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   _cards_scanned(0),
 320   _cards_claimed(0),
 321   _cards_skipped(0),
 322   _rem_set_root_scan_time(),
 323   _rem_set_trim_partially_time(),
 324   _strong_code_root_scan_time(),
 325   _strong_code_trim_partially_time() {
 326 }
 327 
 328 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){
 329   _ct->set_card_claimed(card_index);
 330   _scan_state->add_dirty_region(region_idx_for_card);
 331 }
 332 
 333 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) {
 334   HeapRegion* const card_region = _g1h->region_at(region_idx_for_card);
 335   assert(!card_region->is_young(), "Should not scan card in young region %u", region_idx_for_card);
 336   card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl);
 337   _scan_objs_on_card_cl->trim_queue_partially();
 338   _cards_scanned++;
 339 }
 340 
 341 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) {
 342   EventGCPhaseParallel event;
 343   uint const region_idx = r->hrm_index();
 344 
 345   if (_scan_state->claim_iter(region_idx)) {
 346     // If we ever free the collection set concurrently, we should also
 347     // clear the card table concurrently therefore we won't need to
 348     // add regions of the collection set to the dirty cards region.
 349     _scan_state->add_dirty_region(region_idx);
 350   }
 351 
 352   if (r->rem_set()->cardset_is_empty()) {
 353     return;
 354   }
 355 
 356   // We claim cards in blocks so as to reduce the contention.
 357   size_t const block_size = G1RSetScanBlockSize;
 358 
 359   HeapRegionRemSetIterator iter(r->rem_set());
 360   size_t card_index;
 361 
 362   size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
 363   for (size_t current_card = 0; iter.has_next(card_index); current_card++) {
 364     if (current_card >= claimed_card_block + block_size) {
 365       claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size);
 366     }
 367     if (current_card < claimed_card_block) {
 368       _cards_skipped++;
 369       continue;
 370     }
 371     _cards_claimed++;
 372 
 373     HeapWord* const card_start = _g1h->bot()->address_for_index_raw(card_index);
 374     uint const region_idx_for_card = _g1h->addr_to_region(card_start);
 375 
 376 #ifdef ASSERT
 377     HeapRegion* hr = _g1h->region_at_or_null(region_idx_for_card);
 378     assert(hr == NULL || hr->is_in_reserved(card_start),
 379            "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index());
 380 #endif
 381     HeapWord* const top = _scan_state->scan_top(region_idx_for_card);
 382     if (card_start >= top) {
 383       continue;
 384     }
 385 
 386     // If the card is dirty, then G1 will scan it during Update RS.
 387     if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) {
 388       continue;
 389     }
 390 
 391     // We claim lazily (so races are possible but they're benign), which reduces the
 392     // number of duplicate scans (the rsets of the regions in the cset can intersect).
 393     // Claim the card after checking bounds above: the remembered set may contain
 394     // random cards into current survivor, and we would then have an incorrectly
 395     // claimed card in survivor space. Card table clear does not reset the card table
 396     // of survivor space regions.
 397     claim_card(card_index, region_idx_for_card);
 398 
 399     MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top));
 400 
 401     scan_card(mr, region_idx_for_card);
 402   }
 403   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(_phase));
 404 }
 405 
 406 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) {
 407   EventGCPhaseParallel event;
 408   // We pass a weak code blobs closure to the remembered set scanning because we want to avoid
 409   // treating the nmethods visited to act as roots for concurrent marking.
 410   // We only want to make sure that the oops in the nmethods are adjusted with regard to the
 411   // objects copied by the current evacuation.
 412   r->strong_code_roots_do(_pss->closures()->weak_codeblobs());
 413   event.commit(GCId::current(), _worker_i, G1GCPhaseTimes::phase_name(G1GCPhaseTimes::CodeRoots));
 414 }
 415 
 416 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) {
 417   assert(r->in_collection_set(),
 418          "Should only be called on elements of the collection set but region %u is not.",
 419          r->hrm_index());
 420   uint const region_idx = r->hrm_index();
 421 
 422   // Do an early out if we know we are complete.
 423   if (_scan_state->iter_is_complete(region_idx)) {
 424     return false;
 425   }
 426 
 427   {
 428     G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time);
 429     scan_rem_set_roots(r);
 430   }
 431 
 432   if (_scan_state->set_iter_complete(region_idx)) {
 433     G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time);
 434     // Scan the strong code root list attached to the current region
 435     scan_strong_code_roots(r);
 436   }
 437   return false;
 438 }
 439 
 440 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) {
 441   G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss);
 442   G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, G1GCPhaseTimes::ScanRS, worker_i);
 443   _g1h->collection_set_iterate_from(&cl, worker_i);
 444 
 445   G1GCPhaseTimes* p = _g1p->phase_times();
 446 
 447   p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, cl.rem_set_root_scan_time().seconds());
 448   p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.rem_set_trim_partially_time().seconds());
 449 
 450   p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards);
 451   p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards);
 452   p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards);
 453 
 454   p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time().seconds());
 455   p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, cl.strong_code_root_trim_partially_time().seconds());
 456 }
 457 
 458 // Closure used for updating rem sets. Only called during an evacuation pause.
 459 class G1RefineCardClosure: public G1CardTableEntryClosure {
 460   G1RemSet* _g1rs;
 461   G1ScanObjsDuringUpdateRSClosure* _update_rs_cl;
 462 
 463   size_t _cards_scanned;
 464   size_t _cards_skipped;
 465 public:
 466   G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) :
 467     _g1rs(g1h->rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0)
 468   {}
 469 
 470   bool do_card_ptr(CardValue* card_ptr, uint worker_i) {
 471     // The only time we care about recording cards that
 472     // contain references that point into the collection set
 473     // is during RSet updating within an evacuation pause.
 474     // In this case worker_i should be the id of a GC worker thread.
 475     assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause");
 476 
 477     bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl);
 478 
 479     if (card_scanned) {
 480       _update_rs_cl->trim_queue_partially();
 481       _cards_scanned++;
 482     } else {
 483       _cards_skipped++;
 484     }
 485     return true;
 486   }
 487 
 488   size_t cards_scanned() const { return _cards_scanned; }
 489   size_t cards_skipped() const { return _cards_skipped; }
 490 };
 491 
 492 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) {
 493   G1GCPhaseTimes* p = _g1p->phase_times();
 494 
 495   // Apply closure to log entries in the HCC.
 496   if (G1HotCardCache::default_use_cache()) {
 497     G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i);
 498 
 499     G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss);
 500     G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl);
 501     _g1h->iterate_hcc_closure(&refine_card_cl, worker_i);
 502   }
 503 
 504   // Now apply the closure to all remaining log entries.
 505   {
 506     G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i);
 507 
 508     G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss);
 509     G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl);
 510     _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i);
 511 
 512     p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards);
 513     p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards);
 514   }
 515 }
 516 
 517 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) {
 518   update_rem_set(pss, worker_i);
 519   scan_rem_set(pss, worker_i);;
 520 }
 521 
 522 void G1RemSet::prepare_for_oops_into_collection_set_do() {
 523   G1BarrierSet::dirty_card_queue_set().concatenate_logs();
 524   _scan_state->reset();
 525 }
 526 
 527 void G1RemSet::cleanup_after_oops_into_collection_set_do() {
 528   G1GCPhaseTimes* phase_times = _g1h->phase_times();
 529 
 530   // Set all cards back to clean.
 531   double start = os::elapsedTime();
 532   _scan_state->clear_card_table(_g1h->workers());
 533   phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0);
 534 }
 535 
 536 inline void check_card_ptr(CardTable::CardValue* card_ptr, G1CardTable* ct) {
 537 #ifdef ASSERT
 538   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 539   assert(g1h->is_in_exact(ct->addr_for(card_ptr)),
 540          "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap",
 541          p2i(card_ptr),
 542          ct->index_for(ct->addr_for(card_ptr)),
 543          p2i(ct->addr_for(card_ptr)),
 544          g1h->addr_to_region(ct->addr_for(card_ptr)));
 545 #endif
 546 }
 547 
 548 void G1RemSet::refine_card_concurrently(CardValue* card_ptr,
 549                                         uint worker_i) {
 550   assert(!_g1h->is_gc_active(), "Only call concurrently");
 551 
 552   // Construct the region representing the card.
 553   HeapWord* start = _ct->addr_for(card_ptr);
 554   // And find the region containing it.
 555   HeapRegion* r = _g1h->heap_region_containing_or_null(start);
 556 
 557   // If this is a (stale) card into an uncommitted region, exit.
 558   if (r == NULL) {
 559     return;
 560   }
 561 
 562   check_card_ptr(card_ptr, _ct);
 563 
 564   // If the card is no longer dirty, nothing to do.
 565   if (*card_ptr != G1CardTable::dirty_card_val()) {
 566     return;
 567   }
 568 
 569   // This check is needed for some uncommon cases where we should
 570   // ignore the card.
 571   //
 572   // The region could be young.  Cards for young regions are
 573   // distinctly marked (set to g1_young_gen), so the post-barrier will
 574   // filter them out.  However, that marking is performed
 575   // concurrently.  A write to a young object could occur before the
 576   // card has been marked young, slipping past the filter.
 577   //
 578   // The card could be stale, because the region has been freed since
 579   // the card was recorded. In this case the region type could be
 580   // anything.  If (still) free or (reallocated) young, just ignore
 581   // it.  If (reallocated) old or humongous, the later card trimming
 582   // and additional checks in iteration may detect staleness.  At
 583   // worst, we end up processing a stale card unnecessarily.
 584   //
 585   // In the normal (non-stale) case, the synchronization between the
 586   // enqueueing of the card and processing it here will have ensured
 587   // we see the up-to-date region type here.
 588   if (!r->is_old_or_humongous_or_archive()) {
 589     return;
 590   }
 591 
 592   // The result from the hot card cache insert call is either:
 593   //   * pointer to the current card
 594   //     (implying that the current card is not 'hot'),
 595   //   * null
 596   //     (meaning we had inserted the card ptr into the "hot" card cache,
 597   //     which had some headroom),
 598   //   * a pointer to a "hot" card that was evicted from the "hot" cache.
 599   //
 600 
 601   if (_hot_card_cache->use_cache()) {
 602     assert(!SafepointSynchronize::is_at_safepoint(), "sanity");
 603 
 604     const CardValue* orig_card_ptr = card_ptr;
 605     card_ptr = _hot_card_cache->insert(card_ptr);
 606     if (card_ptr == NULL) {
 607       // There was no eviction. Nothing to do.
 608       return;
 609     } else if (card_ptr != orig_card_ptr) {
 610       // Original card was inserted and an old card was evicted.
 611       start = _ct->addr_for(card_ptr);
 612       r = _g1h->heap_region_containing(start);
 613 
 614       // Check whether the region formerly in the cache should be
 615       // ignored, as discussed earlier for the original card.  The
 616       // region could have been freed while in the cache.
 617       if (!r->is_old_or_humongous_or_archive()) {
 618         return;
 619       }
 620     } // Else we still have the original card.
 621   }
 622 
 623   // Trim the region designated by the card to what's been allocated
 624   // in the region.  The card could be stale, or the card could cover
 625   // (part of) an object at the end of the allocated space and extend
 626   // beyond the end of allocation.
 627 
 628   // Non-humongous objects are only allocated in the old-gen during
 629   // GC, so if region is old then top is stable.  Humongous object
 630   // allocation sets top last; if top has not yet been set, this is
 631   // a stale card and we'll end up with an empty intersection.  If
 632   // this is not a stale card, the synchronization between the
 633   // enqueuing of the card and processing it here will have ensured
 634   // we see the up-to-date top here.
 635   HeapWord* scan_limit = r->top();
 636 
 637   if (scan_limit <= start) {
 638     // If the trimmed region is empty, the card must be stale.
 639     return;
 640   }
 641 
 642   // Okay to clean and process the card now.  There are still some
 643   // stale card cases that may be detected by iteration and dealt with
 644   // as iteration failure.
 645   *const_cast<volatile CardValue*>(card_ptr) = G1CardTable::clean_card_val();
 646 
 647   // This fence serves two purposes.  First, the card must be cleaned
 648   // before processing the contents.  Second, we can't proceed with
 649   // processing until after the read of top, for synchronization with
 650   // possibly concurrent humongous object allocation.  It's okay that
 651   // reading top and reading type were racy wrto each other.  We need
 652   // both set, in any order, to proceed.
 653   OrderAccess::fence();
 654 
 655   // Don't use addr_for(card_ptr + 1) which can ask for
 656   // a card beyond the heap.
 657   HeapWord* end = start + G1CardTable::card_size_in_words;
 658   MemRegion dirty_region(start, MIN2(scan_limit, end));
 659   assert(!dirty_region.is_empty(), "sanity");
 660 
 661   G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i);
 662   if (r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl)) {
 663     _num_conc_refined_cards++; // Unsynchronized update, only used for logging.
 664     return;
 665   }
 666 
 667   // If unable to process the card then we encountered an unparsable
 668   // part of the heap (e.g. a partially allocated object, so only
 669   // temporarily a problem) while processing a stale card.  Despite
 670   // the card being stale, we can't simply ignore it, because we've
 671   // already marked the card cleaned, so taken responsibility for
 672   // ensuring the card gets scanned.
 673   //
 674   // However, the card might have gotten re-dirtied and re-enqueued
 675   // while we worked.  (In fact, it's pretty likely.)
 676   if (*card_ptr == G1CardTable::dirty_card_val()) {
 677     return;
 678   }
 679 
 680   // Re-dirty the card and enqueue in the *shared* queue.  Can't use
 681   // the thread-local queue, because that might be the queue that is
 682   // being processed by us; we could be a Java thread conscripted to
 683   // perform refinement on our queue's current buffer.
 684   *card_ptr = G1CardTable::dirty_card_val();
 685   G1BarrierSet::shared_dirty_card_queue().enqueue(card_ptr);
 686 }
 687 
 688 bool G1RemSet::refine_card_during_gc(CardValue* card_ptr,
 689                                      G1ScanObjsDuringUpdateRSClosure* update_rs_cl) {
 690   assert(_g1h->is_gc_active(), "Only call during GC");
 691 
 692   // Construct the region representing the card.
 693   HeapWord* card_start = _ct->addr_for(card_ptr);
 694   // And find the region containing it.
 695   uint const card_region_idx = _g1h->addr_to_region(card_start);
 696 
 697   HeapWord* scan_limit = _scan_state->scan_top(card_region_idx);
 698   if (scan_limit == NULL) {
 699     // This is a card into an uncommitted region. We need to bail out early as we
 700     // should not access the corresponding card table entry.
 701     return false;
 702   }
 703 
 704   check_card_ptr(card_ptr, _ct);
 705 
 706   // If the card is no longer dirty, nothing to do. This covers cards that were already
 707   // scanned as parts of the remembered sets.
 708   if (*card_ptr != G1CardTable::dirty_card_val()) {
 709     return false;
 710   }
 711 
 712   // We claim lazily (so races are possible but they're benign), which reduces the
 713   // number of potential duplicate scans (multiple threads may enqueue the same card twice).
 714   *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val();
 715 
 716   _scan_state->add_dirty_region(card_region_idx);
 717   if (scan_limit <= card_start) {
 718     // If the card starts above the area in the region containing objects to scan, skip it.
 719     return false;
 720   }
 721 
 722   // Don't use addr_for(card_ptr + 1) which can ask for
 723   // a card beyond the heap.
 724   HeapWord* card_end = card_start + G1CardTable::card_size_in_words;
 725   MemRegion dirty_region(card_start, MIN2(scan_limit, card_end));
 726   assert(!dirty_region.is_empty(), "sanity");
 727 
 728   HeapRegion* const card_region = _g1h->region_at(card_region_idx);
 729   assert(!card_region->is_young(), "Should not scan card in young region %u", card_region_idx);
 730   bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl);
 731   assert(card_processed, "must be");
 732   return true;
 733 }
 734 
 735 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) {
 736   if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) &&
 737       (period_count % G1SummarizeRSetStatsPeriod == 0)) {
 738 
 739     G1RemSetSummary current(this);
 740     _prev_period_summary.subtract_from(&current);
 741 
 742     Log(gc, remset) log;
 743     log.trace("%s", header);
 744     ResourceMark rm;
 745     LogStream ls(log.trace());
 746     _prev_period_summary.print_on(&ls);
 747 
 748     _prev_period_summary.set(&current);
 749   }
 750 }
 751 
 752 void G1RemSet::print_summary_info() {
 753   Log(gc, remset, exit) log;
 754   if (log.is_trace()) {
 755     log.trace(" Cumulative RS summary");
 756     G1RemSetSummary current(this);
 757     ResourceMark rm;
 758     LogStream ls(log.trace());
 759     current.print_on(&ls);
 760   }
 761 }
 762 
 763 class G1RebuildRemSetTask: public AbstractGangTask {
 764   // Aggregate the counting data that was constructed concurrently
 765   // with marking.
 766   class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure {
 767     G1ConcurrentMark* _cm;
 768     G1RebuildRemSetClosure _update_cl;
 769 
 770     // Applies _update_cl to the references of the given object, limiting objArrays
 771     // to the given MemRegion. Returns the amount of words actually scanned.
 772     size_t scan_for_references(oop const obj, MemRegion mr) {
 773       size_t const obj_size = obj->size();
 774       // All non-objArrays and objArrays completely within the mr
 775       // can be scanned without passing the mr.
 776       if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) {
 777         obj->oop_iterate(&_update_cl);
 778         return obj_size;
 779       }
 780       // This path is for objArrays crossing the given MemRegion. Only scan the
 781       // area within the MemRegion.
 782       obj->oop_iterate(&_update_cl, mr);
 783       return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size();
 784     }
 785 
 786     // A humongous object is live (with respect to the scanning) either
 787     // a) it is marked on the bitmap as such
 788     // b) its TARS is larger than TAMS, i.e. has been allocated during marking.
 789     bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const {
 790       return bitmap->is_marked(humongous_obj) || (tars > tams);
 791     }
 792 
 793     // Iterator over the live objects within the given MemRegion.
 794     class LiveObjIterator : public StackObj {
 795       const G1CMBitMap* const _bitmap;
 796       const HeapWord* _tams;
 797       const MemRegion _mr;
 798       HeapWord* _current;
 799 
 800       bool is_below_tams() const {
 801         return _current < _tams;
 802       }
 803 
 804       bool is_live(HeapWord* obj) const {
 805         return !is_below_tams() || _bitmap->is_marked(obj);
 806       }
 807 
 808       HeapWord* bitmap_limit() const {
 809         return MIN2(const_cast<HeapWord*>(_tams), _mr.end());
 810       }
 811 
 812       void move_if_below_tams() {
 813         if (is_below_tams() && has_next()) {
 814           _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
 815         }
 816       }
 817     public:
 818       LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) :
 819           _bitmap(bitmap),
 820           _tams(tams),
 821           _mr(mr),
 822           _current(first_oop_into_mr) {
 823 
 824         assert(_current <= _mr.start(),
 825                "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")",
 826                p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end()));
 827 
 828         // Step to the next live object within the MemRegion if needed.
 829         if (is_live(_current)) {
 830           // Non-objArrays were scanned by the previous part of that region.
 831           if (_current < mr.start() && !oop(_current)->is_objArray()) {
 832             _current += oop(_current)->size();
 833             // We might have positioned _current on a non-live object. Reposition to the next
 834             // live one if needed.
 835             move_if_below_tams();
 836           }
 837         } else {
 838           // The object at _current can only be dead if below TAMS, so we can use the bitmap.
 839           // immediately.
 840           _current = _bitmap->get_next_marked_addr(_current, bitmap_limit());
 841           assert(_current == _mr.end() || is_live(_current),
 842                  "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")",
 843                  p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end()));
 844         }
 845       }
 846 
 847       void move_to_next() {
 848         _current += next()->size();
 849         move_if_below_tams();
 850       }
 851 
 852       oop next() const {
 853         oop result = oop(_current);
 854         assert(is_live(_current),
 855                "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d",
 856                p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result));
 857         return result;
 858       }
 859 
 860       bool has_next() const {
 861         return _current < _mr.end();
 862       }
 863     };
 864 
 865     // Rebuild remembered sets in the part of the region specified by mr and hr.
 866     // Objects between the bottom of the region and the TAMS are checked for liveness
 867     // using the given bitmap. Objects between TAMS and TARS are assumed to be live.
 868     // Returns the number of live words between bottom and TAMS.
 869     size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap,
 870                                      HeapWord* const top_at_mark_start,
 871                                      HeapWord* const top_at_rebuild_start,
 872                                      HeapRegion* hr,
 873                                      MemRegion mr) {
 874       size_t marked_words = 0;
 875 
 876       if (hr->is_humongous()) {
 877         oop const humongous_obj = oop(hr->humongous_start_region()->bottom());
 878         if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) {
 879           // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start);
 880           // however in case of humongous objects it is sufficient to scan the encompassing
 881           // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the
 882           // two areas will be zero sized. I.e. TAMS is either
 883           // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different
 884           // value: this would mean that TAMS points somewhere into the object.
 885           assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start,
 886                  "More than one object in the humongous region?");
 887           humongous_obj->oop_iterate(&_update_cl, mr);
 888           return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0;
 889         } else {
 890           return 0;
 891         }
 892       }
 893 
 894       for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) {
 895         oop obj = it.next();
 896         size_t scanned_size = scan_for_references(obj, mr);
 897         if ((HeapWord*)obj < top_at_mark_start) {
 898           marked_words += scanned_size;
 899         }
 900       }
 901 
 902       return marked_words * HeapWordSize;
 903     }
 904 public:
 905   G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h,
 906                                    G1ConcurrentMark* cm,
 907                                    uint worker_id) :
 908     HeapRegionClosure(),
 909     _cm(cm),
 910     _update_cl(g1h, worker_id) { }
 911 
 912     bool do_heap_region(HeapRegion* hr) {
 913       if (_cm->has_aborted()) {
 914         return true;
 915       }
 916 
 917       uint const region_idx = hr->hrm_index();
 918       DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);)
 919       assert(top_at_rebuild_start_check == NULL ||
 920              top_at_rebuild_start_check > hr->bottom(),
 921              "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)",
 922              p2i(top_at_rebuild_start_check), p2i(hr->bottom()),  region_idx, hr->get_type_str());
 923 
 924       size_t total_marked_bytes = 0;
 925       size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize;
 926 
 927       HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start();
 928 
 929       HeapWord* cur = hr->bottom();
 930       while (cur < hr->end()) {
 931         // After every iteration (yield point) we need to check whether the region's
 932         // TARS changed due to e.g. eager reclaim.
 933         HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);
 934         if (top_at_rebuild_start == NULL) {
 935           return false;
 936         }
 937 
 938         MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words));
 939         if (next_chunk.is_empty()) {
 940           break;
 941         }
 942 
 943         const Ticks start = Ticks::now();
 944         size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(),
 945                                                         top_at_mark_start,
 946                                                         top_at_rebuild_start,
 947                                                         hr,
 948                                                         next_chunk);
 949         Tickspan time = Ticks::now() - start;
 950 
 951         log_trace(gc, remset, tracking)("Rebuilt region %u "
 952                                         "live " SIZE_FORMAT " "
 953                                         "time %.3fms "
 954                                         "marked bytes " SIZE_FORMAT " "
 955                                         "bot " PTR_FORMAT " "
 956                                         "TAMS " PTR_FORMAT " "
 957                                         "TARS " PTR_FORMAT,
 958                                         region_idx,
 959                                         _cm->liveness(region_idx) * HeapWordSize,
 960                                         time.seconds() * 1000.0,
 961                                         marked_bytes,
 962                                         p2i(hr->bottom()),
 963                                         p2i(top_at_mark_start),
 964                                         p2i(top_at_rebuild_start));
 965 
 966         if (marked_bytes > 0) {
 967           total_marked_bytes += marked_bytes;
 968         }
 969         cur += chunk_size_in_words;
 970 
 971         _cm->do_yield_check();
 972         if (_cm->has_aborted()) {
 973           return true;
 974         }
 975       }
 976       // In the final iteration of the loop the region might have been eagerly reclaimed.
 977       // Simply filter out those regions. We can not just use region type because there
 978       // might have already been new allocations into these regions.
 979       DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);)
 980       assert(top_at_rebuild_start == NULL ||
 981              total_marked_bytes == hr->marked_bytes(),
 982              "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " "
 983              "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")",
 984              total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(),
 985              p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start));
 986        // Abort state may have changed after the yield check.
 987       return _cm->has_aborted();
 988     }
 989   };
 990 
 991   HeapRegionClaimer _hr_claimer;
 992   G1ConcurrentMark* _cm;
 993 
 994   uint _worker_id_offset;
 995 public:
 996   G1RebuildRemSetTask(G1ConcurrentMark* cm,
 997                       uint n_workers,
 998                       uint worker_id_offset) :
 999       AbstractGangTask("G1 Rebuild Remembered Set"),
1000       _hr_claimer(n_workers),
1001       _cm(cm),
1002       _worker_id_offset(worker_id_offset) {
1003   }
1004 
1005   void work(uint worker_id) {
1006     SuspendibleThreadSetJoiner sts_join;
1007 
1008     G1CollectedHeap* g1h = G1CollectedHeap::heap();
1009 
1010     G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id);
1011     g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id);
1012   }
1013 };
1014 
1015 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm,
1016                                WorkGang* workers,
1017                                uint worker_id_offset) {
1018   uint num_workers = workers->active_workers();
1019 
1020   G1RebuildRemSetTask cl(cm,
1021                          num_workers,
1022                          worker_id_offset);
1023   workers->run_task(&cl, num_workers);
1024 }