665     return;
 666   }
 667 
 668   // Okay to clean and process the card now.  There are still some
 669   // stale card cases that may be detected by iteration and dealt with
 670   // as iteration failure.
 671   *const_cast<volatile CardValue*>(card_ptr) = G1CardTable::clean_card_val();
 672 
 673   // This fence serves two purposes.  First, the card must be cleaned
 674   // before processing the contents.  Second, we can't proceed with
 675   // processing until after the read of top, for synchronization with
 676   // possibly concurrent humongous object allocation.  It's okay that
 677   // reading top and reading type were racy wrto each other.  We need
 678   // both set, in any order, to proceed.
 679   OrderAccess::fence();
 680 
 681   // Don't use addr_for(card_ptr + 1) which can ask for
 682   // a card beyond the heap.
 683   HeapWord* end = start + G1CardTable::card_size_in_words;
 684   MemRegion dirty_region(start, MIN2(scan_limit, end));
 685   assert(!dirty_region.is_empty(), "sanity");
 686 
 687   G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
 688   if (r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl)) {
 689     _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
 690     return;
 691   }
 692 
 693   // If unable to process the card then we encountered an unparsable
 694   // part of the heap (e.g. a partially allocated object, so only
 695   // temporarily a problem) while processing a stale card.  Despite
 696   // the card being stale, we can't simply ignore it, because we've
 697   // already marked the card cleaned, so taken responsibility for
 698   // ensuring the card gets scanned.
 699   //
 700   // However, the card might have gotten re-dirtied and re-enqueued
 701   // while we worked.  (In fact, it's pretty likely.)
 702   if (*card_ptr == G1CardTable::dirty_card_val()) {
 703     return;
 704   }
 705 
 706   // Re-dirty the card and enqueue in the *shared* queue.  Can't use
 707   // the thread-local queue, because that might be the queue that is
 708   // being processed by us; we could be a Java thread conscripted to
 709   // perform refinement on our queue's current buffer.
 710   *card_ptr = G1CardTable::dirty_card_val();
 711   G1BarrierSet::shared_dirty_card_queue().enqueue(card_ptr);
 712 }
 713 
 714 bool G1RemSet::refine_card_during_gc(CardValue* card_ptr,
 715                                      G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
 716   assert(_g1h->is_gc_active(), "Only call during GC");
 717 
 718   // Construct the region representing the card.
 719   HeapWord* card_start = _ct->addr_for(card_ptr);
 720   // And find the region containing it.
 721   uint const card_region_idx = _g1h->addr_to_region(card_start);
 722 
 723   HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
 724   if (scan_limit == NULL) {
 725     // This is a card into an uncommitted region. We need to bail out early as we
 726     // should not access the corresponding card table entry.
 727     return false;
 728   }
 729 
 730   check_card_ptr(card_ptr, _ct);
 731 
 732   // If the card is no longer dirty, nothing to do. This covers cards that were already
 733   // scanned as parts of the remembered sets.
 734   if (*card_ptr != G1CardTable::dirty_card_val()) {
 735     return false;
 736   }
 737 
 738   // We claim lazily (so races are possible but they're benign), which reduces the
 739   // number of potential duplicate scans (multiple threads may enqueue the same card twice).
 740   *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
 741 
 742   _scan_state->add_dirty_region(card_region_idx);
 743   if (scan_limit <= card_start) {
 744     // If the card starts above the area in the region containing objects to scan, skip it.
 745     return false;
 746   }
 747 
 748   // Don't use addr_for(card_ptr + 1) which can ask for
 749   // a card beyond the heap.
 750   HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
 751   MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
 752   assert(!dirty_region.is_empty(), "sanity");
 753 
 754   HeapRegion* const card_region = _g1h->region_at(card_region_idx);
 755   assert(!card_region->is_young(), "Should not scan card in young region %u", card_region_idx);
 756   bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
 757   assert(card_processed, "must be");
 758   return true;
 759 }
 760 
 761 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
 762   if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
 763       (period_count % G1SummarizeRSetStatsPeriod == 0)) {
 764 
 765     G1RemSetSummary current(this);
 766     _prev_period_summary.subtract_from(&current);
 767 
 768     Log(gc, remset) log;
 769     log.trace("%s", header);
 770     ResourceMark rm;
 771     LogStream ls(log.trace());
 772     _prev_period_summary.print_on(&ls);
 773 
 774     _prev_period_summary.set(&current);
 775   }
 776 }
 777 
 778 void G1RemSet::print_summary_info() {
 779   Log(gc, remset, exit) log;
 780   if (log.is_trace()) {
 781     log.trace(" Cumulative RS summary");
 782     G1RemSetSummary current(this);
 783     ResourceMark rm;
 784     LogStream ls(log.trace());
 785     current.print_on(&ls);
 786   }
 787 }
 788 
 789 class G1RebuildRemSetTask: public AbstractGangTask {
 790   // Aggregate the counting data that was constructed concurrently
 791   // with marking.
 792   class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
 793     G1ConcurrentMark* _cm;
 794     G1RebuildRemSetClosure _update_cl;
 795 
 796     // Applies _update_cl to the references of the given object, limiting objArrays
 797     // to the given MemRegion. Returns the amount of words actually scanned.
 798     size_t scan_for_references(oop const obj, MemRegion mr) {
 799       size_t const obj_size = obj->size();
 800       // All non-objArrays and objArrays completely within the mr
 801       // can be scanned without passing the mr.
 802       if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
 803         obj->oop_iterate(&_update_cl);
 804         return obj_size;
 805       }
 806       // This path is for objArrays crossing the given MemRegion. Only scan the
 807       // area within the MemRegion.
 808       obj->oop_iterate(&_update_cl, mr);
 809       return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
 810     }
 811 
 812     // A humongous object is live (with respect to the scanning) either
 813     // a) it is marked on the bitmap as such
 814     // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
 815     bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
 816       return bitmap->is_marked(humongous_obj) || (tars > tams);
 817     }
 818 
 819     // Iterator over the live objects within the given MemRegion.
 820     class LiveObjIterator : public StackObj {
 821       const G1CMBitMap* const _bitmap;
 822       const HeapWord* _tams;
 823       const MemRegion _mr;
 824       HeapWord* _current;
 825 
 826       bool is_below_tams() const {
 827         return _current < _tams;
 828       }
 829 
 830       bool is_live(HeapWord* obj) const {
 831         return !is_below_tams() || _bitmap->is_marked(obj);
 832       }
 833 
 834       HeapWord* bitmap_limit() const {
 835         return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
 836       }
 837 
 838       void move_if_below_tams() {
 839         if (is_below_tams() && has_next()) {
 840           _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
 841         }
 842       }
 843     public:
 844       LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
 845           _bitmap(bitmap),
 846           _tams(tams),
 847           _mr(mr),
 848           _current(first_oop_into_mr) {
 849 
 850         assert(_current <= _mr.start(),
 851                "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
 852                p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
 853 
 854         // Step to the next live object within the MemRegion if needed.
 855         if (is_live(_current)) {
 856           // Non-objArrays were scanned by the previous part of that region.
 857           if (_current < mr.start() && !oop(_current)->is_objArray()) {
 858             _current += oop(_current)->size();
 859             // We might have positioned _current on a non-live object. Reposition to the next
 860             // live one if needed.
 861             move_if_below_tams();
 862           }
 863         } else {
 864           // The object at _current can only be dead if below TAMS, so we can use the bitmap.
 865           // immediately.
 866           _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
 867           assert(_current == _mr.end() || is_live(_current),
 868                  "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
 869                  p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
 870         }
 871       }
 872 
 873       void move_to_next() {
 874         _current += next()->size();
 875         move_if_below_tams();
 876       }
 877 
 878       oop next() const {
 879         oop result = oop(_current);
 880         assert(is_live(_current),
 881                "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
 882                p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
 883         return result;
 884       }
 885 
 886       bool has_next() const {
 887         return _current < _mr.end();
 888       }
 889     };
 890 
 891     // Rebuild remembered sets in the part of the region specified by mr and hr.
 892     // Objects between the bottom of the region and the TAMS are checked for liveness
 893     // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
 894     // Returns the number of live words between bottom and TAMS.
 895     size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
 896                                      HeapWord* const top_at_mark_start,
 897                                      HeapWord* const top_at_rebuild_start,
 898                                      HeapRegion* hr,
 899                                      MemRegion mr) {
 900       size_t marked_words = 0;
 901 
 902       if (hr->is_humongous()) {
 903         oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
 904         if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
 905           // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
 906           // however in case of humongous objects it is sufficient to scan the encompassing
 907           // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
 908           // two areas will be zero sized. I.e. TAMS is either
 909           // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
 910           // value: this would mean that TAMS points somewhere into the object.
 911           assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
 912                  "More than one object in the humongous region?");
 913           humongous_obj->oop_iterate(&_update_cl, mr);
 914           return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
 915         } else {
 916           return 0;
 917         }
 918       }
 919 
 920       for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
 921         oop obj = it.next();
 922         size_t scanned_size = scan_for_references(obj, mr);
 923         if ((HeapWord*)obj < top_at_mark_start) {
 924           marked_words += scanned_size;
 925         }
 926       }
 927 
 928       return marked_words * HeapWordSize;
 929     }
 930 public:
 931   G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
 932                                    G1ConcurrentMark* cm,
 933                                    uint worker_id) :
 934     HeapRegionClosure(),
 935     _cm(cm),
 936     _update_cl(g1h, worker_id) { }
 937 
 938     bool do_heap_region(HeapRegion* hr) {
 939       if (_cm->has_aborted()) {
 940         return true;
 941       }
 942 
 943       uint const region_idx = hr->hrm_index();
 944       DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
 945       assert(top_at_rebuild_start_check == NULL ||
 946              top_at_rebuild_start_check > hr->bottom(),
 947              "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
 948              p2i(top_at_rebuild_start_check), p2i(hr->bottom()),  region_idx, hr->get_type_str());
 949 
 950       size_t total_marked_bytes = 0;
 951       size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
 952 
 953       HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
 954 
 955       HeapWord* cur = hr->bottom();
 956       while (cur < hr->end()) {
 957         // After every iteration (yield point) we need to check whether the region's
 958         // TARS changed due to e.g. eager reclaim.
 959         HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
 960         if (top_at_rebuild_start == NULL) {
 961           return false;
 962         }
 963 
 964         MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
 965         if (next_chunk.is_empty()) {
 966           break;
 967         }
 968 
 969         const Ticks start = Ticks::now();
 970         size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
 971                                                         top_at_mark_start,
 972                                                         top_at_rebuild_start,
 973                                                         hr,
 974                                                         next_chunk);
 975         Tickspan time = Ticks::now() - start;
 976 
 977         log_trace(gc, remset, tracking)("Rebuilt region %u "
 978                                         "live " SIZE_FORMAT " "
 979                                         "time %.3fms "
 980                                         "marked bytes " SIZE_FORMAT " "
 981                                         "bot " PTR_FORMAT " "
 982                                         "TAMS " PTR_FORMAT " "
 983                                         "TARS " PTR_FORMAT,
 984                                         region_idx,
 985                                         _cm->liveness(region_idx) * HeapWordSize,
 986                                         time.seconds() * 1000.0,
 987                                         marked_bytes,
 988                                         p2i(hr->bottom()),
 989                                         p2i(top_at_mark_start),
 990                                         p2i(top_at_rebuild_start));
 991 
 992         if (marked_bytes > 0) {
 993           total_marked_bytes += marked_bytes;
 994         }
 995         cur += chunk_size_in_words;
 996 
 997         _cm->do_yield_check();
 998         if (_cm->has_aborted()) {
 999           return true;
1000         }
1001       }
1002       // In the final iteration of the loop the region might have been eagerly reclaimed.
1003       // Simply filter out those regions. We can not just use region type because there
1004       // might have already been new allocations into these regions.
1005       DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
1006       assert(top_at_rebuild_start == NULL ||
1007              total_marked_bytes == hr->marked_bytes(),
1008              "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
1009              "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
1010              total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
1011              p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
1012        // Abort state may have changed after the yield check.
1013       return _cm->has_aborted();
1014     }
1015   };
1016 
1017   HeapRegionClaimer _hr_claimer;
1018   G1ConcurrentMark* _cm;
1019 
1020   uint _worker_id_offset;
1021 public:
1022   G1RebuildRemSetTask(G1ConcurrentMark* cm,
1023                       uint n_workers,
1024                       uint worker_id_offset) :
1025       AbstractGangTask("G1 Rebuild Remembered Set"),
1026       _hr_claimer(n_workers),
1027       _cm(cm),
1028       _worker_id_offset(worker_id_offset) {
1029   }
1030 
1031   void work(uint worker_id) {
1032     SuspendibleThreadSetJoiner sts_join;
1033 
1034     G1CollectedHeap* g1h = G1CollectedHeap::heap();
1035 
1036     G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1037     g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1038   }
1039 };
1040 
1041 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1042                                WorkGang* workers,
1043                                uint worker_id_offset) {
1044   uint num_workers = workers->active_workers();
1045 
1046   G1RebuildRemSetTask cl(cm,
1047                          num_workers,
1048                          worker_id_offset);
1049   workers->run_task(&cl, num_workers);
1050 }