1 /* 2 * Copyright Amazon.com Inc. 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 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHSCANREMEMBERED_HPP 26 #define SHARE_GC_SHENANDOAH_SHENANDOAHSCANREMEMBERED_HPP 27 28 // Terminology used within this source file: 29 // 30 // Card Entry: This is the information that identifies whether a 31 // particular card-table entry is Clean or Dirty. A clean 32 // card entry denotes that the associated memory does not 33 // hold references to young-gen memory. 34 // 35 // Card Region, aka 36 // Card Memory: This is the region of memory that is assocated with a 37 // particular card entry. 38 // 39 // Card Cluster: A card cluster represents 64 card entries. A card 40 // cluster is the minimal amount of work performed at a 41 // time by a parallel thread. Note that the work required 42 // to scan a card cluster is somewhat variable in that the 43 // required effort depends on how many cards are dirty, how 44 // many references are held within the objects that span a 45 // DIRTY card's memory, and on the size of the object 46 // that spans the end of a DIRTY card's memory (because 47 // that object, if it's not an array, may need to be scanned in 48 // its entirety, when the object is imprecisely dirtied. Imprecise 49 // dirtying is when the card corresponding to the object header 50 // is dirtied, rather than the card on which the updated field lives). 51 // To better balance work amongst them, parallel worker threads dynamically 52 // claim clusters and are flexible in the number of clusters they 53 // process. 54 // 55 // A cluster represents a "natural" quantum of work to be performed by 56 // a parallel GC thread's background remembered set scanning efforts. 57 // The notion of cluster is similar to the notion of stripe in the 58 // implementation of parallel GC card scanning. However, a cluster is 59 // typically smaller than a stripe, enabling finer grain division of 60 // labor between multiple threads, and potentially better load balancing 61 // when dirty cards are not uniformly distributed in the heap, as is often 62 // the case with generational workloads where more recently promoted objects 63 // may be dirtied more frequently that older objects. 64 // 65 // For illustration, consider the following possible JVM configurations: 66 // 67 // Scenario 1: 68 // RegionSize is 128 MB 69 // Span of a card entry is 512 B 70 // Each card table entry consumes 1 B 71 // Assume one long word (8 B)of the card table represents a cluster. 72 // This long word holds 8 card table entries, spanning a 73 // total of 8*512 B = 4 KB of the heap 74 // The number of clusters per region is 128 MB / 4 KB = 32 K 75 // 76 // Scenario 2: 77 // RegionSize is 128 MB 78 // Span of each card entry is 128 B 79 // Each card table entry consumes 1 bit 80 // Assume one int word (4 B) of the card table represents a cluster. 81 // This int word holds 32 b/1 b = 32 card table entries, spanning a 82 // total of 32 * 128 B = 4 KB of the heap 83 // The number of clusters per region is 128 MB / 4 KB = 32 K 84 // 85 // Scenario 3: 86 // RegionSize is 128 MB 87 // Span of each card entry is 512 B 88 // Each card table entry consumes 1 bit 89 // Assume one long word (8 B) of card table represents a cluster. 90 // This long word holds 64 b/ 1 b = 64 card table entries, spanning a 91 // total of 64 * 512 B = 32 KB of the heap 92 // The number of clusters per region is 128 MB / 32 KB = 4 K 93 // 94 // At the start of a new young-gen concurrent mark pass, the gang of 95 // Shenandoah worker threads collaborate in performing the following 96 // actions: 97 // 98 // Let old_regions = number of ShenandoahHeapRegion comprising 99 // old-gen memory 100 // Let region_size = ShenandoahHeapRegion::region_size_bytes() 101 // represent the number of bytes in each region 102 // Let clusters_per_region = region_size / 512 103 // Let rs represent the relevant RememberedSet implementation 104 // (an instance of ShenandoahDirectCardMarkRememberedSet or an instance 105 // of a to-be-implemented ShenandoahBufferWithSATBRememberedSet) 106 // 107 // for each ShenandoahHeapRegion old_region in the whole heap 108 // determine the cluster number of the first cluster belonging 109 // to that region 110 // for each cluster contained within that region 111 // Assure that exactly one worker thread processes each 112 // cluster, each thread making a series of invocations of the 113 // following: 114 // 115 // rs->process_clusters(worker_id, ReferenceProcessor *, 116 // ShenandoahConcurrentMark *, cluster_no, cluster_count, 117 // HeapWord *end_of_range, OopClosure *oops); 118 // 119 // For efficiency, divide up the clusters so that different threads 120 // are responsible for processing different clusters. Processing costs 121 // may vary greatly between clusters for the following reasons: 122 // 123 // a) some clusters contain mostly dirty cards and other 124 // clusters contain mostly clean cards 125 // b) some clusters contain mostly primitive data and other 126 // clusters contain mostly reference data 127 // c) some clusters are spanned by very large non-array objects that 128 // begin in some other cluster. When a large non-array object 129 // beginning in a preceding cluster spans large portions of 130 // this cluster, then because of imprecise dirtying, the 131 // portion of the object in this cluster may be clean, but 132 // will need to be processed by the worker responsible for 133 // this cluster, potentially increasing its work. 134 // d) in the case that the end of this cluster is spanned by a 135 // very large non-array object, the worker for this cluster will 136 // be responsible for processing the portion of the object 137 // in this cluster. 138 // 139 // Though an initial division of labor between marking threads may 140 // assign equal numbers of clusters to be scanned by each thread, it 141 // should be expected that some threads will finish their assigned 142 // work before others. Therefore, some amount of the full remembered 143 // set scanning effort should be held back and assigned incrementally 144 // to the threads that end up with excess capacity. Consider the 145 // following strategy for dividing labor: 146 // 147 // 1. Assume there are 8 marking threads and 1024 remembered 148 // set clusters to be scanned. 149 // 2. Assign each thread to scan 64 clusters. This leaves 150 // 512 (1024 - (8*64)) clusters to still be scanned. 151 // 3. As the 8 server threads complete previous cluster 152 // scanning assignments, issue each of the next 8 scanning 153 // assignments as units of 32 additional cluster each. 154 // In the case that there is high variance in effort 155 // associated with previous cluster scanning assignments, 156 // multiples of these next assignments may be serviced by 157 // the server threads that were previously assigned lighter 158 // workloads. 159 // 4. Make subsequent scanning assignments as follows: 160 // a) 8 assignments of size 16 clusters 161 // b) 8 assignments of size 8 clusters 162 // c) 16 assignments of size 4 clusters 163 // 164 // When there is no more remembered set processing work to be 165 // assigned to a newly idled worker thread, that thread can move 166 // on to work on other tasks associated with root scanning until such 167 // time as all clusters have been examined. 168 // 169 // Remembered set scanning is designed to run concurrently with 170 // mutator threads, with multiple concurrent workers. Furthermore, the 171 // current implementation of remembered set scanning never clears a 172 // card once it has been marked. 173 // 174 // These limitations will be addressed in future enhancements to the 175 // existing implementation. 176 177 #include "gc/shared/workerThread.hpp" 178 #include "gc/shenandoah/shenandoahCardStats.hpp" 179 #include "gc/shenandoah/shenandoahCardTable.hpp" 180 #include "gc/shenandoah/shenandoahNumberSeq.hpp" 181 #include "gc/shenandoah/shenandoahTaskqueue.hpp" 182 #include "memory/iterator.hpp" 183 #include "utilities/globalDefinitions.hpp" 184 185 class ShenandoahReferenceProcessor; 186 class ShenandoahConcurrentMark; 187 class ShenandoahHeap; 188 class ShenandoahHeapRegion; 189 class ShenandoahRegionIterator; 190 class ShenandoahMarkingContext; 191 192 class CardTable; 193 typedef CardTable::CardValue CardValue; 194 195 class ShenandoahDirectCardMarkRememberedSet: public CHeapObj<mtGC> { 196 197 private: 198 199 // Use symbolic constants defined in cardTable.hpp 200 // CardTable::card_shift = 9; 201 // CardTable::card_size = 512; 202 // CardTable::card_size_in_words = 64; 203 // CardTable::clean_card_val() 204 // CardTable::dirty_card_val() 205 206 const size_t LogCardValsPerIntPtr; // the number of card values (entries) in an intptr_t 207 const size_t LogCardSizeInWords; // the size of a card in heap word units 208 209 ShenandoahHeap *_heap; 210 ShenandoahCardTable *_card_table; 211 size_t _card_shift; 212 size_t _total_card_count; 213 size_t _cluster_count; 214 HeapWord *_whole_heap_base; // Points to first HeapWord of data contained within heap memory 215 CardValue* _byte_map; // Points to first entry within the card table 216 CardValue* _byte_map_base; // Points to byte_map minus the bias computed from address of heap memory 217 218 public: 219 220 // count is the number of cards represented by the card table. 221 ShenandoahDirectCardMarkRememberedSet(ShenandoahCardTable *card_table, size_t total_card_count); 222 223 // Card index is zero-based relative to _byte_map. 224 size_t last_valid_index() const; 225 size_t total_cards() const; 226 size_t card_index_for_addr(HeapWord *p) const; 227 HeapWord *addr_for_card_index(size_t card_index) const; 228 inline const CardValue* get_card_table_byte_map(bool write_table) const; 229 inline bool is_card_dirty(size_t card_index) const; 230 inline bool is_write_card_dirty(size_t card_index) const; 231 inline void mark_card_as_dirty(size_t card_index); 232 inline void mark_range_as_dirty(size_t card_index, size_t num_cards); 233 inline void mark_card_as_clean(size_t card_index); 234 inline void mark_range_as_clean(size_t card_index, size_t num_cards); 235 inline bool is_card_dirty(HeapWord *p) const; 236 inline void mark_card_as_dirty(HeapWord *p); 237 inline void mark_range_as_dirty(HeapWord *p, size_t num_heap_words); 238 inline void mark_card_as_clean(HeapWord *p); 239 inline void mark_range_as_clean(HeapWord *p, size_t num_heap_words); 240 inline size_t cluster_count() const; 241 242 // Called by GC thread at start of concurrent mark to exchange roles of read and write remembered sets. 243 // Not currently used because mutator write barrier does not honor changes to the location of card table. 244 // Instead of swap_remset, the current implementation of concurrent remembered set scanning does reset_remset 245 // in parallel threads, each invocation processing one entire HeapRegion at a time. 246 void swap_remset() { _card_table->swap_card_tables(); } 247 248 // Merge any dirty values from write table into the read table, while leaving 249 // the write table unchanged. 250 void merge_write_table(HeapWord* start, size_t word_count); 251 252 // Destructively copy the write table to the read table, and clean the write table. 253 void reset_remset(HeapWord* start, size_t word_count); 254 255 // Called by GC thread after scanning old remembered set in order to prepare for next GC pass 256 void clear_old_remset() { _card_table->clear_read_table(); } 257 }; 258 259 // A ShenandoahCardCluster represents the minimal unit of work 260 // performed by independent parallel GC threads during scanning of 261 // remembered sets. 262 // 263 // The GC threads that perform card-table remembered set scanning may 264 // overwrite card-table entries to mark them as clean in the case that 265 // the associated memory no longer holds references to young-gen 266 // memory. Rather than access the card-table entries directly, all GC 267 // thread access to card-table information is made by way of the 268 // ShenandoahCardCluster data abstraction. This abstraction 269 // effectively manages access to multiple possible underlying 270 // remembered set implementations, including a traditional card-table 271 // approach and a SATB-based approach. 272 // 273 // The API services represent a compromise between efficiency and 274 // convenience. 275 // 276 // Multiple GC threads that scan the remembered set 277 // in parallel. The desire is to divide the complete scanning effort 278 // into multiple clusters of work that can be independently processed 279 // by individual threads without need for synchronizing efforts 280 // between the work performed by each task. The term "cluster" of 281 // work is similar to the term "stripe" as used in the implementation 282 // of Parallel GC. 283 // 284 // Complexity arises when an object to be scanned crosses the boundary 285 // between adjacent cluster regions. Here is the protocol that we currently 286 // follow: 287 // 288 // 1. The thread responsible for scanning the cards in a cluster modifies 289 // the associated card-table entries. Only cards that are dirty are 290 // processed, except as described below for the case of objects that 291 // straddle more than one card. 292 // 2. Object Arrays are precisely dirtied, so only the portion of the obj-array 293 // that overlaps the range of dirty cards in its cluster are scanned 294 // by each worker thread. This holds for portions of obj-arrays that extend 295 // over clusters processed by different workers, with each worked responsible 296 // for scanning the portion of the obj-array overlapping the dirty cards in 297 // its cluster. 298 // 3. Non-array objects are precisely dirtied by the interpreter and the compilers 299 // For such objects that extend over multiple cards, or even multiple clusters, 300 // the entire object is scanned by the worker that processes the (dirty) card on 301 // which the object's header lies. (However, GC workers should precisely dirty the 302 // cards with inter-regional/inter-generational pointers in the body of this object, 303 // thus making subsequent scans potentially less expensive.) Such larger non-array 304 // objects are relatively rare. 305 // 306 // A possible criticism: 307 // C. The representation of pointer location descriptive information 308 // within Klass representations is not designed for efficient 309 // "random access". An alternative approach to this design would 310 // be to scan very large objects multiple times, once for each 311 // cluster that is spanned by the object's range. This reduces 312 // unnecessary overscan, but it introduces different sorts of 313 // overhead effort: 314 // i) For each spanned cluster, we have to look up the start of 315 // the crossing object. 316 // ii) Each time we scan the very large object, we have to 317 // sequentially walk through its pointer location 318 // descriptors, skipping over all of the pointers that 319 // precede the start of the range of addresses that we 320 // consider relevant. 321 322 323 // Because old-gen heap memory is not necessarily contiguous, and 324 // because cards are not necessarily maintained for young-gen memory, 325 // consecutive card numbers do not necessarily correspond to consecutive 326 // address ranges. For the traditional direct-card-marking 327 // implementation of this interface, consecutive card numbers are 328 // likely to correspond to contiguous regions of memory, but this 329 // should not be assumed. Instead, rely only upon the following: 330 // 331 // 1. All card numbers for cards pertaining to the same 332 // ShenandoahHeapRegion are consecutively numbered. 333 // 2. In the case that neighboring ShenandoahHeapRegions both 334 // represent old-gen memory, the card regions that span the 335 // boundary between these neighboring heap regions will be 336 // consecutively numbered. 337 // 3. (A corollary) In the case that an old-gen object straddles the 338 // boundary between two heap regions, the card regions that 339 // correspond to the span of this object will be consecutively 340 // numbered. 341 // 342 // ShenandoahCardCluster abstracts access to the remembered set 343 // and also keeps track of crossing map information to allow efficient 344 // resolution of object start addresses. 345 // 346 // ShenandoahCardCluster supports all of the services of 347 // RememberedSet, plus it supports register_object() and lookup_object(). 348 // Note that we only need to register the start addresses of the object that 349 // overlays the first address of a card; we need to do this for every card. 350 // In other words, register_object() checks if the object crosses a card boundary, 351 // and updates the offset value for each card that the object crosses into. 352 // For objects that don't straddle cards, nothing needs to be done. 353 // 354 // The RememberedSet template parameter is intended to represent either 355 // ShenandoahDirectCardMarkRememberedSet, or a to-be-implemented 356 // ShenandoahBufferWithSATBRememberedSet. 357 template<typename RememberedSet> 358 class ShenandoahCardCluster: public CHeapObj<mtGC> { 359 360 private: 361 RememberedSet *_rs; 362 363 public: 364 static const size_t CardsPerCluster = 64; 365 366 private: 367 typedef struct cross_map { uint8_t first; uint8_t last; } xmap; 368 typedef union crossing_info { uint16_t short_word; xmap offsets; } crossing_info; 369 370 // ObjectStartsInCardRegion bit is set within a crossing_info.offsets.start iff at least one object starts within 371 // a particular card region. We pack this bit into start byte under assumption that start byte is accessed less 372 // frequently than last byte. This is true when number of clean cards is greater than number of dirty cards. 373 static const uint16_t ObjectStartsInCardRegion = 0x80; 374 static const uint16_t FirstStartBits = 0x7f; 375 376 // Check that we have enough bits to store the largest possible offset into a card for an object start. 377 // The value for maximum card size is based on the constraints for GCCardSizeInBytes in gc_globals.hpp. 378 static const int MaxCardSize = NOT_LP64(512) LP64_ONLY(1024); 379 STATIC_ASSERT((MaxCardSize / HeapWordSize) - 1 <= FirstStartBits); 380 381 crossing_info *object_starts; 382 383 public: 384 // If we're setting first_start, assume the card has an object. 385 inline void set_first_start(size_t card_index, uint8_t value) { 386 object_starts[card_index].offsets.first = ObjectStartsInCardRegion | value; 387 } 388 389 inline void set_last_start(size_t card_index, uint8_t value) { 390 object_starts[card_index].offsets.last = value; 391 } 392 393 inline void set_starts_object_bit(size_t card_index) { 394 object_starts[card_index].offsets.first |= ObjectStartsInCardRegion; 395 } 396 397 inline void clear_starts_object_bit(size_t card_index) { 398 object_starts[card_index].offsets.first &= ~ObjectStartsInCardRegion; 399 } 400 401 // Returns true iff an object is known to start within the card memory associated with card card_index. 402 inline bool starts_object(size_t card_index) const { 403 return (object_starts[card_index].offsets.first & ObjectStartsInCardRegion) != 0; 404 } 405 406 inline void clear_objects_in_range(HeapWord *addr, size_t num_words) { 407 size_t card_index = _rs->card_index_for_addr(addr); 408 size_t last_card_index = _rs->card_index_for_addr(addr + num_words - 1); 409 while (card_index <= last_card_index) 410 object_starts[card_index++].short_word = 0; 411 } 412 413 ShenandoahCardCluster(RememberedSet *rs) { 414 _rs = rs; 415 // TODO: We don't really need object_starts entries for every card entry. We only need these for 416 // the card entries that correspond to old-gen memory. But for now, let's be quick and dirty. 417 object_starts = NEW_C_HEAP_ARRAY(crossing_info, rs->total_cards(), mtGC); 418 for (size_t i = 0; i < rs->total_cards(); i++) { 419 object_starts[i].short_word = 0; 420 } 421 } 422 423 ~ShenandoahCardCluster() { 424 FREE_C_HEAP_ARRAY(crossing_info, object_starts); 425 object_starts = nullptr; 426 } 427 428 // There is one entry within the object_starts array for each card entry. 429 // 430 // Suppose multiple garbage objects are coalesced during GC sweep 431 // into a single larger "free segment". As each two objects are 432 // coalesced together, the start information pertaining to the second 433 // object must be removed from the objects_starts array. If the 434 // second object had been been the first object within card memory, 435 // the new first object is the object that follows that object if 436 // that starts within the same card memory, or NoObject if the 437 // following object starts within the following cluster. If the 438 // second object had been the last object in the card memory, 439 // replace this entry with the newly coalesced object if it starts 440 // within the same card memory, or with NoObject if it starts in a 441 // preceding card's memory. 442 // 443 // Suppose a large free segment is divided into a smaller free 444 // segment and a new object. The second part of the newly divided 445 // memory must be registered as a new object, overwriting at most 446 // one first_start and one last_start entry. Note that one of the 447 // newly divided two objects might be a new GCLAB. 448 // 449 // Suppose postprocessing of a GCLAB finds that the original GCLAB 450 // has been divided into N objects. Each of the N newly allocated 451 // objects will be registered, overwriting at most one first_start 452 // and one last_start entries. 453 // 454 // No object registration operations are linear in the length of 455 // the registered objects. 456 // 457 // Consider further the following observations regarding object 458 // registration costs: 459 // 460 // 1. The cost is paid once for each old-gen object (Except when 461 // an object is demoted and repromoted, in which case we would 462 // pay the cost again). 463 // 2. The cost can be deferred so that there is no urgency during 464 // mutator copy-on-first-access promotion. Background GC 465 // threads will update the object_starts array by post- 466 // processing the contents of retired PLAB buffers. 467 // 3. The bet is that these costs are paid relatively rarely 468 // because: 469 // a) Most objects die young and objects that die in young-gen 470 // memory never need to be registered with the object_starts 471 // array. 472 // b) Most objects that are promoted into old-gen memory live 473 // there without further relocation for a relatively long 474 // time, so we get a lot of benefit from each investment 475 // in registering an object. 476 477 public: 478 479 // The starting locations of objects contained within old-gen memory 480 // are registered as part of the remembered set implementation. This 481 // information is required when scanning dirty card regions that are 482 // spanned by objects beginning within preceding card regions. It 483 // is necessary to find the first and last objects that begin within 484 // this card region. Starting addresses of objects are required to 485 // find the object headers, and object headers provide information 486 // about which fields within the object hold addresses. 487 // 488 // The old-gen memory allocator invokes register_object() for any 489 // object that is allocated within old-gen memory. This identifies 490 // the starting addresses of objects that span boundaries between 491 // card regions. 492 // 493 // It is not necessary to invoke register_object at the very instant 494 // an object is allocated. It is only necessary to invoke it 495 // prior to the next start of a garbage collection concurrent mark 496 // or concurrent update-references phase. An "ideal" time to register 497 // objects is during post-processing of a GCLAB after the GCLAB is 498 // retired due to depletion of its memory. 499 // 500 // register_object() does not perform synchronization. In the case 501 // that multiple threads are registering objects whose starting 502 // addresses are within the same cluster, races between these 503 // threads may result in corruption of the object-start data 504 // structures. Parallel GC threads should avoid registering objects 505 // residing within the same cluster by adhering to the following 506 // coordination protocols: 507 // 508 // 1. Align thread-local GCLAB buffers with some TBD multiple of 509 // card clusters. The card cluster size is 32 KB. If the 510 // desired GCLAB size is 128 KB, align the buffer on a multiple 511 // of 4 card clusters. 512 // 2. Post-process the contents of GCLAB buffers to register the 513 // objects allocated therein. Allow one GC thread at a 514 // time to do the post-processing of each GCLAB. 515 // 3. Since only one GC thread at a time is registering objects 516 // belonging to a particular allocation buffer, no locking 517 // is performed when registering these objects. 518 // 4. Any remnant of unallocated memory within an expended GC 519 // allocation buffer is not returned to the old-gen allocation 520 // pool until after the GC allocation buffer has been post 521 // processed. Before any remnant memory is returned to the 522 // old-gen allocation pool, the GC thread that scanned this GC 523 // allocation buffer performs a write-commit memory barrier. 524 // 5. Background GC threads that perform tenuring of young-gen 525 // objects without a GCLAB use a CAS lock before registering 526 // each tenured object. The CAS lock assures both mutual 527 // exclusion and memory coherency/visibility. Note that an 528 // object tenured by a background GC thread will not overlap 529 // with any of the clusters that are receiving tenured objects 530 // by way of GCLAB buffers. Multiple independent GC threads may 531 // attempt to tenure objects into a shared cluster. This is why 532 // sychronization may be necessary. Consider the following 533 // scenarios: 534 // 535 // a) If two objects are tenured into the same card region, each 536 // registration may attempt to modify the first-start or 537 // last-start information associated with that card region. 538 // Furthermore, because the representations of first-start 539 // and last-start information within the object_starts array 540 // entry uses different bits of a shared uint_16 to represent 541 // each, it is necessary to lock the entire card entry 542 // before modifying either the first-start or last-start 543 // information within the entry. 544 // b) Suppose GC thread X promotes a tenured object into 545 // card region A and this tenured object spans into 546 // neighboring card region B. Suppose GC thread Y (not equal 547 // to X) promotes a tenured object into cluster B. GC thread X 548 // will update the object_starts information for card A. No 549 // synchronization is required. 550 // c) In summary, when background GC threads register objects 551 // newly tenured into old-gen memory, they must acquire a 552 // mutual exclusion lock on the card that holds the starting 553 // address of the newly tenured object. This can be achieved 554 // by using a CAS instruction to assure that the previous 555 // values of first-offset and last-offset have not been 556 // changed since the same thread inquired as to their most 557 // current values. 558 // 559 // One way to minimize the need for synchronization between 560 // background tenuring GC threads is for each tenuring GC thread 561 // to promote young-gen objects into distinct dedicated cluster 562 // ranges. 563 // 6. The object_starts information is only required during the 564 // starting of concurrent marking and concurrent evacuation 565 // phases of GC. Before we start either of these GC phases, the 566 // JVM enters a safe point and all GC threads perform 567 // commit-write barriers to assure that access to the 568 // object_starts information is coherent. 569 570 571 // Notes on synchronization of register_object(): 572 // 573 // 1. For efficiency, there is no locking in the implementation of register_object() 574 // 2. Thus, it is required that users of this service assure that concurrent/parallel invocations of 575 // register_object() do pertain to the same card's memory range. See discussion below to understand 576 // the risks. 577 // 3. When allocating from a TLAB or GCLAB, the mutual exclusion can be guaranteed by assuring that each 578 // LAB's start and end are aligned on card memory boundaries. 579 // 4. Use the same lock that guarantees exclusivity when performing free-list allocation within heap regions. 580 // 581 // Register the newly allocated object while we're holding the global lock since there's no synchronization 582 // built in to the implementation of register_object(). There are potential races when multiple independent 583 // threads are allocating objects, some of which might span the same card region. For example, consider 584 // a card table's memory region within which three objects are being allocated by three different threads: 585 // 586 // objects being "concurrently" allocated: 587 // [-----a------][-----b-----][--------------c------------------] 588 // [---- card table memory range --------------] 589 // 590 // Before any objects are allocated, this card's memory range holds no objects. Note that: 591 // allocation of object a wants to set the has-object, first-start, and last-start attributes of the preceding card region. 592 // allocation of object b wants to set the has-object, first-start, and last-start attributes of this card region. 593 // allocation of object c also wants to set the has-object, first-start, and last-start attributes of this card region. 594 // 595 // The thread allocating b and the thread allocating c can "race" in various ways, resulting in confusion, such as last-start 596 // representing object b while first-start represents object c. This is why we need to require all register_object() 597 // invocations associated with objects that are allocated from "free lists" to provide their own mutual exclusion locking 598 // mechanism. 599 600 // Reset the starts_object() information to false for all cards in the range between from and to. 601 void reset_object_range(HeapWord *from, HeapWord *to); 602 603 // register_object() requires that the caller hold the heap lock 604 // before calling it. 605 void register_object(HeapWord* address); 606 607 // register_object_without_lock() does not require that the caller hold 608 // the heap lock before calling it, under the assumption that the 609 // caller has assure no other thread will endeavor to concurrently 610 // register objects that start within the same card's memory region 611 // as address. 612 void register_object_without_lock(HeapWord* address); 613 614 // During the reference updates phase of GC, we walk through each old-gen memory region that was 615 // not part of the collection set and we invalidate all unmarked objects. As part of this effort, 616 // we coalesce neighboring dead objects in order to make future remembered set scanning more 617 // efficient (since future remembered set scanning of any card region containing consecutive 618 // dead objects can skip over all of them at once by reading only a single dead object header 619 // instead of having to read the header of each of the coalesced dead objects. 620 // 621 // At some future time, we may implement a further optimization: satisfy future allocation requests 622 // by carving new objects out of the range of memory that represents the coalesced dead objects. 623 // 624 // Suppose we want to combine several dead objects into a single coalesced object. How does this 625 // impact our representation of crossing map information? 626 // 1. If the newly coalesced range is contained entirely within a card range, that card's last 627 // start entry either remains the same or it is changed to the start of the coalesced region. 628 // 2. For the card that holds the start of the coalesced object, it will not impact the first start 629 // but it may impact the last start. 630 // 3. For following cards spanned entirely by the newly coalesced object, it will change starts_object 631 // to false (and make first-start and last-start "undefined"). 632 // 4. For a following card that is spanned patially by the newly coalesced object, it may change 633 // first-start value, but it will not change the last-start value. 634 // 635 // The range of addresses represented by the arguments to coalesce_objects() must represent a range 636 // of memory that was previously occupied exactly by one or more previously registered objects. For 637 // convenience, it is legal to invoke coalesce_objects() with arguments that span a single previously 638 // registered object. 639 // 640 // The role of coalesce_objects is to change the crossing map information associated with all of the coalesced 641 // objects. 642 void coalesce_objects(HeapWord* address, size_t length_in_words); 643 644 // The typical use case is going to look something like this: 645 // for each heapregion that comprises old-gen memory 646 // for each card number that corresponds to this heap region 647 // scan the objects contained therein if the card is dirty 648 // To avoid excessive lookups in a sparse array, the API queries 649 // the card number pertaining to a particular address and then uses the 650 // card number for subsequent information lookups and stores. 651 652 // If starts_object(card_index), this returns the word offset within this card 653 // memory at which the first object begins. If !starts_object(card_index), the 654 // result is a don't care value -- asserts in a debug build. 655 size_t get_first_start(size_t card_index) const; 656 657 // If starts_object(card_index), this returns the word offset within this card 658 // memory at which the last object begins. If !starts_object(card_index), the 659 // result is a don't care value. 660 size_t get_last_start(size_t card_index) const; 661 662 663 // Given a card_index, return the starting address of the first block in the heap 664 // that straddles into the card. If the card is co-initial with an object, then 665 // this would return the starting address of the heap that this card covers. 666 // Expects to be called for a card affiliated with the old generation in 667 // generational mode. 668 HeapWord* block_start(size_t card_index) const; 669 }; 670 671 // ShenandoahScanRemembered is a concrete class representing the 672 // ability to scan the old-gen remembered set for references to 673 // objects residing in young-gen memory. 674 // 675 // Scanning normally begins with an invocation of numRegions and ends 676 // after all clusters of all regions have been scanned. 677 // 678 // Throughout the scanning effort, the number of regions does not 679 // change. 680 // 681 // Even though the regions that comprise old-gen memory are not 682 // necessarily contiguous, the abstraction represented by this class 683 // identifies each of the old-gen regions with an integer value 684 // in the range from 0 to (numRegions() - 1) inclusive. 685 // 686 687 template<typename RememberedSet> 688 class ShenandoahScanRemembered: public CHeapObj<mtGC> { 689 690 private: 691 RememberedSet* _rs; 692 ShenandoahCardCluster<RememberedSet>* _scc; 693 694 // Global card stats (cumulative) 695 HdrSeq _card_stats_scan_rs[MAX_CARD_STAT_TYPE]; 696 HdrSeq _card_stats_update_refs[MAX_CARD_STAT_TYPE]; 697 // Per worker card stats (multiplexed by phase) 698 HdrSeq** _card_stats; 699 700 // The types of card metrics that we gather 701 const char* _card_stats_name[MAX_CARD_STAT_TYPE] = { 702 "dirty_run", "clean_run", 703 "dirty_cards", "clean_cards", 704 "max_dirty_run", "max_clean_run", 705 "dirty_scan_objs", 706 "alternations" 707 }; 708 709 // The statistics are collected and logged separately for 710 // card-scans for initial marking, and for updating refs. 711 const char* _card_stat_log_type[MAX_CARD_STAT_LOG_TYPE] = { 712 "Scan Remembered Set", "Update Refs" 713 }; 714 715 int _card_stats_log_counter[2] = {0, 0}; 716 717 public: 718 // How to instantiate this object? 719 // ShenandoahDirectCardMarkRememberedSet *rs = 720 // new ShenandoahDirectCardMarkRememberedSet(); 721 // scr = new 722 // ShenandoahScanRememberd<ShenandoahDirectCardMarkRememberedSet>(rs); 723 // 724 // or, after the planned implementation of 725 // ShenandoahBufferWithSATBRememberedSet has been completed: 726 // 727 // ShenandoahBufferWithSATBRememberedSet *rs = 728 // new ShenandoahBufferWithSATBRememberedSet(); 729 // scr = new 730 // ShenandoahScanRememberd<ShenandoahBufferWithSATBRememberedSet>(rs); 731 732 733 ShenandoahScanRemembered(RememberedSet *rs) { 734 _rs = rs; 735 _scc = new ShenandoahCardCluster<RememberedSet>(rs); 736 737 // We allocate ParallelGCThreads worth even though we usually only 738 // use up to ConcGCThreads, because degenerate collections may employ 739 // ParallelGCThreads for remembered set scanning. 740 if (ShenandoahEnableCardStats) { 741 _card_stats = NEW_C_HEAP_ARRAY(HdrSeq*, ParallelGCThreads, mtGC); 742 for (uint i = 0; i < ParallelGCThreads; i++) { 743 _card_stats[i] = new HdrSeq[MAX_CARD_STAT_TYPE]; 744 } 745 } else { 746 _card_stats = nullptr; 747 } 748 } 749 750 ~ShenandoahScanRemembered() { 751 delete _scc; 752 if (ShenandoahEnableCardStats) { 753 for (uint i = 0; i < ParallelGCThreads; i++) { 754 delete _card_stats[i]; 755 } 756 FREE_C_HEAP_ARRAY(HdrSeq*, _card_stats); 757 _card_stats = nullptr; 758 } 759 assert(_card_stats == nullptr, "Error"); 760 } 761 762 HdrSeq* card_stats(uint worker_id) { 763 assert(worker_id < ParallelGCThreads, "Error"); 764 assert(ShenandoahEnableCardStats == (_card_stats != nullptr), "Error"); 765 return ShenandoahEnableCardStats ? _card_stats[worker_id] : nullptr; 766 } 767 768 HdrSeq* card_stats_for_phase(CardStatLogType t) { 769 switch (t) { 770 case CARD_STAT_SCAN_RS: 771 return _card_stats_scan_rs; 772 case CARD_STAT_UPDATE_REFS: 773 return _card_stats_update_refs; 774 default: 775 guarantee(false, "No such CardStatLogType"); 776 } 777 return nullptr; // Quiet compiler 778 } 779 780 // TODO: We really don't want to share all of these APIs with arbitrary consumers of the ShenandoahScanRemembered abstraction. 781 // But in the spirit of quick and dirty for the time being, I'm going to go ahead and publish everything for right now. Some 782 // of existing code already depends on having access to these services (because existing code has not been written to honor 783 // full abstraction of remembered set scanning. In the not too distant future, we want to try to make most, if not all, of 784 // these services private. Two problems with publicizing: 785 // 1. Allowing arbitrary users to reach beneath the hood allows the users to make assumptions about underlying implementation. 786 // This will make it more difficult to change underlying implementation at a future time, such as when we eventually experiment 787 // with SATB-based implementation of remembered set representation. 788 // 2. If we carefully control sharing of certain of these services, we can reduce the overhead of synchronization by assuring 789 // that all users follow protocols that avoid contention that might require synchronization. When we publish these APIs, we 790 // lose control over who and how the data is accessed. As a result, we are required to insert more defensive measures into 791 // the implementation, including synchronization locks. 792 793 794 // Card index is zero-based relative to first spanned card region. 795 size_t last_valid_index(); 796 size_t total_cards(); 797 size_t card_index_for_addr(HeapWord *p); 798 HeapWord *addr_for_card_index(size_t card_index); 799 bool is_card_dirty(size_t card_index); 800 bool is_write_card_dirty(size_t card_index) { return _rs->is_write_card_dirty(card_index); } 801 void mark_card_as_dirty(size_t card_index); 802 void mark_range_as_dirty(size_t card_index, size_t num_cards); 803 void mark_card_as_clean(size_t card_index); 804 void mark_range_as_clean(size_t card_index, size_t num_cards); 805 bool is_card_dirty(HeapWord *p); 806 void mark_card_as_dirty(HeapWord *p); 807 void mark_range_as_dirty(HeapWord *p, size_t num_heap_words); 808 void mark_card_as_clean(HeapWord *p); 809 void mark_range_as_clean(HeapWord *p, size_t num_heap_words); 810 size_t cluster_count(); 811 812 // Called by GC thread at start of concurrent mark to exchange roles of read and write remembered sets. 813 void swap_remset() { _rs->swap_remset(); } 814 815 void reset_remset(HeapWord* start, size_t word_count) { _rs->reset_remset(start, word_count); } 816 817 void merge_write_table(HeapWord* start, size_t word_count) { _rs->merge_write_table(start, word_count); } 818 819 // Called by GC thread after scanning old remembered set in order to prepare for next GC pass 820 void clear_old_remset() { _rs->clear_old_remset(); } 821 822 size_t cluster_for_addr(HeapWord *addr); 823 HeapWord* addr_for_cluster(size_t cluster_no); 824 825 void reset_object_range(HeapWord *from, HeapWord *to); 826 void register_object(HeapWord *addr); 827 void register_object_without_lock(HeapWord *addr); 828 void coalesce_objects(HeapWord *addr, size_t length_in_words); 829 830 HeapWord* first_object_in_card(size_t card_index) { 831 if (_scc->starts_object(card_index)) { 832 return addr_for_card_index(card_index) + _scc->get_first_start(card_index); 833 } else { 834 return nullptr; 835 } 836 } 837 838 // Return true iff this object is "properly" registered. 839 bool verify_registration(HeapWord* address, ShenandoahMarkingContext* ctx); 840 841 // clear the cards to clean, and clear the object_starts info to no objects 842 void mark_range_as_empty(HeapWord *addr, size_t length_in_words); 843 844 // process_clusters() scans a portion of the remembered set 845 // for references from old gen into young. Several worker threads 846 // scan different portions of the remembered set by making parallel invocations 847 // of process_clusters() with each invocation scanning different 848 // "clusters" of the remembered set. 849 // 850 // An invocation of process_clusters() examines all of the 851 // intergenerational references spanned by `count` clusters starting 852 // with `first_cluster`. The `oops` argument is a worker-thread-local 853 // OopClosure that is applied to all "valid" references in the remembered set. 854 // 855 // A side-effect of executing process_clusters() is to update the remembered 856 // set entries (e.g. marking dirty cards clean if they no longer 857 // hold references to young-gen memory). 858 // 859 // An implementation of process_clusters() may choose to efficiently 860 // address more typical scenarios in the structure of remembered sets. E.g. 861 // in the generational setting, one might expect remembered sets to be very sparse 862 // (low mutation rates in the old generation leading to sparse dirty cards, 863 // each with very few intergenerational pointers). Specific implementations 864 // may choose to degrade gracefully as the sparsity assumption fails to hold, 865 // such as when there are sudden spikes in (premature) promotion or in the 866 // case of an underprovisioned, poorly-tuned, or poorly-shaped heap. 867 // 868 // At the start of a concurrent young generation marking cycle, we invoke process_clusters 869 // with ClosureType ShenandoahInitMarkRootsClosure. 870 // 871 // At the start of a concurrent evacuation phase, we invoke process_clusters with 872 // ClosureType ShenandoahEvacuateUpdateRootsClosure. 873 874 // All template expansions require methods to be defined in the inline.hpp file, but larger 875 // such methods need not be declared as inline. 876 template <typename ClosureType> 877 void process_clusters(size_t first_cluster, size_t count, HeapWord *end_of_range, ClosureType *oops, 878 bool use_write_table, uint worker_id); 879 880 template <typename ClosureType> 881 inline void process_humongous_clusters(ShenandoahHeapRegion* r, size_t first_cluster, size_t count, 882 HeapWord *end_of_range, ClosureType *oops, bool use_write_table); 883 884 template <typename ClosureType> 885 inline void process_region_slice(ShenandoahHeapRegion* region, size_t offset, size_t clusters, HeapWord* end_of_range, 886 ClosureType *cl, bool use_write_table, uint worker_id); 887 888 // To Do: 889 // Create subclasses of ShenandoahInitMarkRootsClosure and 890 // ShenandoahEvacuateUpdateRootsClosure and any other closures 891 // that need to participate in remembered set scanning. Within the 892 // subclasses, add a (probably templated) instance variable that 893 // refers to the associated ShenandoahCardCluster object. Use this 894 // ShenandoahCardCluster instance to "enhance" the do_oops 895 // processing so that we can: 896 // 897 // 1. Avoid processing references that correspond to clean card 898 // regions, and 899 // 2. Set card status to CLEAN when the associated card region no 900 // longer holds inter-generatioanal references. 901 // 902 // To enable efficient implementation of these behaviors, we 903 // probably also want to add a few fields into the 904 // ShenandoahCardCluster object that allow us to precompute and 905 // remember the addresses at which card status is going to change 906 // from dirty to clean and clean to dirty. The do_oops 907 // implementations will want to update this value each time they 908 // cross one of these boundaries. 909 void roots_do(OopIterateClosure* cl); 910 911 // Log stats related to card/RS stats for given phase t 912 void log_card_stats(uint nworkers, CardStatLogType t) PRODUCT_RETURN; 913 private: 914 // Log stats for given worker id related into given summary card/RS stats 915 void log_worker_card_stats(uint worker_id, HdrSeq* sum_stats) PRODUCT_RETURN; 916 917 // Log given stats 918 inline void log_card_stats(HdrSeq* stats) PRODUCT_RETURN; 919 920 // Merge the stats from worked_id into the given summary stats, and clear the worker_id's stats. 921 void merge_worker_card_stats_cumulative(HdrSeq* worker_stats, HdrSeq* sum_stats) PRODUCT_RETURN; 922 }; 923 924 925 // A ShenandoahRegionChunk represents a contiguous interval of a ShenandoahHeapRegion, typically representing 926 // work to be done by a worker thread. 927 struct ShenandoahRegionChunk { 928 ShenandoahHeapRegion *_r; // The region of which this represents a chunk 929 size_t _chunk_offset; // HeapWordSize offset 930 size_t _chunk_size; // HeapWordSize qty 931 }; 932 933 // ShenandoahRegionChunkIterator divides the total remembered set scanning effort into ShenandoahRegionChunks 934 // that are assigned one at a time to worker threads. (Here, we use the terms `assignments` and `chunks` 935 // interchangeably.) Note that the effort required to scan a range of memory is not necessarily a linear 936 // function of the size of the range. Some memory ranges hold only a small number of live objects. 937 // Some ranges hold primarily primitive (non-pointer) data. We start with larger chunk sizes because larger chunks 938 // reduce coordination overhead. We expect that the GC worker threads that receive more difficult assignments 939 // will work longer on those chunks. Meanwhile, other worker will threads repeatedly accept and complete multiple 940 // easier chunks. As the total amount of work remaining to be completed decreases, we decrease the size of chunks 941 // given to individual threads. This reduces the likelihood of significant imbalance between worker thread assignments 942 // when there is less meaningful work to be performed by the remaining worker threads while they wait for 943 // worker threads with difficult assignments to finish, reducing the overall duration of the phase. 944 945 class ShenandoahRegionChunkIterator : public StackObj { 946 private: 947 // The largest chunk size is 4 MiB, measured in words. Otherwise, remembered set scanning may become too unbalanced. 948 // If the largest chunk size is too small, there is too much overhead sifting out assignments to individual worker threads. 949 static const size_t _maximum_chunk_size_words = (4 * 1024 * 1024) / HeapWordSize; 950 951 static const size_t _clusters_in_smallest_chunk = 4; 952 953 // smallest_chunk_size is 4 clusters. Each cluster spans 128 KiB. 954 // This is computed from CardTable::card_size_in_words() * 955 // ShenandoahCardCluster<ShenandoahDirectCardMarkRememberedSet>::CardsPerCluster; 956 static size_t smallest_chunk_size_words() { 957 return _clusters_in_smallest_chunk * CardTable::card_size_in_words() * 958 ShenandoahCardCluster<ShenandoahDirectCardMarkRememberedSet>::CardsPerCluster; 959 } 960 961 // The total remembered set scanning effort is divided into chunks of work that are assigned to individual worker tasks. 962 // The chunks of assigned work are divided into groups, where the size of the typical group (_regular_group_size) is half the 963 // total number of regions. The first group may be larger than 964 // _regular_group_size in the case that the first group's chunk 965 // size is less than the region size. The last group may be larger 966 // than _regular_group_size because no group is allowed to 967 // have smaller assignments than _smallest_chunk_size, which is 128 KB. 968 969 // Under normal circumstances, no configuration needs more than _maximum_groups (default value of 16). 970 // The first group "effectively" processes chunks of size 1 MiB (or smaller for smaller region sizes). 971 // The last group processes chunks of size 128 KiB. There are four groups total. 972 973 // group[0] is 4 MiB chunk size (_maximum_chunk_size_words) 974 // group[1] is 2 MiB chunk size 975 // group[2] is 1 MiB chunk size 976 // group[3] is 512 KiB chunk size 977 // group[4] is 256 KiB chunk size 978 // group[5] is 128 Kib shunk size (_smallest_chunk_size_words = 4 * 64 * 64 979 static const size_t _maximum_groups = 6; 980 981 const ShenandoahHeap* _heap; 982 983 const size_t _regular_group_size; // Number of chunks in each group 984 const size_t _first_group_chunk_size_b4_rebalance; 985 const size_t _num_groups; // Number of groups in this configuration 986 const size_t _total_chunks; 987 988 shenandoah_padding(0); 989 volatile size_t _index; 990 shenandoah_padding(1); 991 992 size_t _region_index[_maximum_groups]; // The region index for the first region spanned by this group 993 size_t _group_offset[_maximum_groups]; // The offset at which group begins within first region spanned by this group 994 size_t _group_chunk_size[_maximum_groups]; // The size of each chunk within this group 995 size_t _group_entries[_maximum_groups]; // Total chunks spanned by this group and the ones before it. 996 997 // No implicit copying: iterators should be passed by reference to capture the state 998 NONCOPYABLE(ShenandoahRegionChunkIterator); 999 1000 // Makes use of _heap. 1001 size_t calc_regular_group_size(); 1002 1003 // Makes use of _regular_group_size, which must be initialized before call. 1004 size_t calc_first_group_chunk_size_b4_rebalance(); 1005 1006 // Makes use of _regular_group_size and _first_group_chunk_size_b4_rebalance, both of which must be initialized before call. 1007 size_t calc_num_groups(); 1008 1009 // Makes use of _regular_group_size, _first_group_chunk_size_b4_rebalance, which must be initialized before call. 1010 size_t calc_total_chunks(); 1011 1012 public: 1013 ShenandoahRegionChunkIterator(size_t worker_count); 1014 ShenandoahRegionChunkIterator(ShenandoahHeap* heap, size_t worker_count); 1015 1016 // Reset iterator to default state 1017 void reset(); 1018 1019 // Fills in assignment with next chunk of work and returns true iff there is more work. 1020 // Otherwise, returns false. This is multi-thread-safe. 1021 inline bool next(struct ShenandoahRegionChunk *assignment); 1022 1023 // This is *not* MT safe. However, in the absence of multithreaded access, it 1024 // can be used to determine if there is more work to do. 1025 inline bool has_next() const; 1026 }; 1027 1028 typedef ShenandoahScanRemembered<ShenandoahDirectCardMarkRememberedSet> RememberedScanner; 1029 1030 class ShenandoahScanRememberedTask : public WorkerTask { 1031 private: 1032 ShenandoahObjToScanQueueSet* _queue_set; 1033 ShenandoahObjToScanQueueSet* _old_queue_set; 1034 ShenandoahReferenceProcessor* _rp; 1035 ShenandoahRegionChunkIterator* _work_list; 1036 bool _is_concurrent; 1037 1038 public: 1039 ShenandoahScanRememberedTask(ShenandoahObjToScanQueueSet* queue_set, 1040 ShenandoahObjToScanQueueSet* old_queue_set, 1041 ShenandoahReferenceProcessor* rp, 1042 ShenandoahRegionChunkIterator* work_list, 1043 bool is_concurrent); 1044 1045 void work(uint worker_id); 1046 void do_work(uint worker_id); 1047 }; 1048 1049 // After Full GC is done, reconstruct the remembered set by iterating over OLD regions, 1050 // registering all objects between bottom() and top(), and dirtying the cards containing 1051 // cross-generational pointers. 1052 class ShenandoahReconstructRememberedSetTask : public WorkerTask { 1053 private: 1054 ShenandoahRegionIterator* _regions; 1055 1056 public: 1057 explicit ShenandoahReconstructRememberedSetTask(ShenandoahRegionIterator* regions); 1058 1059 void work(uint worker_id) override; 1060 }; 1061 1062 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHSCANREMEMBERED_HPP