1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/g1/dirtyCardQueue.hpp" 27 #include "gc/g1/g1BlockOffsetTable.inline.hpp" 28 #include "gc/g1/g1CardTable.inline.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1ConcurrentRefine.hpp" 31 #include "gc/g1/g1FromCardCache.hpp" 32 #include "gc/g1/g1GCPhaseTimes.hpp" 33 #include "gc/g1/g1HotCardCache.hpp" 34 #include "gc/g1/g1OopClosures.inline.hpp" 35 #include "gc/g1/g1RootClosures.hpp" 36 #include "gc/g1/g1RemSet.hpp" 37 #include "gc/g1/heapRegion.inline.hpp" 38 #include "gc/g1/heapRegionManager.inline.hpp" 39 #include "gc/g1/heapRegionRemSet.hpp" 40 #include "gc/shared/gcTraceTime.inline.hpp" 41 #include "gc/shared/suspendibleThreadSet.hpp" 42 #include "memory/iterator.hpp" 43 #include "memory/resourceArea.hpp" 44 #include "oops/access.inline.hpp" 45 #include "oops/oop.inline.hpp" 46 #include "utilities/align.hpp" 47 #include "utilities/globalDefinitions.hpp" 48 #include "utilities/intHisto.hpp" 49 #include "utilities/stack.inline.hpp" 50 #include "utilities/ticks.inline.hpp" 51 52 // Collects information about the overall remembered set scan progress during an evacuation. 53 class G1RemSetScanState : public CHeapObj<mtGC> { 54 private: 55 class G1ClearCardTableTask : public AbstractGangTask { 56 G1CollectedHeap* _g1h; 57 uint* _dirty_region_list; 58 size_t _num_dirty_regions; 59 size_t _chunk_length; 60 61 size_t volatile _cur_dirty_regions; 62 public: 63 G1ClearCardTableTask(G1CollectedHeap* g1h, 64 uint* dirty_region_list, 65 size_t num_dirty_regions, 66 size_t chunk_length) : 67 AbstractGangTask("G1 Clear Card Table Task"), 68 _g1h(g1h), 69 _dirty_region_list(dirty_region_list), 70 _num_dirty_regions(num_dirty_regions), 71 _chunk_length(chunk_length), 72 _cur_dirty_regions(0) { 73 74 assert(chunk_length > 0, "must be"); 75 } 76 77 static size_t chunk_size() { return M; } 78 79 void work(uint worker_id) { 80 while (_cur_dirty_regions < _num_dirty_regions) { 81 size_t next = Atomic::add(_chunk_length, &_cur_dirty_regions) - _chunk_length; 82 size_t max = MIN2(next + _chunk_length, _num_dirty_regions); 83 84 for (size_t i = next; i < max; i++) { 85 HeapRegion* r = _g1h->region_at(_dirty_region_list[i]); 86 if (!r->is_survivor()) { 87 r->clear_cardtable(); 88 } 89 } 90 } 91 } 92 }; 93 94 size_t _max_regions; 95 96 // Scan progress for the remembered set of a single region. Transitions from 97 // Unclaimed -> Claimed -> Complete. 98 // At each of the transitions the thread that does the transition needs to perform 99 // some special action once. This is the reason for the extra "Claimed" state. 100 typedef jint G1RemsetIterState; 101 102 static const G1RemsetIterState Unclaimed = 0; // The remembered set has not been scanned yet. 103 static const G1RemsetIterState Claimed = 1; // The remembered set is currently being scanned. 104 static const G1RemsetIterState Complete = 2; // The remembered set has been completely scanned. 105 106 G1RemsetIterState volatile* _iter_states; 107 // The current location where the next thread should continue scanning in a region's 108 // remembered set. 109 size_t volatile* _iter_claims; 110 111 // Temporary buffer holding the regions we used to store remembered set scan duplicate 112 // information. These are also called "dirty". Valid entries are from [0.._cur_dirty_region) 113 uint* _dirty_region_buffer; 114 115 typedef jbyte IsDirtyRegionState; 116 static const IsDirtyRegionState Clean = 0; 117 static const IsDirtyRegionState Dirty = 1; 118 // Holds a flag for every region whether it is in the _dirty_region_buffer already 119 // to avoid duplicates. Uses jbyte since there are no atomic instructions for bools. 120 IsDirtyRegionState* _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()) { 134 _scan_top[hrm_index] = r->top(); 135 } else { 136 _scan_top[hrm_index] = r->bottom(); 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(IsDirtyRegionState, _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(IsDirtyRegionState, 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 } 192 193 G1ResetScanTopClosure cl(_scan_top); 194 G1CollectedHeap::heap()->heap_region_iterate(&cl); 195 196 memset((void*)_iter_claims, 0, _max_regions * sizeof(size_t)); 197 memset(_in_dirty_region_buffer, Clean, _max_regions * sizeof(IsDirtyRegionState)); 198 _cur_dirty_region = 0; 199 } 200 201 // Attempt to claim the remembered set of the region for iteration. Returns true 202 // if this call caused the transition from Unclaimed to Claimed. 203 inline bool claim_iter(uint region) { 204 assert(region < _max_regions, "Tried to access invalid region %u", region); 205 if (_iter_states[region] != Unclaimed) { 206 return false; 207 } 208 G1RemsetIterState res = Atomic::cmpxchg(Claimed, &_iter_states[region], Unclaimed); 209 return (res == Unclaimed); 210 } 211 212 // Try to atomically sets the iteration state to "complete". Returns true for the 213 // thread that caused the transition. 214 inline bool set_iter_complete(uint region) { 215 if (iter_is_complete(region)) { 216 return false; 217 } 218 G1RemsetIterState res = Atomic::cmpxchg(Complete, &_iter_states[region], Claimed); 219 return (res == Claimed); 220 } 221 222 // Returns true if the region's iteration is complete. 223 inline bool iter_is_complete(uint region) const { 224 assert(region < _max_regions, "Tried to access invalid region %u", region); 225 return _iter_states[region] == Complete; 226 } 227 228 // The current position within the remembered set of the given region. 229 inline size_t iter_claimed(uint region) const { 230 assert(region < _max_regions, "Tried to access invalid region %u", region); 231 return _iter_claims[region]; 232 } 233 234 // Claim the next block of cards within the remembered set of the region with 235 // step size. 236 inline size_t iter_claimed_next(uint region, size_t step) { 237 return Atomic::add(step, &_iter_claims[region]) - step; 238 } 239 240 void add_dirty_region(uint region) { 241 if (_in_dirty_region_buffer[region] == Dirty) { 242 return; 243 } 244 245 bool marked_as_dirty = Atomic::cmpxchg(Dirty, &_in_dirty_region_buffer[region], Clean) == Clean; 246 if (marked_as_dirty) { 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 _g1h(g1h), 284 _scan_state(new G1RemSetScanState()), 285 _num_conc_refined_cards(0), 286 _ct(ct), 287 _g1p(_g1h->g1_policy()), 288 _hot_card_cache(hot_card_cache), 289 _prev_period_summary() { 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 DirtyCardQueueSet::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 uint worker_i) : 311 _g1h(G1CollectedHeap::heap()), 312 _ct(_g1h->card_table()), 313 _pss(pss), 314 _scan_objs_on_card_cl(scan_obj_on_card), 315 _scan_state(scan_state), 316 _worker_i(worker_i), 317 _cards_claimed(0), 318 _cards_scanned(0), 319 _cards_skipped(0), 320 _rem_set_root_scan_time(), 321 _rem_set_trim_partially_time(), 322 _strong_code_root_scan_time(), 323 _strong_code_trim_partially_time() { 324 } 325 326 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){ 327 _ct->set_card_claimed(card_index); 328 _scan_state->add_dirty_region(region_idx_for_card); 329 } 330 331 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) { 332 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card); 333 _scan_objs_on_card_cl->set_region(card_region); 334 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl); 335 _scan_objs_on_card_cl->trim_queue_partially(); 336 _cards_scanned++; 337 } 338 339 void G1ScanRSForRegionClosure::scan_rem_set_roots(HeapRegion* r) { 340 uint const region_idx = r->hrm_index(); 341 342 if (_scan_state->claim_iter(region_idx)) { 343 // If we ever free the collection set concurrently, we should also 344 // clear the card table concurrently therefore we won't need to 345 // add regions of the collection set to the dirty cards region. 346 _scan_state->add_dirty_region(region_idx); 347 } 348 349 // We claim cards in blocks so as to reduce the contention. 350 size_t const block_size = G1RSetScanBlockSize; 351 352 HeapRegionRemSetIterator iter(r->rem_set()); 353 size_t card_index; 354 355 size_t claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size); 356 for (size_t current_card = 0; iter.has_next(card_index); current_card++) { 357 if (current_card >= claimed_card_block + block_size) { 358 claimed_card_block = _scan_state->iter_claimed_next(region_idx, block_size); 359 } 360 if (current_card < claimed_card_block) { 361 _cards_skipped++; 362 continue; 363 } 364 _cards_claimed++; 365 366 // If the card is dirty, then G1 will scan it during Update RS. 367 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) { 368 continue; 369 } 370 371 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index); 372 uint const region_idx_for_card = _g1h->addr_to_region(card_start); 373 374 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start), 375 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index()); 376 HeapWord* const top = _scan_state->scan_top(region_idx_for_card); 377 if (card_start >= top) { 378 continue; 379 } 380 381 // We claim lazily (so races are possible but they're benign), which reduces the 382 // number of duplicate scans (the rsets of the regions in the cset can intersect). 383 // Claim the card after checking bounds above: the remembered set may contain 384 // random cards into current survivor, and we would then have an incorrectly 385 // claimed card in survivor space. Card table clear does not reset the card table 386 // of survivor space regions. 387 claim_card(card_index, region_idx_for_card); 388 389 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top)); 390 391 scan_card(mr, region_idx_for_card); 392 } 393 } 394 395 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) { 396 Ticks const scan_start = Ticks::now(); 397 r->strong_code_roots_do(_pss->closures()->weak_codeblobs()); 398 _strong_code_root_scan_time += (Ticks::now() - scan_start); 399 } 400 401 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) { 402 assert(r->in_collection_set(), 403 "Should only be called on elements of the collection set but region %u is not.", 404 r->hrm_index()); 405 uint const region_idx = r->hrm_index(); 406 407 // Do an early out if we know we are complete. 408 if (_scan_state->iter_is_complete(region_idx)) { 409 return false; 410 } 411 412 { 413 Ticks const start = Ticks::now(); 414 scan_rem_set_roots(r); 415 Tickspan const trim_partially_time = _pss->trim_ticks_and_reset(); 416 _rem_set_root_scan_time += (Ticks::now() - start) - trim_partially_time; 417 _rem_set_trim_partially_time += trim_partially_time; 418 } 419 420 if (_scan_state->set_iter_complete(region_idx)) { 421 Ticks const start = Ticks::now(); 422 // Scan the strong code root list attached to the current region 423 scan_strong_code_roots(r); 424 Tickspan const trim_partially_time = _pss->trim_ticks_and_reset(); 425 _strong_code_root_scan_time += (Ticks::now() - start) - trim_partially_time; 426 _strong_code_trim_partially_time += trim_partially_time; 427 } 428 return false; 429 } 430 431 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, uint worker_i) { 432 G1ScanObjsDuringScanRSClosure scan_cl(_g1h, pss); 433 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, pss, worker_i); 434 _g1h->collection_set_iterate_from(&cl, worker_i); 435 436 G1GCPhaseTimes* p = _g1p->phase_times(); 437 438 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, TicksToTimeHelper::seconds(cl.rem_set_root_scan_time())); 439 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, TicksToTimeHelper::seconds(cl.rem_set_trim_partially_time())); 440 441 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards); 442 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards); 443 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards); 444 445 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, TicksToTimeHelper::seconds(cl.strong_code_root_scan_time())); 446 p->add_time_secs(G1GCPhaseTimes::ObjCopy, worker_i, TicksToTimeHelper::seconds(cl.strong_code_root_trim_partially_time())); 447 } 448 449 // Closure used for updating rem sets. Only called during an evacuation pause. 450 class G1RefineCardClosure: public CardTableEntryClosure { 451 G1RemSet* _g1rs; 452 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl; 453 454 size_t _cards_scanned; 455 size_t _cards_skipped; 456 public: 457 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) : 458 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0) 459 {} 460 461 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 462 // The only time we care about recording cards that 463 // contain references that point into the collection set 464 // is during RSet updating within an evacuation pause. 465 // In this case worker_i should be the id of a GC worker thread. 466 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause"); 467 468 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl); 469 470 if (card_scanned) { 471 _update_rs_cl->trim_queue_partially(); 472 _cards_scanned++; 473 } else { 474 _cards_skipped++; 475 } 476 return true; 477 } 478 479 size_t cards_scanned() const { return _cards_scanned; } 480 size_t cards_skipped() const { return _cards_skipped; } 481 }; 482 483 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) { 484 G1GCPhaseTimes* p = _g1p->phase_times(); 485 486 // Apply closure to log entries in the HCC. 487 if (G1HotCardCache::default_use_cache()) { 488 { 489 G1GCParPhaseTimesTracker x(p, G1GCPhaseTimes::ScanHCC, worker_i); 490 491 G1ScanObjsDuringUpdateRSClosure scan_hcc_cl(_g1h, pss, worker_i); 492 G1RefineCardClosure refine_card_cl(_g1h, &scan_hcc_cl); 493 494 _g1h->iterate_hcc_closure(&refine_card_cl, worker_i); 495 } 496 double const scan_hcc_trim_queue_time = TicksToTimeHelper::seconds(pss->trim_ticks_and_reset()); 497 p->move_time_secs(G1GCPhaseTimes::ScanHCC, G1GCPhaseTimes::ObjCopy, worker_i, scan_hcc_trim_queue_time); 498 } 499 500 // Now apply the closure to all remaining log entries. 501 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1h, pss, worker_i); 502 G1RefineCardClosure refine_card_cl(_g1h, &update_rs_cl); 503 { 504 G1GCParPhaseTimesTracker x(p, G1GCPhaseTimes::UpdateRS, worker_i); 505 _g1h->iterate_dirty_card_closure(&refine_card_cl, worker_i); 506 } 507 508 double const update_rs_trim_queue_time = TicksToTimeHelper::seconds(pss->trim_ticks_and_reset()); 509 p->move_time_secs(G1GCPhaseTimes::UpdateRS, G1GCPhaseTimes::ObjCopy, worker_i, update_rs_trim_queue_time); 510 511 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards); 512 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards); 513 } 514 515 void G1RemSet::cleanupHRRS() { 516 HeapRegionRemSet::cleanup(); 517 } 518 519 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, uint worker_i) { 520 update_rem_set(pss, worker_i); 521 scan_rem_set(pss, worker_i);; 522 } 523 524 void G1RemSet::prepare_for_oops_into_collection_set_do() { 525 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 526 dcqs.concatenate_logs(); 527 528 _scan_state->reset(); 529 } 530 531 void G1RemSet::cleanup_after_oops_into_collection_set_do() { 532 G1GCPhaseTimes* phase_times = _g1h->g1_policy()->phase_times(); 533 534 // Set all cards back to clean. 535 double start = os::elapsedTime(); 536 _scan_state->clear_card_table(_g1h->workers()); 537 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0); 538 } 539 540 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) { 541 #ifdef ASSERT 542 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 543 assert(g1h->is_in_exact(ct->addr_for(card_ptr)), 544 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap", 545 p2i(card_ptr), 546 ct->index_for(ct->addr_for(card_ptr)), 547 p2i(ct->addr_for(card_ptr)), 548 g1h->addr_to_region(ct->addr_for(card_ptr))); 549 #endif 550 } 551 552 void G1RemSet::refine_card_concurrently(jbyte* card_ptr, 553 uint worker_i) { 554 assert(!_g1h->is_gc_active(), "Only call concurrently"); 555 556 check_card_ptr(card_ptr, _ct); 557 558 // If the card is no longer dirty, nothing to do. 559 if (*card_ptr != G1CardTable::dirty_card_val()) { 560 return; 561 } 562 563 // Construct the region representing the card. 564 HeapWord* start = _ct->addr_for(card_ptr); 565 // And find the region containing it. 566 HeapRegion* r = _g1h->heap_region_containing(start); 567 568 // This check is needed for some uncommon cases where we should 569 // ignore the card. 570 // 571 // The region could be young. Cards for young regions are 572 // distinctly marked (set to g1_young_gen), so the post-barrier will 573 // filter them out. However, that marking is performed 574 // concurrently. A write to a young object could occur before the 575 // card has been marked young, slipping past the filter. 576 // 577 // The card could be stale, because the region has been freed since 578 // the card was recorded. In this case the region type could be 579 // anything. If (still) free or (reallocated) young, just ignore 580 // it. If (reallocated) old or humongous, the later card trimming 581 // and additional checks in iteration may detect staleness. At 582 // worst, we end up processing a stale card unnecessarily. 583 // 584 // In the normal (non-stale) case, the synchronization between the 585 // enqueueing of the card and processing it here will have ensured 586 // we see the up-to-date region type here. 587 if (!r->is_old_or_humongous()) { 588 return; 589 } 590 591 // The result from the hot card cache insert call is either: 592 // * pointer to the current card 593 // (implying that the current card is not 'hot'), 594 // * null 595 // (meaning we had inserted the card ptr into the "hot" card cache, 596 // which had some headroom), 597 // * a pointer to a "hot" card that was evicted from the "hot" cache. 598 // 599 600 if (_hot_card_cache->use_cache()) { 601 assert(!SafepointSynchronize::is_at_safepoint(), "sanity"); 602 603 const jbyte* orig_card_ptr = card_ptr; 604 card_ptr = _hot_card_cache->insert(card_ptr); 605 if (card_ptr == NULL) { 606 // There was no eviction. Nothing to do. 607 return; 608 } else if (card_ptr != orig_card_ptr) { 609 // Original card was inserted and an old card was evicted. 610 start = _ct->addr_for(card_ptr); 611 r = _g1h->heap_region_containing(start); 612 613 // Check whether the region formerly in the cache should be 614 // ignored, as discussed earlier for the original card. The 615 // region could have been freed while in the cache. 616 if (!r->is_old_or_humongous()) { 617 return; 618 } 619 } // Else we still have the original card. 620 } 621 622 // Trim the region designated by the card to what's been allocated 623 // in the region. The card could be stale, or the card could cover 624 // (part of) an object at the end of the allocated space and extend 625 // beyond the end of allocation. 626 627 // Non-humongous objects are only allocated in the old-gen during 628 // GC, so if region is old then top is stable. Humongous object 629 // allocation sets top last; if top has not yet been set, this is 630 // a stale card and we'll end up with an empty intersection. If 631 // this is not a stale card, the synchronization between the 632 // enqueuing of the card and processing it here will have ensured 633 // we see the up-to-date top here. 634 HeapWord* scan_limit = r->top(); 635 636 if (scan_limit <= start) { 637 // If the trimmed region is empty, the card must be stale. 638 return; 639 } 640 641 // Okay to clean and process the card now. There are still some 642 // stale card cases that may be detected by iteration and dealt with 643 // as iteration failure. 644 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val(); 645 646 // This fence serves two purposes. First, the card must be cleaned 647 // before processing the contents. Second, we can't proceed with 648 // processing until after the read of top, for synchronization with 649 // possibly concurrent humongous object allocation. It's okay that 650 // reading top and reading type were racy wrto each other. We need 651 // both set, in any order, to proceed. 652 OrderAccess::fence(); 653 654 // Don't use addr_for(card_ptr + 1) which can ask for 655 // a card beyond the heap. 656 HeapWord* end = start + G1CardTable::card_size_in_words; 657 MemRegion dirty_region(start, MIN2(scan_limit, end)); 658 assert(!dirty_region.is_empty(), "sanity"); 659 660 G1ConcurrentRefineOopClosure conc_refine_cl(_g1h, worker_i); 661 662 bool card_processed = 663 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl); 664 665 // If unable to process the card then we encountered an unparsable 666 // part of the heap (e.g. a partially allocated object) while 667 // processing a stale card. Despite the card being stale, redirty 668 // and re-enqueue, because we've already cleaned the card. Without 669 // this we could incorrectly discard a non-stale card. 670 if (!card_processed) { 671 // The card might have gotten re-dirtied and re-enqueued while we 672 // worked. (In fact, it's pretty likely.) 673 if (*card_ptr != G1CardTable::dirty_card_val()) { 674 *card_ptr = G1CardTable::dirty_card_val(); 675 MutexLockerEx x(Shared_DirtyCardQ_lock, 676 Mutex::_no_safepoint_check_flag); 677 DirtyCardQueue* sdcq = 678 JavaThread::dirty_card_queue_set().shared_dirty_card_queue(); 679 sdcq->enqueue(card_ptr); 680 } 681 } else { 682 _num_conc_refined_cards++; // Unsynchronized update, only used for logging. 683 } 684 } 685 686 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr, 687 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) { 688 assert(_g1h->is_gc_active(), "Only call during GC"); 689 690 check_card_ptr(card_ptr, _ct); 691 692 // If the card is no longer dirty, nothing to do. This covers cards that were already 693 // scanned as parts of the remembered sets. 694 if (*card_ptr != G1CardTable::dirty_card_val()) { 695 return false; 696 } 697 698 // We claim lazily (so races are possible but they're benign), which reduces the 699 // number of potential duplicate scans (multiple threads may enqueue the same card twice). 700 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val(); 701 702 // Construct the region representing the card. 703 HeapWord* card_start = _ct->addr_for(card_ptr); 704 // And find the region containing it. 705 uint const card_region_idx = _g1h->addr_to_region(card_start); 706 707 _scan_state->add_dirty_region(card_region_idx); 708 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx); 709 if (scan_limit <= card_start) { 710 // If the card starts above the area in the region containing objects to scan, skip it. 711 return false; 712 } 713 714 // Don't use addr_for(card_ptr + 1) which can ask for 715 // a card beyond the heap. 716 HeapWord* card_end = card_start + G1CardTable::card_size_in_words; 717 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end)); 718 assert(!dirty_region.is_empty(), "sanity"); 719 720 HeapRegion* const card_region = _g1h->region_at(card_region_idx); 721 update_rs_cl->set_region(card_region); 722 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl); 723 assert(card_processed, "must be"); 724 return true; 725 } 726 727 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) { 728 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) && 729 (period_count % G1SummarizeRSetStatsPeriod == 0)) { 730 731 G1RemSetSummary current(this); 732 _prev_period_summary.subtract_from(¤t); 733 734 Log(gc, remset) log; 735 log.trace("%s", header); 736 ResourceMark rm; 737 LogStream ls(log.trace()); 738 _prev_period_summary.print_on(&ls); 739 740 _prev_period_summary.set(¤t); 741 } 742 } 743 744 void G1RemSet::print_summary_info() { 745 Log(gc, remset, exit) log; 746 if (log.is_trace()) { 747 log.trace(" Cumulative RS summary"); 748 G1RemSetSummary current(this); 749 ResourceMark rm; 750 LogStream ls(log.trace()); 751 current.print_on(&ls); 752 } 753 } 754 755 class G1RebuildRemSetTask: public AbstractGangTask { 756 // Aggregate the counting data that was constructed concurrently 757 // with marking. 758 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure { 759 G1ConcurrentMark* _cm; 760 G1RebuildRemSetClosure _update_cl; 761 762 // Applies _update_cl to the references of the given object, limiting objArrays 763 // to the given MemRegion. Returns the amount of words actually scanned. 764 size_t scan_for_references(oop const obj, MemRegion mr) { 765 size_t const obj_size = obj->size(); 766 // All non-objArrays and objArrays completely within the mr 767 // can be scanned without passing the mr. 768 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) { 769 obj->oop_iterate(&_update_cl); 770 return obj_size; 771 } 772 // This path is for objArrays crossing the given MemRegion. Only scan the 773 // area within the MemRegion. 774 obj->oop_iterate(&_update_cl, mr); 775 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size(); 776 } 777 778 // A humongous object is live (with respect to the scanning) either 779 // a) it is marked on the bitmap as such 780 // b) its TARS is larger than TAMS, i.e. has been allocated during marking. 781 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const { 782 return bitmap->is_marked(humongous_obj) || (tars > tams); 783 } 784 785 // Iterator over the live objects within the given MemRegion. 786 class LiveObjIterator : public StackObj { 787 const G1CMBitMap* const _bitmap; 788 const HeapWord* _tams; 789 const MemRegion _mr; 790 HeapWord* _current; 791 792 bool is_below_tams() const { 793 return _current < _tams; 794 } 795 796 bool is_live(HeapWord* obj) const { 797 return !is_below_tams() || _bitmap->is_marked(obj); 798 } 799 800 HeapWord* bitmap_limit() const { 801 return MIN2(const_cast<HeapWord*>(_tams), _mr.end()); 802 } 803 804 void move_if_below_tams() { 805 if (is_below_tams() && has_next()) { 806 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 807 } 808 } 809 public: 810 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) : 811 _bitmap(bitmap), 812 _tams(tams), 813 _mr(mr), 814 _current(first_oop_into_mr) { 815 816 assert(_current <= _mr.start(), 817 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")", 818 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end())); 819 820 // Step to the next live object within the MemRegion if needed. 821 if (is_live(_current)) { 822 // Non-objArrays were scanned by the previous part of that region. 823 if (_current < mr.start() && !oop(_current)->is_objArray()) { 824 _current += oop(_current)->size(); 825 // We might have positioned _current on a non-live object. Reposition to the next 826 // live one if needed. 827 move_if_below_tams(); 828 } 829 } else { 830 // The object at _current can only be dead if below TAMS, so we can use the bitmap. 831 // immediately. 832 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 833 assert(_current == _mr.end() || is_live(_current), 834 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")", 835 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end())); 836 } 837 } 838 839 void move_to_next() { 840 _current += next()->size(); 841 move_if_below_tams(); 842 } 843 844 oop next() const { 845 oop result = oop(_current); 846 assert(is_live(_current), 847 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d", 848 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result)); 849 return result; 850 } 851 852 bool has_next() const { 853 return _current < _mr.end(); 854 } 855 }; 856 857 // Rebuild remembered sets in the part of the region specified by mr and hr. 858 // Objects between the bottom of the region and the TAMS are checked for liveness 859 // using the given bitmap. Objects between TAMS and TARS are assumed to be live. 860 // Returns the number of live words between bottom and TAMS. 861 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap, 862 HeapWord* const top_at_mark_start, 863 HeapWord* const top_at_rebuild_start, 864 HeapRegion* hr, 865 MemRegion mr) { 866 size_t marked_words = 0; 867 868 if (hr->is_humongous()) { 869 oop const humongous_obj = oop(hr->humongous_start_region()->bottom()); 870 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) { 871 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start); 872 // however in case of humongous objects it is sufficient to scan the encompassing 873 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the 874 // two areas will be zero sized. I.e. TAMS is either 875 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different 876 // value: this would mean that TAMS points somewhere into the object. 877 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start, 878 "More than one object in the humongous region?"); 879 humongous_obj->oop_iterate(&_update_cl, mr); 880 return top_at_mark_start != hr->bottom() ? mr.intersection(MemRegion((HeapWord*)humongous_obj, humongous_obj->size())).byte_size() : 0; 881 } else { 882 return 0; 883 } 884 } 885 886 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) { 887 oop obj = it.next(); 888 size_t scanned_size = scan_for_references(obj, mr); 889 if ((HeapWord*)obj < top_at_mark_start) { 890 marked_words += scanned_size; 891 } 892 } 893 894 return marked_words * HeapWordSize; 895 } 896 public: 897 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h, 898 G1ConcurrentMark* cm, 899 uint worker_id) : 900 HeapRegionClosure(), 901 _cm(cm), 902 _update_cl(g1h, worker_id) { } 903 904 bool do_heap_region(HeapRegion* hr) { 905 if (_cm->has_aborted()) { 906 return true; 907 } 908 909 uint const region_idx = hr->hrm_index(); 910 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);) 911 assert(top_at_rebuild_start_check == NULL || 912 top_at_rebuild_start_check > hr->bottom(), 913 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)", 914 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str()); 915 916 size_t total_marked_bytes = 0; 917 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize; 918 919 HeapWord* const top_at_mark_start = hr->prev_top_at_mark_start(); 920 921 HeapWord* cur = hr->bottom(); 922 while (cur < hr->end()) { 923 // After every iteration (yield point) we need to check whether the region's 924 // TARS changed due to e.g. eager reclaim. 925 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx); 926 if (top_at_rebuild_start == NULL) { 927 return false; 928 } 929 930 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words)); 931 if (next_chunk.is_empty()) { 932 break; 933 } 934 935 const Ticks start = Ticks::now(); 936 size_t marked_bytes = rebuild_rem_set_in_region(_cm->prev_mark_bitmap(), 937 top_at_mark_start, 938 top_at_rebuild_start, 939 hr, 940 next_chunk); 941 Tickspan time = Ticks::now() - start; 942 943 log_trace(gc, remset, tracking)("Rebuilt region %u " 944 "live " SIZE_FORMAT " " 945 "time %.3fms " 946 "marked bytes " SIZE_FORMAT " " 947 "bot " PTR_FORMAT " " 948 "TAMS " PTR_FORMAT " " 949 "TARS " PTR_FORMAT, 950 region_idx, 951 _cm->liveness(region_idx) * HeapWordSize, 952 TicksToTimeHelper::seconds(time) * 1000.0, 953 marked_bytes, 954 p2i(hr->bottom()), 955 p2i(top_at_mark_start), 956 p2i(top_at_rebuild_start)); 957 958 if (marked_bytes > 0) { 959 total_marked_bytes += marked_bytes; 960 } 961 cur += chunk_size_in_words; 962 963 _cm->do_yield_check(); 964 if (_cm->has_aborted()) { 965 return true; 966 } 967 } 968 // In the final iteration of the loop the region might have been eagerly reclaimed. 969 // Simply filter out those regions. We can not just use region type because there 970 // might have already been new allocations into these regions. 971 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);) 972 assert(top_at_rebuild_start == NULL || 973 total_marked_bytes == hr->marked_bytes(), 974 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match calculated marked bytes " SIZE_FORMAT " " 975 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")", 976 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), hr->marked_bytes(), 977 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start)); 978 // Abort state may have changed after the yield check. 979 return _cm->has_aborted(); 980 } 981 }; 982 983 HeapRegionClaimer _hr_claimer; 984 G1ConcurrentMark* _cm; 985 986 uint _worker_id_offset; 987 public: 988 G1RebuildRemSetTask(G1ConcurrentMark* cm, 989 uint n_workers, 990 uint worker_id_offset) : 991 AbstractGangTask("G1 Rebuild Remembered Set"), 992 _cm(cm), 993 _hr_claimer(n_workers), 994 _worker_id_offset(worker_id_offset) { 995 } 996 997 void work(uint worker_id) { 998 SuspendibleThreadSetJoiner sts_join; 999 1000 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1001 1002 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id); 1003 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id); 1004 } 1005 }; 1006 1007 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm, 1008 WorkGang* workers, 1009 uint worker_id_offset) { 1010 uint num_workers = workers->active_workers(); 1011 1012 G1RebuildRemSetTask cl(cm, 1013 num_workers, 1014 worker_id_offset); 1015 workers->run_task(&cl, num_workers); 1016 }