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