1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/dirtyCardQueue.hpp" 27 #include "gc/g1/g1BarrierSet.hpp" 28 #include "gc/g1/g1BlockOffsetTable.inline.hpp" 29 #include "gc/g1/g1CardTable.inline.hpp" 30 #include "gc/g1/g1CollectedHeap.inline.hpp" 31 #include "gc/g1/g1ConcurrentRefine.hpp" 32 #include "gc/g1/g1FromCardCache.hpp" 33 #include "gc/g1/g1GCPhaseTimes.hpp" 34 #include "gc/g1/g1HotCardCache.hpp" 35 #include "gc/g1/g1OopClosures.inline.hpp" 36 #include "gc/g1/g1RootClosures.hpp" 37 #include "gc/g1/g1RemSet.hpp" 38 #include "gc/g1/heapRegion.inline.hpp" 39 #include "gc/g1/heapRegionManager.inline.hpp" 40 #include "gc/g1/heapRegionRemSet.hpp" 41 #include "gc/shared/gcTraceTime.inline.hpp" 42 #include "gc/shared/suspendibleThreadSet.hpp" 43 #include "memory/iterator.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "oops/access.inline.hpp" 46 #include "oops/oop.inline.hpp" 47 #include "utilities/align.hpp" 48 #include "utilities/globalDefinitions.hpp" 49 #include "utilities/intHisto.hpp" 50 #include "utilities/stack.inline.hpp" 51 #include "utilities/ticks.inline.hpp" 52 53 // Collects information about the overall remembered set scan progress during an evacuation. 54 class G1RemSetScanState : public CHeapObj<mtGC> { 55 private: 56 class G1ClearCardTableTask : public AbstractGangTask { 57 G1CollectedHeap* _g1h; 58 uint* _dirty_region_list; 59 size_t _num_dirty_regions; 60 size_t _chunk_length; 61 62 size_t volatile _cur_dirty_regions; 63 public: 64 G1ClearCardTableTask(G1CollectedHeap* g1h, 65 uint* dirty_region_list, 66 size_t num_dirty_regions, 67 size_t chunk_length) : 68 AbstractGangTask("G1 Clear Card Table Task"), 69 _g1h(g1h), 70 _dirty_region_list(dirty_region_list), 71 _num_dirty_regions(num_dirty_regions), 72 _chunk_length(chunk_length), 73 _cur_dirty_regions(0) { 74 75 assert(chunk_length > 0, "must be"); 76 } 77 78 static size_t chunk_size() { return M; } 79 80 void work(uint worker_id) { 81 while (_cur_dirty_regions < _num_dirty_regions) { 82 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length; 83 size_t max = MIN2(next + _chunk_length, _num_dirty_regions); 84 85 for (size_t i = next; i < max; i++) { 86 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]); 87 if (!r->is_survivor()) { 88 r->clear_cardtable(); 89 } 90 } 91 } 92 } 93 }; 94 95 size_t _max_regions; 96 97 // Scan progress for the remembered set of a single region. Transitions from 98 // Unclaimed -> Claimed -> Complete. 99 // At each of the transitions the thread that does the transition needs to perform 100 // some special action once. This is the reason for the extra "Claimed" state. 101 typedef jint G1RemsetIterState; 102 103 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet. 104 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned. 105 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned. 106 107 G1RemsetIterState volatile* _iter_states; 108 // The current location where the next thread should continue scanning in a region's 109 // remembered set. 110 size_t volatile* _iter_claims; 111 112 // Temporary buffer holding the regions we used to store remembered set scan duplicate 113 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region) 114 uint* _dirty_region_buffer; 115 116 typedef jbyte IsDirtyRegionState; 117 static const IsDirtyRegionState Clean = 0; 118 static const IsDirtyRegionState Dirty = 1; 119 // Holds a flag for every region whether it is in the _dirty_region_buffer already 120 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools. 121 IsDirtyRegionState* _in_dirty_region_buffer; 122 size_t _cur_dirty_region; 123 124 // Creates a snapshot of the current _top values at the start of collection to 125 // filter out card marks that we do not want to scan. 126 class G1ResetScanTopClosure : public HeapRegionClosure { 127 private: 128 HeapWord** _scan_top; 129 public: 130 G1ResetScanTopClosure(HeapWord** scan_top) : _scan_top(scan_top) { } 131 132 virtual bool do_heap_region(HeapRegion* r) { 133 uint hrm_index = r->hrm_index(); 134 if (!r->in_collection_set() && r->is_old_or_humongous()) { 135 _scan_top[hrm_index] = r->top(); 136 } else { 137 _scan_top[hrm_index] = r->bottom(); 138 } 139 return false; 140 } 141 }; 142 143 // For each region, contains the maximum top() value to be used during this garbage 144 // collection. Subsumes common checks like filtering out everything but old and 145 // humongous regions outside the collection set. 146 // This is valid because we are not interested in scanning stray remembered set 147 // entries from free or archive regions. 148 HeapWord** _scan_top; 149 public: 150 G1RemSetScanState() : 151 _max_regions(0), 152 _iter_states(NULL), 153 _iter_claims(NULL), 154 _dirty_region_buffer(NULL), 155 _in_dirty_region_buffer(NULL), 156 _cur_dirty_region(0), 157 _scan_top(NULL) { 158 } 159 160 ~G1RemSetScanState() { 161 if (_iter_states != NULL) { 162 FREE_C_HEAP_ARRAY(G1RemsetIterState, _iter_states); 163 } 164 if (_iter_claims != NULL) { 165 FREE_C_HEAP_ARRAY(size_t, _iter_claims); 166 } 167 if (_dirty_region_buffer != NULL) { 168 FREE_C_HEAP_ARRAY(uint, _dirty_region_buffer); 169 } 170 if (_in_dirty_region_buffer != NULL) { 171 FREE_C_HEAP_ARRAY(IsDirtyRegionState, _in_dirty_region_buffer); 172 } 173 if (_scan_top != NULL) { 174 FREE_C_HEAP_ARRAY(HeapWord*, _scan_top); 175 } 176 } 177 178 void initialize(uint max_regions) { 179 assert(_iter_states == NULL, "Must not be initialized twice"); 180 assert(_iter_claims == NULL, "Must not be initialized twice"); 181 _max_regions = max_regions; 182 _iter_states = NEW_C_HEAP_ARRAY(G1RemsetIterState, max_regions, mtGC); 183 _iter_claims = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC); 184 _dirty_region_buffer = NEW_C_HEAP_ARRAY(uint, max_regions, mtGC); 185 _in_dirty_region_buffer = NEW_C_HEAP_ARRAY(IsDirtyRegionState, max_regions, mtGC); 186 _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_regions, mtGC); 187 } 188 189 void reset() { 190 for (uint i = 0; i < _max_regions; i++) { 191 _iter_states[i] = Unclaimed; 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(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState)); 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] == Dirty) { 243 return; 244 } 245 246 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean; 247 if (marked_as_dirty) { 248 size_t allocated = Atomic::add(1u, &_cur_dirty_region) - 1; 249 _dirty_region_buffer[allocated] = region; 250 } 251 } 252 253 HeapWord* scan_top(uint region_idx) const { 254 return _scan_top[region_idx]; 255 } 256 257 // Clear the card table of "dirty" regions. 258 void clear_card_table(WorkGang* workers) { 259 if (_cur_dirty_region == 0) { 260 return; 261 } 262 263 size_t const num_chunks = align_up(_cur_dirty_region * HeapRegion::CardsPerRegion, G1ClearCardTableTask::chunk_size()) / G1ClearCardTableTask::chunk_size(); 264 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers()); 265 size_t const chunk_length = G1ClearCardTableTask::chunk_size() / HeapRegion::CardsPerRegion; 266 267 // Iterate over the dirty cards region list. 268 G1ClearCardTableTask cl(G1CollectedHeap::heap(), _dirty_region_buffer, _cur_dirty_region, chunk_length); 269 270 log_debug(gc, ergo)("Running %s using %u workers for " SIZE_FORMAT " " 271 "units of work for " SIZE_FORMAT " regions.", 272 cl.name(), num_workers, num_chunks, _cur_dirty_region); 273 workers->run_task(&cl, num_workers); 274 275 #ifndef PRODUCT 276 G1CollectedHeap::heap()->verifier()->verify_card_table_cleanup(); 277 #endif 278 } 279 }; 280 281 G1RemSet::G1RemSet(G1CollectedHeap* g1h, 282 G1CardTable* ct, 283 G1HotCardCache* hot_card_cache) : 284 _g1h(g1h), 285 _scan_state(new G1RemSetScanState()), 286 _num_conc_refined_cards(0), 287 _ct(ct), 288 _g1p(_g1h->g1_policy()), 289 _hot_card_cache(hot_card_cache), 290 _prev_period_summary() { 291 } 292 293 G1RemSet::~G1RemSet() { 294 if (_scan_state != NULL) { 295 delete _scan_state; 296 } 297 } 298 299 uint G1RemSet::num_par_rem_sets() { 300 return DirtyCardQueueSet::num_par_ids() + G1ConcurrentRefine::max_num_threads() + MAX2(ConcGCThreads, ParallelGCThreads); 301 } 302 303 void G1RemSet::initialize(size_t capacity, uint max_regions) { 304 G1FromCardCache::initialize(num_par_rem_sets(), max_regions); 305 _scan_state->initialize(max_regions); 306 } 307 308 G1ScanRSForRegionClosure::G1ScanRSForRegionClosure(G1RemSetScanState* scan_state, 309 G1ScanObjsDuringScanRSClosure* scan_obj_on_card, 310 G1ParScanThreadState* pss, 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 _worker_i(worker_i), 318 _cards_claimed(0), 319 _cards_scanned(0), 320 _cards_skipped(0), 321 _rem_set_root_scan_time(), 322 _rem_set_trim_partially_time(), 323 _strong_code_root_scan_time(), 324 _strong_code_trim_partially_time() { 325 } 326 327 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){ 328 _ct->set_card_claimed(card_index); 329 _scan_state->add_dirty_region(region_idx_for_card); 330 } 331 332 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) { 333 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card); 334 _scan_objs_on_card_cl->set_region(card_region); 335 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl); 336 _scan_objs_on_card_cl->trim_queue_partially(); 337 _cards_scanned++; 338 } 339 340 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) { 341 uint const region_idx = r->hrm_index(); 342 343 if (_scan_state->claim_iter(region_idx)) { 344 // If we ever free the collection set concurrently, we should also 345 // clear the card table concurrently therefore we won't need to 346 // add regions of the collection set to the dirty cards region. 347 _scan_state->add_dirty_region(region_idx); 348 } 349 350 // We claim cards in blocks so as to reduce the contention. 351 size_t const block_size = G1RSetScanBlockSize; 352 353 HeapRegionRemSetIterator iter(r->rem_set()); 354 size_t card_index; 355 356 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size); 357 for (size_t current_card = 0; iter.has_next(card_index); current_card++) { 358 if (current_card >= claimed_card_block + block_size) { 359 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size); 360 } 361 if (current_card < claimed_card_block) { 362 _cards_skipped++; 363 continue; 364 } 365 _cards_claimed++; 366 367 // If the card is dirty, then G1 will scan it during Update RS. 368 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) { 369 continue; 370 } 371 372 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index); 373 uint const region_idx_for_card = _g1h->addr_to_region(card_start); 374 375 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start), 376 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index()); 377 HeapWord* const top = _scan_state->scan_top(region_idx_for_card); 378 if (card_start >= top) { 379 continue; 380 } 381 382 // We claim lazily (so races are possible but they're benign), which reduces the 383 // number of duplicate scans (the rsets of the regions in the cset can intersect). 384 // Claim the card after checking bounds above: the remembered set may contain 385 // random cards into current survivor, and we would then have an incorrectly 386 // claimed card in survivor space. Card table clear does not reset the card table 387 // of survivor space regions. 388 claim_card(card_index, region_idx_for_card); 389 390 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top)); 391 392 scan_card(mr, region_idx_for_card); 393 } 394 } 395 396 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) { 397 r->strong_code_roots_do(_pss->closures()->weak_codeblobs()); 398 } 399 400 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) { 401 assert(r->in_collection_set(), 402 "Should only be called on elements of the collection set but region %u is not.", 403 r->hrm_index()); 404 uint const region_idx = r->hrm_index(); 405 406 // Do an early out if we know we are complete. 407 if (_scan_state->iter_is_complete(region_idx)) { 408 return false; 409 } 410 411 { 412 G1EvacPhaseWithTrimTimeTracker timer(_pss, _rem_set_root_scan_time, _rem_set_trim_partially_time); 413 scan_rem_set_roots(r); 414 } 415 416 if (_scan_state->set_iter_complete(region_idx)) { 417 G1EvacPhaseWithTrimTimeTracker timer(_pss, _strong_code_root_scan_time, _strong_code_trim_partially_time); 418 // Scan the strong code root list attached to the current region 419 scan_strong_code_roots(r); 420 } 421 return false; 422 } 423 424 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) { 425 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss); 426 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i); 427 _g1h->collection_set_iterate_from(&cl, worker_i); 428 429 G1GCPhaseTimes* p = _g1p->phase_times(); 430 431 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, TicksToTimeHelper::seconds(cl.rem_set_root_scan_time())); 432 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, TicksToTimeHelper::seconds(cl.rem_set_trim_partially_time())); 433 434 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards); 435 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards); 436 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards); 437 438 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, TicksToTimeHelper::seconds(cl.strong_code_root_scan_time())); 439 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, TicksToTimeHelper::seconds(cl.strong_code_root_trim_partially_time())); 440 } 441 442 // Closure used for updating rem sets. Only called during an evacuation pause. 443 class G1RefineCardClosure: public CardTableEntryClosure { 444 G1RemSet* _g1rs; 445 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl; 446 447 size_t _cards_scanned; 448 size_t _cards_skipped; 449 public: 450 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) : 451 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0) 452 {} 453 454 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 455 // The only time we care about recording cards that 456 // contain references that point into the collection set 457 // is during RSet updating within an evacuation pause. 458 // In this case worker_i should be the id of a GC worker thread. 459 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause"); 460 461 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl); 462 463 if (card_scanned) { 464 _update_rs_cl->trim_queue_partially(); 465 _cards_scanned++; 466 } else { 467 _cards_skipped++; 468 } 469 return true; 470 } 471 472 size_t cards_scanned() const { return _cards_scanned; } 473 size_t cards_skipped() const { return _cards_skipped; } 474 }; 475 476 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) { 477 G1GCPhaseTimes* p = _g1p->phase_times(); 478 479 // Apply closure to log entries in the HCC. 480 if (G1HotCardCache::default_use_cache()) { 481 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::ScanHCC, worker_i); 482 483 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss, worker_i); 484 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl); 485 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i); 486 } 487 488 // Now apply the closure to all remaining log entries. 489 { 490 G1EvacPhaseTimesTracker x(p, pss, G1GCPhaseTimes::UpdateRS, worker_i); 491 492 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i); 493 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl); 494 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i); 495 496 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards); 497 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards); 498 } 499 } 500 501 void G1RemSet::cleanupHRRS() { 502 HeapRegionRemSet::cleanup(); 503 } 504 505 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) { 506 update_rem_set(pss, worker_i); 507 scan_rem_set(pss, worker_i);; 508 } 509 510 void G1RemSet::prepare_for_oops_into_collection_set_do() { 511 DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set(); 512 dcqs.concatenate_logs(); 513 514 _scan_state->reset(); 515 } 516 517 void G1RemSet::cleanup_after_oops_into_collection_set_do() { 518 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times(); 519 520 // Set all cards back to clean. 521 double start = os::elapsedTime(); 522 _scan_state->clear_card_table(_g1h->workers()); 523 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0); 524 } 525 526 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) { 527 #ifdef ASSERT 528 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 529 assert(g1h->is_in_exact(ct->addr_for(card_ptr)), 530 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap", 531 p2i(card_ptr), 532 ct->index_for(ct->addr_for(card_ptr)), 533 p2i(ct->addr_for(card_ptr)), 534 g1h->addr_to_region(ct->addr_for(card_ptr))); 535 #endif 536 } 537 538 void G1RemSet::refine_card_concurrently(jbyte* card_ptr, 539 uint worker_i) { 540 assert(!_g1h->is_gc_active(), "Only call concurrently"); 541 542 check_card_ptr(card_ptr, _ct); 543 544 // If the card is no longer dirty, nothing to do. 545 if (*card_ptr != G1CardTable::dirty_card_val()) { 546 return; 547 } 548 549 // Construct the region representing the card. 550 HeapWord* start = _ct->addr_for(card_ptr); 551 // And find the region containing it. 552 HeapRegion* r = _g1h->heap_region_containing(start); 553 554 // This check is needed for some uncommon cases where we should 555 // ignore the card. 556 // 557 // The region could be young. Cards for young regions are 558 // distinctly marked (set to g1_young_gen), so the post-barrier will 559 // filter them out. However, that marking is performed 560 // concurrently. A write to a young object could occur before the 561 // card has been marked young, slipping past the filter. 562 // 563 // The card could be stale, because the region has been freed since 564 // the card was recorded. In this case the region type could be 565 // anything. If (still) free or (reallocated) young, just ignore 566 // it. If (reallocated) old or humongous, the later card trimming 567 // and additional checks in iteration may detect staleness. At 568 // worst, we end up processing a stale card unnecessarily. 569 // 570 // In the normal (non-stale) case, the synchronization between the 571 // enqueueing of the card and processing it here will have ensured 572 // we see the up-to-date region type here. 573 if (!r->is_old_or_humongous()) { 574 return; 575 } 576 577 // The result from the hot card cache insert call is either: 578 // * pointer to the current card 579 // (implying that the current card is not 'hot'), 580 // * null 581 // (meaning we had inserted the card ptr into the "hot" card cache, 582 // which had some headroom), 583 // * a pointer to a "hot" card that was evicted from the "hot" cache. 584 // 585 586 if (_hot_card_cache->use_cache()) { 587 assert(!SafepointSynchronize::is_at_safepoint(), "sanity"); 588 589 const jbyte* orig_card_ptr = card_ptr; 590 card_ptr = _hot_card_cache->insert(card_ptr); 591 if (card_ptr == NULL) { 592 // There was no eviction. Nothing to do. 593 return; 594 } else if (card_ptr != orig_card_ptr) { 595 // Original card was inserted and an old card was evicted. 596 start = _ct->addr_for(card_ptr); 597 r = _g1h->heap_region_containing(start); 598 599 // Check whether the region formerly in the cache should be 600 // ignored, as discussed earlier for the original card. The 601 // region could have been freed while in the cache. 602 if (!r->is_old_or_humongous()) { 603 return; 604 } 605 } // Else we still have the original card. 606 } 607 608 // Trim the region designated by the card to what's been allocated 609 // in the region. The card could be stale, or the card could cover 610 // (part of) an object at the end of the allocated space and extend 611 // beyond the end of allocation. 612 613 // Non-humongous objects are only allocated in the old-gen during 614 // GC, so if region is old then top is stable. Humongous object 615 // allocation sets top last; if top has not yet been set, this is 616 // a stale card and we'll end up with an empty intersection. If 617 // this is not a stale card, the synchronization between the 618 // enqueuing of the card and processing it here will have ensured 619 // we see the up-to-date top here. 620 HeapWord* scan_limit = r->top(); 621 622 if (scan_limit <= start) { 623 // If the trimmed region is empty, the card must be stale. 624 return; 625 } 626 627 // Okay to clean and process the card now. There are still some 628 // stale card cases that may be detected by iteration and dealt with 629 // as iteration failure. 630 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val(); 631 632 // This fence serves two purposes. First, the card must be cleaned 633 // before processing the contents. Second, we can't proceed with 634 // processing until after the read of top, for synchronization with 635 // possibly concurrent humongous object allocation. It's okay that 636 // reading top and reading type were racy wrto each other. We need 637 // both set, in any order, to proceed. 638 OrderAccess::fence(); 639 640 // Don't use addr_for(card_ptr + 1) which can ask for 641 // a card beyond the heap. 642 HeapWord* end = start + G1CardTable::card_size_in_words; 643 MemRegion dirty_region(start, MIN2(scan_limit, end)); 644 assert(!dirty_region.is_empty(), "sanity"); 645 646 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i); 647 648 bool card_processed = 649 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl); 650 651 // If unable to process the card then we encountered an unparsable 652 // part of the heap (e.g. a partially allocated object) while 653 // processing a stale card. Despite the card being stale, redirty 654 // and re-enqueue, because we've already cleaned the card. Without 655 // this we could incorrectly discard a non-stale card. 656 if (!card_processed) { 657 // The card might have gotten re-dirtied and re-enqueued while we 658 // worked. (In fact, it's pretty likely.) 659 if (*card_ptr != G1CardTable::dirty_card_val()) { 660 *card_ptr = G1CardTable::dirty_card_val(); 661 MutexLockerEx x(Shared_DirtyCardQ_lock, 662 Mutex::_no_safepoint_check_flag); 663 DirtyCardQueue* sdcq = 664 G1BarrierSet::dirty_card_queue_set().shared_dirty_card_queue(); 665 sdcq->enqueue(card_ptr); 666 } 667 } else { 668 _num_conc_refined_cards++; // Unsynchronized update, only used for logging. 669 } 670 } 671 672 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr, 673 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) { 674 assert(_g1h->is_gc_active(), "Only call during GC"); 675 676 check_card_ptr(card_ptr, _ct); 677 678 // If the card is no longer dirty, nothing to do. This covers cards that were already 679 // scanned as parts of the remembered sets. 680 if (*card_ptr != G1CardTable::dirty_card_val()) { 681 return false; 682 } 683 684 // We claim lazily (so races are possible but they're benign), which reduces the 685 // number of potential duplicate scans (multiple threads may enqueue the same card twice). 686 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val(); 687 688 // Construct the region representing the card. 689 HeapWord* card_start = _ct->addr_for(card_ptr); 690 // And find the region containing it. 691 uint const card_region_idx = _g1h->addr_to_region(card_start); 692 693 _scan_state->add_dirty_region(card_region_idx); 694 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx); 695 if (scan_limit <= card_start) { 696 // If the card starts above the area in the region containing objects to scan, skip it. 697 return false; 698 } 699 700 // Don't use addr_for(card_ptr + 1) which can ask for 701 // a card beyond the heap. 702 HeapWord* card_end = card_start + G1CardTable::card_size_in_words; 703 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end)); 704 assert(!dirty_region.is_empty(), "sanity"); 705 706 HeapRegion* const card_region = _g1h->region_at(card_region_idx); 707 update_rs_cl->set_region(card_region); 708 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl); 709 assert(card_processed, "must be"); 710 return true; 711 } 712 713 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) { 714 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) && 715 (period_count % G1SummarizeRSetStatsPeriod == 0)) { 716 717 G1RemSetSummary current(this); 718 _prev_period_summary.subtract_from(¤t); 719 720 Log(gc, remset) log; 721 log.trace("%s", header); 722 ResourceMark rm; 723 LogStream ls(log.trace()); 724 _prev_period_summary.print_on(&ls); 725 726 _prev_period_summary.set(¤t); 727 } 728 } 729 730 void G1RemSet::print_summary_info() { 731 Log(gc, remset, exit) log; 732 if (log.is_trace()) { 733 log.trace(" Cumulative RS summary"); 734 G1RemSetSummary current(this); 735 ResourceMark rm; 736 LogStream ls(log.trace()); 737 current.print_on(&ls); 738 } 739 } 740 741 class G1RebuildRemSetTask: public AbstractGangTask { 742 // Aggregate the counting data that was constructed concurrently 743 // with marking. 744 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure { 745 G1ConcurrentMark* _cm; 746 G1RebuildRemSetClosure _update_cl; 747 748 // Applies _update_cl to the references of the given object, limiting objArrays 749 // to the given MemRegion. Returns the amount of words actually scanned. 750 size_t scan_for_references(oop const obj, MemRegion mr) { 751 size_t const obj_size = obj->size(); 752 // All non-objArrays and objArrays completely within the mr 753 // can be scanned without passing the mr. 754 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) { 755 obj->oop_iterate(&_update_cl); 756 return obj_size; 757 } 758 // This path is for objArrays crossing the given MemRegion. Only scan the 759 // area within the MemRegion. 760 obj->oop_iterate(&_update_cl, mr); 761 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size(); 762 } 763 764 // A humongous object is live (with respect to the scanning) either 765 // a) it is marked on the bitmap as such 766 // b) its TARS is larger than TAMS, i.e. has been allocated during marking. 767 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const { 768 return bitmap->is_marked(humongous_obj) || (tars > tams); 769 } 770 771 // Iterator over the live objects within the given MemRegion. 772 class LiveObjIterator : public StackObj { 773 const G1CMBitMap* const _bitmap; 774 const HeapWord* _tams; 775 const MemRegion _mr; 776 HeapWord* _current; 777 778 bool is_below_tams() const { 779 return _current < _tams; 780 } 781 782 bool is_live(HeapWord* obj) const { 783 return !is_below_tams() || _bitmap->is_marked(obj); 784 } 785 786 HeapWord* bitmap_limit() const { 787 return MIN2(const_cast<HeapWord*>(_tams), _mr.end()); 788 } 789 790 void move_if_below_tams() { 791 if (is_below_tams() && has_next()) { 792 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 793 } 794 } 795 public: 796 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) : 797 _bitmap(bitmap), 798 _tams(tams), 799 _mr(mr), 800 _current(first_oop_into_mr) { 801 802 assert(_current <= _mr.start(), 803 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")", 804 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end())); 805 806 // Step to the next live object within the MemRegion if needed. 807 if (is_live(_current)) { 808 // Non-objArrays were scanned by the previous part of that region. 809 if (_current < mr.start() && !oop(_current)->is_objArray()) { 810 _current += oop(_current)->size(); 811 // We might have positioned _current on a non-live object. Reposition to the next 812 // live one if needed. 813 move_if_below_tams(); 814 } 815 } else { 816 // The object at _current can only be dead if below TAMS, so we can use the bitmap. 817 // immediately. 818 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 819 assert(_current == _mr.end() || is_live(_current), 820 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")", 821 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end())); 822 } 823 } 824 825 void move_to_next() { 826 _current += next()->size(); 827 move_if_below_tams(); 828 } 829 830 oop next() const { 831 oop result = oop(_current); 832 assert(is_live(_current), 833 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d", 834 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result)); 835 return result; 836 } 837 838 bool has_next() const { 839 return _current < _mr.end(); 840 } 841 }; 842 843 // Rebuild remembered sets in the part of the region specified by mr and hr. 844 // Objects between the bottom of the region and the TAMS are checked for liveness 845 // using the given bitmap. Objects between TAMS and TARS are assumed to be live. 846 // Returns the number of live words between bottom and TAMS. 847 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap, 848 HeapWord* const top_at_mark_start, 849 HeapWord* const top_at_rebuild_start, 850 HeapRegion* hr, 851 MemRegion mr) { 852 size_t marked_words = 0; 853 854 if (hr->is_humongous()) { 855 oop const humongous_obj = oop(hr->humongous_start_region()->bottom()); 856 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) { 857 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start); 858 // however in case of humongous objects it is sufficient to scan the encompassing 859 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the 860 // two areas will be zero sized. I.e. TAMS is either 861 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different 862 // value: this would mean that TAMS points somewhere into the object. 863 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start, 864 "More than one object in the humongous region?"); 865 humongous_obj->oop_iterate(&_update_cl, mr); 866 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0; 867 } else { 868 return 0; 869 } 870 } 871 872 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) { 873 oop obj = it.next(); 874 size_t scanned_size = scan_for_references(obj, mr); 875 if ((HeapWord*)obj < top_at_mark_start) { 876 marked_words += scanned_size; 877 } 878 } 879 880 return marked_words * HeapWordSize; 881 } 882 public: 883 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h, 884 G1ConcurrentMark* cm, 885 uint worker_id) : 886 HeapRegionClosure(), 887 _cm(cm), 888 _update_cl(g1h, worker_id) { } 889 890 bool do_heap_region(HeapRegion* hr) { 891 if (_cm->has_aborted()) { 892 return true; 893 } 894 895 uint const region_idx = hr->hrm_index(); 896 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);) 897 assert(top_at_rebuild_start_check == NULL || 898 top_at_rebuild_start_check > hr->bottom(), 899 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)", 900 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str()); 901 902 size_t total_marked_bytes = 0; 903 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize; 904 905 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start(); 906 907 HeapWord* cur = hr->bottom(); 908 while (cur < hr->end()) { 909 // After every iteration (yield point) we need to check whether the region's 910 // TARS changed due to e.g. eager reclaim. 911 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx); 912 if (top_at_rebuild_start == NULL) { 913 return false; 914 } 915 916 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words)); 917 if (next_chunk.is_empty()) { 918 break; 919 } 920 921 const Ticks start = Ticks::now(); 922 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(), 923 top_at_mark_start, 924 top_at_rebuild_start, 925 hr, 926 next_chunk); 927 Tickspan time = Ticks::now() - start; 928 929 log_trace(gc, remset, tracking)("Rebuilt region %u " 930 "live " SIZE_FORMAT " " 931 "time %.3fms " 932 "marked bytes " SIZE_FORMAT " " 933 "bot " PTR_FORMAT " " 934 "TAMS " PTR_FORMAT " " 935 "TARS " PTR_FORMAT, 936 region_idx, 937 _cm->liveness(region_idx) * HeapWordSize, 938 TicksToTimeHelper::seconds(time) * 1000.0, 939 marked_bytes, 940 p2i(hr->bottom()), 941 p2i(top_at_mark_start), 942 p2i(top_at_rebuild_start)); 943 944 if (marked_bytes > 0) { 945 total_marked_bytes += marked_bytes; 946 } 947 cur += chunk_size_in_words; 948 949 _cm->do_yield_check(); 950 if (_cm->has_aborted()) { 951 return true; 952 } 953 } 954 // In the final iteration of the loop the region might have been eagerly reclaimed. 955 // Simply filter out those regions. We can not just use region type because there 956 // might have already been new allocations into these regions. 957 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);) 958 assert(top_at_rebuild_start == NULL || 959 total_marked_bytes == hr->marked_bytes(), 960 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " " 961 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")", 962 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(), 963 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start)); 964 // Abort state may have changed after the yield check. 965 return _cm->has_aborted(); 966 } 967 }; 968 969 HeapRegionClaimer _hr_claimer; 970 G1ConcurrentMark* _cm; 971 972 uint _worker_id_offset; 973 public: 974 G1RebuildRemSetTask(G1ConcurrentMark* cm, 975 uint n_workers, 976 uint worker_id_offset) : 977 AbstractGangTask("G1 Rebuild Remembered Set"), 978 _cm(cm), 979 _hr_claimer(n_workers), 980 _worker_id_offset(worker_id_offset) { 981 } 982 983 void work(uint worker_id) { 984 SuspendibleThreadSetJoiner sts_join; 985 986 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 987 988 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id); 989 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id); 990 } 991 }; 992 993 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm, 994 WorkGang* workers, 995 uint worker_id_offset) { 996 uint num_workers = workers->active_workers(); 997 998 G1RebuildRemSetTask cl(cm, 999 num_workers, 1000 worker_id_offset); 1001 workers->run_task(&cl, num_workers); 1002 }