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/g1RemSet.hpp" 36 #include "gc/g1/heapRegion.inline.hpp" 37 #include "gc/g1/heapRegionManager.inline.hpp" 38 #include "gc/g1/heapRegionRemSet.hpp" 39 #include "gc/shared/gcTraceTime.inline.hpp" 40 #include "gc/shared/suspendibleThreadSet.hpp" 41 #include "memory/iterator.hpp" 42 #include "memory/resourceArea.hpp" 43 #include "oops/access.inline.hpp" 44 #include "oops/oop.inline.hpp" 45 #include "runtime/os.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* g1, 281 G1CardTable* ct, 282 G1HotCardCache* hot_card_cache) : 283 _g1(g1), 284 _scan_state(new G1RemSetScanState()), 285 _num_conc_refined_cards(0), 286 _ct(ct), 287 _g1p(_g1->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 CodeBlobClosure* code_root_cl, 310 uint worker_i) : 311 _scan_state(scan_state), 312 _scan_objs_on_card_cl(scan_obj_on_card), 313 _code_root_cl(code_root_cl), 314 _strong_code_root_scan_time_sec(0.0), 315 _cards_claimed(0), 316 _cards_scanned(0), 317 _cards_skipped(0), 318 _worker_i(worker_i) { 319 _g1h = G1CollectedHeap::heap(); 320 _bot = _g1h->bot(); 321 _ct = _g1h->card_table(); 322 } 323 324 void G1ScanRSForRegionClosure::scan_card(MemRegion mr, uint region_idx_for_card) { 325 HeapRegion* const card_region = _g1h->region_at(region_idx_for_card); 326 _scan_objs_on_card_cl->set_region(card_region); 327 card_region->oops_on_card_seq_iterate_careful<true>(mr, _scan_objs_on_card_cl); 328 _cards_scanned++; 329 } 330 331 void G1ScanRSForRegionClosure::scan_strong_code_roots(HeapRegion* r) { 332 double scan_start = os::elapsedTime(); 333 r->strong_code_roots_do(_code_root_cl); 334 _strong_code_root_scan_time_sec += (os::elapsedTime() - scan_start); 335 } 336 337 void G1ScanRSForRegionClosure::claim_card(size_t card_index, const uint region_idx_for_card){ 338 _ct->set_card_claimed(card_index); 339 _scan_state->add_dirty_region(region_idx_for_card); 340 } 341 342 bool G1ScanRSForRegionClosure::do_heap_region(HeapRegion* r) { 343 assert(r->in_collection_set(), "should only be called on elements of CS."); 344 uint region_idx = r->hrm_index(); 345 346 if (_scan_state->iter_is_complete(region_idx)) { 347 return false; 348 } 349 if (_scan_state->claim_iter(region_idx)) { 350 // If we ever free the collection set concurrently, we should also 351 // clear the card table concurrently therefore we won't need to 352 // add regions of the collection set to the dirty cards region. 353 _scan_state->add_dirty_region(region_idx); 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 // If the card is dirty, then G1 will scan it during Update RS. 374 if (_ct->is_card_claimed(card_index) || _ct->is_card_dirty(card_index)) { 375 continue; 376 } 377 378 HeapWord* const card_start = _g1h->bot()->address_for_index(card_index); 379 uint const region_idx_for_card = _g1h->addr_to_region(card_start); 380 381 assert(_g1h->region_at(region_idx_for_card)->is_in_reserved(card_start), 382 "Card start " PTR_FORMAT " to scan outside of region %u", p2i(card_start), _g1h->region_at(region_idx_for_card)->hrm_index()); 383 HeapWord* const top = _scan_state->scan_top(region_idx_for_card); 384 if (card_start >= top) { 385 continue; 386 } 387 388 // We claim lazily (so races are possible but they're benign), which reduces the 389 // number of duplicate scans (the rsets of the regions in the cset can intersect). 390 // Claim the card after checking bounds above: the remembered set may contain 391 // random cards into current survivor, and we would then have an incorrectly 392 // claimed card in survivor space. Card table clear does not reset the card table 393 // of survivor space regions. 394 claim_card(card_index, region_idx_for_card); 395 396 MemRegion const mr(card_start, MIN2(card_start + BOTConstants::N_words, top)); 397 398 scan_card(mr, region_idx_for_card); 399 } 400 if (_scan_state->set_iter_complete(region_idx)) { 401 // Scan the strong code root list attached to the current region 402 scan_strong_code_roots(r); 403 } 404 return false; 405 } 406 407 void G1RemSet::scan_rem_set(G1ParScanThreadState* pss, 408 CodeBlobClosure* heap_region_codeblobs, 409 uint worker_i) { 410 double rs_time_start = os::elapsedTime(); 411 412 G1ScanObjsDuringScanRSClosure scan_cl(_g1, pss); 413 G1ScanRSForRegionClosure cl(_scan_state, &scan_cl, heap_region_codeblobs, worker_i); 414 _g1->collection_set_iterate_from(&cl, worker_i); 415 416 double scan_rs_time_sec = (os::elapsedTime() - rs_time_start) - 417 cl.strong_code_root_scan_time_sec(); 418 419 G1GCPhaseTimes* p = _g1p->phase_times(); 420 421 p->record_time_secs(G1GCPhaseTimes::ScanRS, worker_i, scan_rs_time_sec); 422 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_scanned(), G1GCPhaseTimes::ScanRSScannedCards); 423 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_claimed(), G1GCPhaseTimes::ScanRSClaimedCards); 424 p->record_thread_work_item(G1GCPhaseTimes::ScanRS, worker_i, cl.cards_skipped(), G1GCPhaseTimes::ScanRSSkippedCards); 425 426 p->record_time_secs(G1GCPhaseTimes::CodeRoots, worker_i, cl.strong_code_root_scan_time_sec()); 427 } 428 429 // Closure used for updating rem sets. Only called during an evacuation pause. 430 class G1RefineCardClosure: public CardTableEntryClosure { 431 G1RemSet* _g1rs; 432 G1ScanObjsDuringUpdateRSClosure* _update_rs_cl; 433 434 size_t _cards_scanned; 435 size_t _cards_skipped; 436 public: 437 G1RefineCardClosure(G1CollectedHeap* g1h, G1ScanObjsDuringUpdateRSClosure* update_rs_cl) : 438 _g1rs(g1h->g1_rem_set()), _update_rs_cl(update_rs_cl), _cards_scanned(0), _cards_skipped(0) 439 {} 440 441 bool do_card_ptr(jbyte* card_ptr, uint worker_i) { 442 // The only time we care about recording cards that 443 // contain references that point into the collection set 444 // is during RSet updating within an evacuation pause. 445 // In this case worker_i should be the id of a GC worker thread. 446 assert(SafepointSynchronize::is_at_safepoint(), "not during an evacuation pause"); 447 448 bool card_scanned = _g1rs->refine_card_during_gc(card_ptr, _update_rs_cl); 449 450 if (card_scanned) { 451 _cards_scanned++; 452 } else { 453 _cards_skipped++; 454 } 455 return true; 456 } 457 458 size_t cards_scanned() const { return _cards_scanned; } 459 size_t cards_skipped() const { return _cards_skipped; } 460 }; 461 462 void G1RemSet::update_rem_set(G1ParScanThreadState* pss, uint worker_i) { 463 G1ScanObjsDuringUpdateRSClosure update_rs_cl(_g1, pss, worker_i); 464 G1RefineCardClosure refine_card_cl(_g1, &update_rs_cl); 465 466 G1GCParPhaseTimesTracker x(_g1p->phase_times(), G1GCPhaseTimes::UpdateRS, worker_i); 467 if (G1HotCardCache::default_use_cache()) { 468 // Apply the closure to the entries of the hot card cache. 469 G1GCParPhaseTimesTracker y(_g1p->phase_times(), G1GCPhaseTimes::ScanHCC, worker_i); 470 _g1->iterate_hcc_closure(&refine_card_cl, worker_i); 471 } 472 // Apply the closure to all remaining log entries. 473 _g1->iterate_dirty_card_closure(&refine_card_cl, worker_i); 474 475 G1GCPhaseTimes* p = _g1p->phase_times(); 476 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_scanned(), G1GCPhaseTimes::UpdateRSScannedCards); 477 p->record_thread_work_item(G1GCPhaseTimes::UpdateRS, worker_i, refine_card_cl.cards_skipped(), G1GCPhaseTimes::UpdateRSSkippedCards); 478 } 479 480 void G1RemSet::cleanupHRRS() { 481 HeapRegionRemSet::cleanup(); 482 } 483 484 void G1RemSet::oops_into_collection_set_do(G1ParScanThreadState* pss, 485 CodeBlobClosure* heap_region_codeblobs, 486 uint worker_i) { 487 update_rem_set(pss, worker_i); 488 scan_rem_set(pss, heap_region_codeblobs, worker_i);; 489 } 490 491 void G1RemSet::prepare_for_oops_into_collection_set_do() { 492 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 493 dcqs.concatenate_logs(); 494 495 _scan_state->reset(); 496 } 497 498 void G1RemSet::cleanup_after_oops_into_collection_set_do() { 499 G1GCPhaseTimes* phase_times = _g1->g1_policy()->phase_times(); 500 501 // Set all cards back to clean. 502 double start = os::elapsedTime(); 503 _scan_state->clear_card_table(_g1->workers()); 504 phase_times->record_clear_ct_time((os::elapsedTime() - start) * 1000.0); 505 } 506 507 inline void check_card_ptr(jbyte* card_ptr, G1CardTable* ct) { 508 #ifdef ASSERT 509 G1CollectedHeap* g1 = G1CollectedHeap::heap(); 510 assert(g1->is_in_exact(ct->addr_for(card_ptr)), 511 "Card at " PTR_FORMAT " index " SIZE_FORMAT " representing heap at " PTR_FORMAT " (%u) must be in committed heap", 512 p2i(card_ptr), 513 ct->index_for(ct->addr_for(card_ptr)), 514 p2i(ct->addr_for(card_ptr)), 515 g1->addr_to_region(ct->addr_for(card_ptr))); 516 #endif 517 } 518 519 void G1RemSet::refine_card_concurrently(jbyte* card_ptr, 520 uint worker_i) { 521 assert(!_g1->is_gc_active(), "Only call concurrently"); 522 523 check_card_ptr(card_ptr, _ct); 524 525 // If the card is no longer dirty, nothing to do. 526 if (*card_ptr != G1CardTable::dirty_card_val()) { 527 return; 528 } 529 530 // Construct the region representing the card. 531 HeapWord* start = _ct->addr_for(card_ptr); 532 // And find the region containing it. 533 HeapRegion* r = _g1->heap_region_containing(start); 534 535 // This check is needed for some uncommon cases where we should 536 // ignore the card. 537 // 538 // The region could be young. Cards for young regions are 539 // distinctly marked (set to g1_young_gen), so the post-barrier will 540 // filter them out. However, that marking is performed 541 // concurrently. A write to a young object could occur before the 542 // card has been marked young, slipping past the filter. 543 // 544 // The card could be stale, because the region has been freed since 545 // the card was recorded. In this case the region type could be 546 // anything. If (still) free or (reallocated) young, just ignore 547 // it. If (reallocated) old or humongous, the later card trimming 548 // and additional checks in iteration may detect staleness. At 549 // worst, we end up processing a stale card unnecessarily. 550 // 551 // In the normal (non-stale) case, the synchronization between the 552 // enqueueing of the card and processing it here will have ensured 553 // we see the up-to-date region type here. 554 if (!r->is_old_or_humongous()) { 555 return; 556 } 557 558 // The result from the hot card cache insert call is either: 559 // * pointer to the current card 560 // (implying that the current card is not 'hot'), 561 // * null 562 // (meaning we had inserted the card ptr into the "hot" card cache, 563 // which had some headroom), 564 // * a pointer to a "hot" card that was evicted from the "hot" cache. 565 // 566 567 if (_hot_card_cache->use_cache()) { 568 assert(!SafepointSynchronize::is_at_safepoint(), "sanity"); 569 570 const jbyte* orig_card_ptr = card_ptr; 571 card_ptr = _hot_card_cache->insert(card_ptr); 572 if (card_ptr == NULL) { 573 // There was no eviction. Nothing to do. 574 return; 575 } else if (card_ptr != orig_card_ptr) { 576 // Original card was inserted and an old card was evicted. 577 start = _ct->addr_for(card_ptr); 578 r = _g1->heap_region_containing(start); 579 580 // Check whether the region formerly in the cache should be 581 // ignored, as discussed earlier for the original card. The 582 // region could have been freed while in the cache. 583 if (!r->is_old_or_humongous()) { 584 return; 585 } 586 } // Else we still have the original card. 587 } 588 589 // Trim the region designated by the card to what's been allocated 590 // in the region. The card could be stale, or the card could cover 591 // (part of) an object at the end of the allocated space and extend 592 // beyond the end of allocation. 593 594 // Non-humongous objects are only allocated in the old-gen during 595 // GC, so if region is old then top is stable. Humongous object 596 // allocation sets top last; if top has not yet been set, this is 597 // a stale card and we'll end up with an empty intersection. If 598 // this is not a stale card, the synchronization between the 599 // enqueuing of the card and processing it here will have ensured 600 // we see the up-to-date top here. 601 HeapWord* scan_limit = r->top(); 602 603 if (scan_limit <= start) { 604 // If the trimmed region is empty, the card must be stale. 605 return; 606 } 607 608 // Okay to clean and process the card now. There are still some 609 // stale card cases that may be detected by iteration and dealt with 610 // as iteration failure. 611 *const_cast<volatile jbyte*>(card_ptr) = G1CardTable::clean_card_val(); 612 613 // This fence serves two purposes. First, the card must be cleaned 614 // before processing the contents. Second, we can't proceed with 615 // processing until after the read of top, for synchronization with 616 // possibly concurrent humongous object allocation. It's okay that 617 // reading top and reading type were racy wrto each other. We need 618 // both set, in any order, to proceed. 619 OrderAccess::fence(); 620 621 // Don't use addr_for(card_ptr + 1) which can ask for 622 // a card beyond the heap. 623 HeapWord* end = start + G1CardTable::card_size_in_words; 624 MemRegion dirty_region(start, MIN2(scan_limit, end)); 625 assert(!dirty_region.is_empty(), "sanity"); 626 627 G1ConcurrentRefineOopClosure conc_refine_cl(_g1, worker_i); 628 629 bool card_processed = 630 r->oops_on_card_seq_iterate_careful<false>(dirty_region, &conc_refine_cl); 631 632 // If unable to process the card then we encountered an unparsable 633 // part of the heap (e.g. a partially allocated object) while 634 // processing a stale card. Despite the card being stale, redirty 635 // and re-enqueue, because we've already cleaned the card. Without 636 // this we could incorrectly discard a non-stale card. 637 if (!card_processed) { 638 // The card might have gotten re-dirtied and re-enqueued while we 639 // worked. (In fact, it's pretty likely.) 640 if (*card_ptr != G1CardTable::dirty_card_val()) { 641 *card_ptr = G1CardTable::dirty_card_val(); 642 MutexLockerEx x(Shared_DirtyCardQ_lock, 643 Mutex::_no_safepoint_check_flag); 644 DirtyCardQueue* sdcq = 645 JavaThread::dirty_card_queue_set().shared_dirty_card_queue(); 646 sdcq->enqueue(card_ptr); 647 } 648 } else { 649 _num_conc_refined_cards++; // Unsynchronized update, only used for logging. 650 } 651 } 652 653 bool G1RemSet::refine_card_during_gc(jbyte* card_ptr, 654 G1ScanObjsDuringUpdateRSClosure* update_rs_cl) { 655 assert(_g1->is_gc_active(), "Only call during GC"); 656 657 check_card_ptr(card_ptr, _ct); 658 659 // If the card is no longer dirty, nothing to do. This covers cards that were already 660 // scanned as parts of the remembered sets. 661 if (*card_ptr != G1CardTable::dirty_card_val()) { 662 return false; 663 } 664 665 // We claim lazily (so races are possible but they're benign), which reduces the 666 // number of potential duplicate scans (multiple threads may enqueue the same card twice). 667 *card_ptr = G1CardTable::clean_card_val() | G1CardTable::claimed_card_val(); 668 669 // Construct the region representing the card. 670 HeapWord* card_start = _ct->addr_for(card_ptr); 671 // And find the region containing it. 672 uint const card_region_idx = _g1->addr_to_region(card_start); 673 674 _scan_state->add_dirty_region(card_region_idx); 675 HeapWord* scan_limit = _scan_state->scan_top(card_region_idx); 676 if (scan_limit <= card_start) { 677 // If the card starts above the area in the region containing objects to scan, skip it. 678 return false; 679 } 680 681 // Don't use addr_for(card_ptr + 1) which can ask for 682 // a card beyond the heap. 683 HeapWord* card_end = card_start + G1CardTable::card_size_in_words; 684 MemRegion dirty_region(card_start, MIN2(scan_limit, card_end)); 685 assert(!dirty_region.is_empty(), "sanity"); 686 687 HeapRegion* const card_region = _g1->region_at(card_region_idx); 688 update_rs_cl->set_region(card_region); 689 bool card_processed = card_region->oops_on_card_seq_iterate_careful<true>(dirty_region, update_rs_cl); 690 assert(card_processed, "must be"); 691 return true; 692 } 693 694 void G1RemSet::print_periodic_summary_info(const char* header, uint period_count) { 695 if ((G1SummarizeRSetStatsPeriod > 0) && log_is_enabled(Trace, gc, remset) && 696 (period_count % G1SummarizeRSetStatsPeriod == 0)) { 697 698 G1RemSetSummary current(this); 699 _prev_period_summary.subtract_from(¤t); 700 701 Log(gc, remset) log; 702 log.trace("%s", header); 703 ResourceMark rm; 704 LogStream ls(log.trace()); 705 _prev_period_summary.print_on(&ls); 706 707 _prev_period_summary.set(¤t); 708 } 709 } 710 711 void G1RemSet::print_summary_info() { 712 Log(gc, remset, exit) log; 713 if (log.is_trace()) { 714 log.trace(" Cumulative RS summary"); 715 G1RemSetSummary current(this); 716 ResourceMark rm; 717 LogStream ls(log.trace()); 718 current.print_on(&ls); 719 } 720 } 721 722 class G1RebuildRemSetTask: public AbstractGangTask { 723 // Aggregate the counting data that was constructed concurrently 724 // with marking. 725 class G1RebuildRemSetHeapRegionClosure : public HeapRegionClosure { 726 G1ConcurrentMark* _cm; 727 G1RebuildRemSetClosure _update_cl; 728 729 // Applies _update_cl to the references of the given object, limiting objArrays 730 // to the given MemRegion. Returns the amount of words actually scanned. 731 size_t scan_for_references(oop const obj, MemRegion mr) { 732 size_t const obj_size = obj->size(); 733 // All non-objArrays and objArrays completely within the mr 734 // can be scanned without passing the mr. 735 if (!obj->is_objArray() || mr.contains(MemRegion((HeapWord*)obj, obj_size))) { 736 obj->oop_iterate(&_update_cl); 737 return obj_size; 738 } 739 // This path is for objArrays crossing the given MemRegion. Only scan the 740 // area within the MemRegion. 741 obj->oop_iterate(&_update_cl, mr); 742 return mr.intersection(MemRegion((HeapWord*)obj, obj_size)).word_size(); 743 } 744 745 // A humongous object is live (with respect to the scanning) either 746 // a) it is marked on the bitmap as such 747 // b) its TARS is larger than TAMS, i.e. has been allocated during marking. 748 bool is_humongous_live(oop const humongous_obj, const G1CMBitMap* const bitmap, HeapWord* tams, HeapWord* tars) const { 749 return bitmap->is_marked(humongous_obj) || (tars > tams); 750 } 751 752 // Iterator over the live objects within the given MemRegion. 753 class LiveObjIterator : public StackObj { 754 const G1CMBitMap* const _bitmap; 755 const HeapWord* _tams; 756 const MemRegion _mr; 757 HeapWord* _current; 758 759 bool is_below_tams() const { 760 return _current < _tams; 761 } 762 763 bool is_live(HeapWord* obj) const { 764 return !is_below_tams() || _bitmap->is_marked(obj); 765 } 766 767 HeapWord* bitmap_limit() const { 768 return MIN2(const_cast<HeapWord*>(_tams), _mr.end()); 769 } 770 771 void move_if_below_tams() { 772 if (is_below_tams() && has_next()) { 773 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 774 } 775 } 776 public: 777 LiveObjIterator(const G1CMBitMap* const bitmap, const HeapWord* tams, const MemRegion mr, HeapWord* first_oop_into_mr) : 778 _bitmap(bitmap), 779 _tams(tams), 780 _mr(mr), 781 _current(first_oop_into_mr) { 782 783 assert(_current <= _mr.start(), 784 "First oop " PTR_FORMAT " should extend into mr [" PTR_FORMAT ", " PTR_FORMAT ")", 785 p2i(first_oop_into_mr), p2i(mr.start()), p2i(mr.end())); 786 787 // Step to the next live object within the MemRegion if needed. 788 if (is_live(_current)) { 789 // Non-objArrays were scanned by the previous part of that region. 790 if (_current < mr.start() && !oop(_current)->is_objArray()) { 791 _current += oop(_current)->size(); 792 // We might have positioned _current on a non-live object. Reposition to the next 793 // live one if needed. 794 move_if_below_tams(); 795 } 796 } else { 797 // The object at _current can only be dead if below TAMS, so we can use the bitmap. 798 // immediately. 799 _current = _bitmap->get_next_marked_addr(_current, bitmap_limit()); 800 assert(_current == _mr.end() || is_live(_current), 801 "Current " PTR_FORMAT " should be live (%s) or beyond the end of the MemRegion (" PTR_FORMAT ")", 802 p2i(_current), BOOL_TO_STR(is_live(_current)), p2i(_mr.end())); 803 } 804 } 805 806 void move_to_next() { 807 _current += next()->size(); 808 move_if_below_tams(); 809 } 810 811 oop next() const { 812 oop result = oop(_current); 813 assert(is_live(_current), 814 "Object " PTR_FORMAT " must be live TAMS " PTR_FORMAT " below %d mr " PTR_FORMAT " " PTR_FORMAT " outside %d", 815 p2i(_current), p2i(_tams), _tams > _current, p2i(_mr.start()), p2i(_mr.end()), _mr.contains(result)); 816 return result; 817 } 818 819 bool has_next() const { 820 return _current < _mr.end(); 821 } 822 }; 823 824 // Rebuild remembered sets in the part of the region specified by mr and hr. 825 // Objects between the bottom of the region and the TAMS are checked for liveness 826 // using the given bitmap. Objects between TAMS and TARS are assumed to be live. 827 // Returns the number of live words between bottom and TAMS. 828 size_t rebuild_rem_set_in_region(const G1CMBitMap* const bitmap, 829 HeapWord* const top_at_mark_start, 830 HeapWord* const top_at_rebuild_start, 831 HeapRegion* hr, 832 MemRegion mr) { 833 size_t marked_words = 0; 834 835 if (hr->is_humongous()) { 836 oop const humongous_obj = oop(hr->humongous_start_region()->bottom()); 837 if (is_humongous_live(humongous_obj, bitmap, top_at_mark_start, top_at_rebuild_start)) { 838 // We need to scan both [bottom, TAMS) and [TAMS, top_at_rebuild_start); 839 // however in case of humongous objects it is sufficient to scan the encompassing 840 // area (top_at_rebuild_start is always larger or equal to TAMS) as one of the 841 // two areas will be zero sized. I.e. TAMS is either 842 // the same as bottom or top(_at_rebuild_start). There is no way TAMS has a different 843 // value: this would mean that TAMS points somewhere into the object. 844 assert(hr->top() == top_at_mark_start || hr->top() == top_at_rebuild_start, 845 "More than one object in the humongous region?"); 846 humongous_obj->oop_iterate(&_update_cl, mr); 847 return top_at_mark_start != hr->bottom() ? mr.byte_size() : 0; 848 } else { 849 return 0; 850 } 851 } 852 853 for (LiveObjIterator it(bitmap, top_at_mark_start, mr, hr->block_start(mr.start())); it.has_next(); it.move_to_next()) { 854 oop obj = it.next(); 855 size_t scanned_size = scan_for_references(obj, mr); 856 if ((HeapWord*)obj < top_at_mark_start) { 857 marked_words += scanned_size; 858 } 859 } 860 861 return marked_words * HeapWordSize; 862 } 863 public: 864 G1RebuildRemSetHeapRegionClosure(G1CollectedHeap* g1h, 865 G1ConcurrentMark* cm, 866 uint worker_id) : 867 HeapRegionClosure(), 868 _cm(cm), 869 _update_cl(g1h, worker_id) { } 870 871 bool do_heap_region(HeapRegion* hr) { 872 if (_cm->has_aborted()) { 873 return true; 874 } 875 876 uint const region_idx = hr->hrm_index(); 877 DEBUG_ONLY(HeapWord* const top_at_rebuild_start_check = _cm->top_at_rebuild_start(region_idx);) 878 assert(top_at_rebuild_start_check == NULL || 879 top_at_rebuild_start_check > hr->bottom(), 880 "A TARS (" PTR_FORMAT ") == bottom() (" PTR_FORMAT ") indicates the old region %u is empty (%s)", 881 p2i(top_at_rebuild_start_check), p2i(hr->bottom()), region_idx, hr->get_type_str()); 882 883 size_t total_marked_bytes = 0; 884 size_t const chunk_size_in_words = G1RebuildRemSetChunkSize / HeapWordSize; 885 886 HeapWord* const top_at_mark_start = hr->next_top_at_mark_start(); 887 888 HeapWord* cur = hr->bottom(); 889 while (cur < hr->end()) { 890 // After every iteration (yield point) we need to check whether the region's 891 // TARS changed due to e.g. eager reclaim. 892 HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx); 893 if (top_at_rebuild_start == NULL) { 894 return false; 895 } 896 897 MemRegion next_chunk = MemRegion(hr->bottom(), top_at_rebuild_start).intersection(MemRegion(cur, chunk_size_in_words)); 898 if (next_chunk.is_empty()) { 899 break; 900 } 901 902 const Ticks start = Ticks::now(); 903 size_t marked_bytes = rebuild_rem_set_in_region(_cm->next_mark_bitmap(), 904 top_at_mark_start, 905 top_at_rebuild_start, 906 hr, 907 next_chunk); 908 Tickspan time = Ticks::now() - start; 909 910 log_trace(gc, remset, tracking)("Rebuilt region %u " 911 "live " SIZE_FORMAT " " 912 "time %.3fms " 913 "marked bytes " SIZE_FORMAT " " 914 "bot " PTR_FORMAT " " 915 "TAMS " PTR_FORMAT " " 916 "TARS " PTR_FORMAT, 917 region_idx, 918 _cm->liveness(region_idx) * HeapWordSize, 919 (time.value() / os::elapsed_frequency()) * 1000.0, 920 marked_bytes, 921 p2i(hr->bottom()), 922 p2i(top_at_mark_start), 923 p2i(top_at_rebuild_start)); 924 925 if (marked_bytes > 0) { 926 hr->add_to_marked_bytes(marked_bytes); 927 total_marked_bytes += marked_bytes; 928 } 929 cur += chunk_size_in_words; 930 931 _cm->do_yield_check(); 932 if (_cm->has_aborted()) { 933 return true; 934 } 935 } 936 // In the final iteration of the loop the region might have been eagerly reclaimed. 937 // Simply filter out those regions. We can not just use region type because there 938 // might have already been new allocations into these regions. 939 DEBUG_ONLY(HeapWord* const top_at_rebuild_start = _cm->top_at_rebuild_start(region_idx);) 940 assert(!hr->is_old() || 941 top_at_rebuild_start == NULL || 942 total_marked_bytes == _cm->liveness(region_idx) * HeapWordSize, 943 "Marked bytes " SIZE_FORMAT " for region %u (%s) in [bottom, TAMS) do not match liveness during mark " SIZE_FORMAT " " 944 "(" PTR_FORMAT " " PTR_FORMAT " " PTR_FORMAT ")", 945 total_marked_bytes, hr->hrm_index(), hr->get_type_str(), _cm->liveness(region_idx) * HeapWordSize, 946 p2i(hr->bottom()), p2i(top_at_mark_start), p2i(top_at_rebuild_start)); 947 // Abort state may have changed after the yield check. 948 return _cm->has_aborted(); 949 } 950 }; 951 952 HeapRegionClaimer _hr_claimer; 953 G1ConcurrentMark* _cm; 954 955 uint _worker_id_offset; 956 public: 957 G1RebuildRemSetTask(G1ConcurrentMark* cm, 958 uint n_workers, 959 uint worker_id_offset) : 960 AbstractGangTask("G1 Rebuild Remembered Set"), 961 _cm(cm), 962 _hr_claimer(n_workers), 963 _worker_id_offset(worker_id_offset) { 964 } 965 966 void work(uint worker_id) { 967 SuspendibleThreadSetJoiner sts_join; 968 969 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 970 971 G1RebuildRemSetHeapRegionClosure cl(g1h, _cm, _worker_id_offset + worker_id); 972 g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hr_claimer, worker_id); 973 } 974 }; 975 976 void G1RemSet::rebuild_rem_set(G1ConcurrentMark* cm, 977 WorkGang* workers, 978 uint worker_id_offset) { 979 uint num_workers = workers->active_workers(); 980 981 G1RebuildRemSetTask cl(cm, 982 num_workers, 983 worker_id_offset); 984 workers->run_task(&cl, num_workers); 985 }