1 /*
2 * Copyright (c) 2014, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/altHashing.hpp"
27 #include "classfile/javaClasses.inline.hpp"
28 #include "gc/g1/g1CollectedHeap.inline.hpp"
29 #include "gc/g1/g1BarrierSet.hpp"
30 #include "gc/g1/g1StringDedup.hpp"
31 #include "gc/g1/g1StringDedupTable.hpp"
32 #include "gc/shared/gcLocker.hpp"
33 #include "logging/log.hpp"
34 #include "memory/padded.inline.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "oops/typeArrayOop.hpp"
37 #include "runtime/mutexLocker.hpp"
38
39 //
40 // List of deduplication table entries. Links table
41 // entries together using their _next fields.
42 //
43 class G1StringDedupEntryList : public CHeapObj<mtGC> {
44 private:
45 G1StringDedupEntry* _list;
46 size_t _length;
47
48 public:
49 G1StringDedupEntryList() :
50 _list(NULL),
51 _length(0) {
52 }
53
54 void add(G1StringDedupEntry* entry) {
55 entry->set_next(_list);
56 _list = entry;
57 _length++;
58 }
59
60 G1StringDedupEntry* remove() {
61 G1StringDedupEntry* entry = _list;
62 if (entry != NULL) {
63 _list = entry->next();
64 _length--;
65 }
66 return entry;
67 }
68
69 G1StringDedupEntry* remove_all() {
70 G1StringDedupEntry* list = _list;
71 _list = NULL;
72 return list;
73 }
74
75 size_t length() {
76 return _length;
77 }
78 };
79
80 //
81 // Cache of deduplication table entries. This cache provides fast allocation and
82 // reuse of table entries to lower the pressure on the underlying allocator.
83 // But more importantly, it provides fast/deferred freeing of table entries. This
84 // is important because freeing of table entries is done during stop-the-world
85 // phases and it is not uncommon for large number of entries to be freed at once.
86 // Tables entries that are freed during these phases are placed onto a freelist in
87 // the cache. The deduplication thread, which executes in a concurrent phase, will
88 // later reuse or free the underlying memory for these entries.
89 //
90 // The cache allows for single-threaded allocations and multi-threaded frees.
91 // Allocations are synchronized by StringDedupTable_lock as part of a table
92 // modification.
93 //
94 class G1StringDedupEntryCache : public CHeapObj<mtGC> {
95 private:
96 // One cache/overflow list per GC worker to allow lock less freeing of
97 // entries while doing a parallel scan of the table. Using PaddedEnd to
98 // avoid false sharing.
99 size_t _nlists;
100 size_t _max_list_length;
101 PaddedEnd<G1StringDedupEntryList>* _cached;
102 PaddedEnd<G1StringDedupEntryList>* _overflowed;
103
104 public:
105 G1StringDedupEntryCache(size_t max_size);
106 ~G1StringDedupEntryCache();
107
108 // Set max number of table entries to cache.
109 void set_max_size(size_t max_size);
110
111 // Get a table entry from the cache, or allocate a new entry if the cache is empty.
112 G1StringDedupEntry* alloc();
113
114 // Insert a table entry into the cache.
115 void free(G1StringDedupEntry* entry, uint worker_id);
116
117 // Returns current number of entries in the cache.
118 size_t size();
119
120 // Deletes overflowed entries.
121 void delete_overflowed();
122 };
123
124 G1StringDedupEntryCache::G1StringDedupEntryCache(size_t max_size) :
125 _nlists(ParallelGCThreads),
126 _max_list_length(0),
127 _cached(PaddedArray<G1StringDedupEntryList, mtGC>::create_unfreeable((uint)_nlists)),
128 _overflowed(PaddedArray<G1StringDedupEntryList, mtGC>::create_unfreeable((uint)_nlists)) {
129 set_max_size(max_size);
130 }
131
132 G1StringDedupEntryCache::~G1StringDedupEntryCache() {
133 ShouldNotReachHere();
134 }
135
136 void G1StringDedupEntryCache::set_max_size(size_t size) {
137 _max_list_length = size / _nlists;
138 }
139
140 G1StringDedupEntry* G1StringDedupEntryCache::alloc() {
141 for (size_t i = 0; i < _nlists; i++) {
142 G1StringDedupEntry* entry = _cached[i].remove();
143 if (entry != NULL) {
144 return entry;
145 }
146 }
147 return new G1StringDedupEntry();
148 }
149
150 void G1StringDedupEntryCache::free(G1StringDedupEntry* entry, uint worker_id) {
151 assert(entry->obj() != NULL, "Double free");
152 assert(worker_id < _nlists, "Invalid worker id");
153
154 entry->set_obj(NULL);
155 entry->set_hash(0);
156
157 if (_cached[worker_id].length() < _max_list_length) {
158 // Cache is not full
159 _cached[worker_id].add(entry);
160 } else {
161 // Cache is full, add to overflow list for later deletion
162 _overflowed[worker_id].add(entry);
163 }
164 }
165
166 size_t G1StringDedupEntryCache::size() {
167 size_t size = 0;
168 for (size_t i = 0; i < _nlists; i++) {
169 size += _cached[i].length();
170 }
171 return size;
172 }
173
174 void G1StringDedupEntryCache::delete_overflowed() {
175 double start = os::elapsedTime();
176 uintx count = 0;
177
178 for (size_t i = 0; i < _nlists; i++) {
179 G1StringDedupEntry* entry;
180
181 {
182 // The overflow list can be modified during safepoints, therefore
183 // we temporarily join the suspendible thread set while removing
184 // all entries from the list.
185 SuspendibleThreadSetJoiner sts_join;
186 entry = _overflowed[i].remove_all();
187 }
188
189 // Delete all entries
190 while (entry != NULL) {
191 G1StringDedupEntry* next = entry->next();
192 delete entry;
193 entry = next;
194 count++;
195 }
196 }
197
198 double end = os::elapsedTime();
199 log_trace(gc, stringdedup)("Deleted " UINTX_FORMAT " entries, " G1_STRDEDUP_TIME_FORMAT_MS,
200 count, G1_STRDEDUP_TIME_PARAM_MS(end - start));
201 }
202
203 G1StringDedupTable* G1StringDedupTable::_table = NULL;
204 G1StringDedupEntryCache* G1StringDedupTable::_entry_cache = NULL;
205
206 const size_t G1StringDedupTable::_min_size = (1 << 10); // 1024
207 const size_t G1StringDedupTable::_max_size = (1 << 24); // 16777216
208 const double G1StringDedupTable::_grow_load_factor = 2.0; // Grow table at 200% load
209 const double G1StringDedupTable::_shrink_load_factor = _grow_load_factor / 3.0; // Shrink table at 67% load
210 const double G1StringDedupTable::_max_cache_factor = 0.1; // Cache a maximum of 10% of the table size
211 const uintx G1StringDedupTable::_rehash_multiple = 60; // Hash bucket has 60 times more collisions than expected
212 const uintx G1StringDedupTable::_rehash_threshold = (uintx)(_rehash_multiple * _grow_load_factor);
213
214 uintx G1StringDedupTable::_entries_added = 0;
215 uintx G1StringDedupTable::_entries_removed = 0;
216 uintx G1StringDedupTable::_resize_count = 0;
217 uintx G1StringDedupTable::_rehash_count = 0;
218
219 G1StringDedupTable::G1StringDedupTable(size_t size, jint hash_seed) :
220 _size(size),
221 _entries(0),
222 _grow_threshold((uintx)(size * _grow_load_factor)),
223 _shrink_threshold((uintx)(size * _shrink_load_factor)),
224 _rehash_needed(false),
225 _hash_seed(hash_seed) {
226 assert(is_power_of_2(size), "Table size must be a power of 2");
227 _buckets = NEW_C_HEAP_ARRAY(G1StringDedupEntry*, _size, mtGC);
228 memset(_buckets, 0, _size * sizeof(G1StringDedupEntry*));
229 }
230
231 G1StringDedupTable::~G1StringDedupTable() {
232 FREE_C_HEAP_ARRAY(G1StringDedupEntry*, _buckets);
233 }
234
235 void G1StringDedupTable::create() {
236 assert(_table == NULL, "One string deduplication table allowed");
237 _entry_cache = new G1StringDedupEntryCache(_min_size * _max_cache_factor);
238 _table = new G1StringDedupTable(_min_size);
239 }
240
241 void G1StringDedupTable::add(typeArrayOop value, bool latin1, unsigned int hash, G1StringDedupEntry** list) {
242 G1StringDedupEntry* entry = _entry_cache->alloc();
243 entry->set_obj(value);
244 entry->set_hash(hash);
245 entry->set_latin1(latin1);
246 entry->set_next(*list);
247 *list = entry;
248 _entries++;
249 }
250
251 void G1StringDedupTable::remove(G1StringDedupEntry** pentry, uint worker_id) {
252 G1StringDedupEntry* entry = *pentry;
253 *pentry = entry->next();
254 _entry_cache->free(entry, worker_id);
255 }
256
257 void G1StringDedupTable::transfer(G1StringDedupEntry** pentry, G1StringDedupTable* dest) {
258 G1StringDedupEntry* entry = *pentry;
259 *pentry = entry->next();
260 unsigned int hash = entry->hash();
261 size_t index = dest->hash_to_index(hash);
262 G1StringDedupEntry** list = dest->bucket(index);
263 entry->set_next(*list);
264 *list = entry;
265 }
266
267 bool G1StringDedupTable::equals(typeArrayOop value1, typeArrayOop value2) {
268 return (value1 == value2 ||
269 (value1->length() == value2->length() &&
270 (!memcmp(value1->base(T_BYTE),
271 value2->base(T_BYTE),
272 value1->length() * sizeof(jbyte)))));
273 }
274
275 typeArrayOop G1StringDedupTable::lookup(typeArrayOop value, bool latin1, unsigned int hash,
276 G1StringDedupEntry** list, uintx &count) {
277 for (G1StringDedupEntry* entry = *list; entry != NULL; entry = entry->next()) {
278 if (entry->hash() == hash && entry->latin1() == latin1) {
279 typeArrayOop existing_value = entry->obj();
280 if (equals(value, existing_value)) {
281 // Match found
282 return existing_value;
283 }
284 }
285 count++;
286 }
287
288 // Not found
289 return NULL;
290 }
291
292 typeArrayOop G1StringDedupTable::lookup_or_add_inner(typeArrayOop value, bool latin1, unsigned int hash) {
293 size_t index = hash_to_index(hash);
294 G1StringDedupEntry** list = bucket(index);
295 uintx count = 0;
296
297 // Lookup in list
298 typeArrayOop existing_value = lookup(value, latin1, hash, list, count);
299
300 // Check if rehash is needed
301 if (count > _rehash_threshold) {
302 _rehash_needed = true;
303 }
304
305 if (existing_value == NULL) {
306 // Not found, add new entry
307 add(value, latin1, hash, list);
308
309 // Update statistics
310 _entries_added++;
311 }
312
313 return existing_value;
314 }
315
316 unsigned int G1StringDedupTable::hash_code(typeArrayOop value, bool latin1) {
317 unsigned int hash;
318 int length = value->length();
319 if (latin1) {
320 const jbyte* data = (jbyte*)value->base(T_BYTE);
321 if (use_java_hash()) {
322 hash = java_lang_String::hash_code(data, length);
323 } else {
324 hash = AltHashing::murmur3_32(_table->_hash_seed, data, length);
325 }
326 } else {
327 length /= sizeof(jchar) / sizeof(jbyte); // Convert number of bytes to number of chars
328 const jchar* data = (jchar*)value->base(T_CHAR);
329 if (use_java_hash()) {
330 hash = java_lang_String::hash_code(data, length);
331 } else {
332 hash = AltHashing::murmur3_32(_table->_hash_seed, data, length);
333 }
334 }
335
336 return hash;
337 }
338
339 void G1StringDedupTable::deduplicate(oop java_string, G1StringDedupStat& stat) {
340 assert(java_lang_String::is_instance(java_string), "Must be a string");
341 NoSafepointVerifier nsv;
342
343 stat.inc_inspected();
344
345 typeArrayOop value = java_lang_String::value(java_string);
346 if (value == NULL) {
347 // String has no value
348 stat.inc_skipped();
349 return;
350 }
351
352 bool latin1 = java_lang_String::is_latin1(java_string);
353 unsigned int hash = 0;
354
355 if (use_java_hash()) {
356 // Get hash code from cache
357 hash = java_lang_String::hash(java_string);
358 }
359
360 if (hash == 0) {
361 // Compute hash
362 hash = hash_code(value, latin1);
363 stat.inc_hashed();
364
365 if (use_java_hash() && hash != 0) {
366 // Store hash code in cache
367 java_lang_String::set_hash(java_string, hash);
368 }
369 }
370
371 typeArrayOop existing_value = lookup_or_add(value, latin1, hash);
372 if (existing_value == value) {
373 // Same value, already known
374 stat.inc_known();
375 return;
376 }
377
378 // Get size of value array
379 uintx size_in_bytes = value->size() * HeapWordSize;
380 stat.inc_new(size_in_bytes);
381
382 if (existing_value != NULL) {
383 // Enqueue the reference to make sure it is kept alive. Concurrent mark might
384 // otherwise declare it dead if there are no other strong references to this object.
385 G1BarrierSet::satb_enqueue(existing_value);
386
387 // Existing value found, deduplicate string
388 java_lang_String::set_value(java_string, existing_value);
389
390 if (G1CollectedHeap::heap()->is_in_young(value)) {
391 stat.inc_deduped_young(size_in_bytes);
392 } else {
393 stat.inc_deduped_old(size_in_bytes);
394 }
395 }
396 }
397
398 G1StringDedupTable* G1StringDedupTable::prepare_resize() {
399 size_t size = _table->_size;
400
401 // Check if the hashtable needs to be resized
402 if (_table->_entries > _table->_grow_threshold) {
403 // Grow table, double the size
404 size *= 2;
405 if (size > _max_size) {
406 // Too big, don't resize
407 return NULL;
408 }
409 } else if (_table->_entries < _table->_shrink_threshold) {
410 // Shrink table, half the size
411 size /= 2;
412 if (size < _min_size) {
413 // Too small, don't resize
414 return NULL;
415 }
416 } else if (StringDeduplicationResizeALot) {
417 // Force grow
418 size *= 2;
419 if (size > _max_size) {
420 // Too big, force shrink instead
421 size /= 4;
422 }
423 } else {
424 // Resize not needed
425 return NULL;
426 }
427
428 // Update statistics
429 _resize_count++;
430
431 // Update max cache size
432 _entry_cache->set_max_size(size * _max_cache_factor);
433
434 // Allocate the new table. The new table will be populated by workers
435 // calling unlink_or_oops_do() and finally installed by finish_resize().
436 return new G1StringDedupTable(size, _table->_hash_seed);
437 }
438
439 void G1StringDedupTable::finish_resize(G1StringDedupTable* resized_table) {
440 assert(resized_table != NULL, "Invalid table");
441
442 resized_table->_entries = _table->_entries;
443
444 // Free old table
445 delete _table;
446
447 // Install new table
448 _table = resized_table;
449 }
450
451 void G1StringDedupTable::unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl, uint worker_id) {
452 // The table is divided into partitions to allow lock-less parallel processing by
453 // multiple worker threads. A worker thread first claims a partition, which ensures
454 // exclusive access to that part of the table, then continues to process it. To allow
455 // shrinking of the table in parallel we also need to make sure that the same worker
456 // thread processes all partitions where entries will hash to the same destination
457 // partition. Since the table size is always a power of two and we always shrink by
458 // dividing the table in half, we know that for a given partition there is only one
459 // other partition whoes entries will hash to the same destination partition. That
460 // other partition is always the sibling partition in the second half of the table.
461 // For example, if the table is divided into 8 partitions, the sibling of partition 0
462 // is partition 4, the sibling of partition 1 is partition 5, etc.
463 size_t table_half = _table->_size / 2;
464
465 // Let each partition be one page worth of buckets
466 size_t partition_size = MIN2(table_half, os::vm_page_size() / sizeof(G1StringDedupEntry*));
467 assert(table_half % partition_size == 0, "Invalid partition size");
468
469 // Number of entries removed during the scan
470 uintx removed = 0;
471
472 for (;;) {
473 // Grab next partition to scan
474 size_t partition_begin = cl->claim_table_partition(partition_size);
475 size_t partition_end = partition_begin + partition_size;
476 if (partition_begin >= table_half) {
477 // End of table
478 break;
479 }
480
481 // Scan the partition followed by the sibling partition in the second half of the table
482 removed += unlink_or_oops_do(cl, partition_begin, partition_end, worker_id);
483 removed += unlink_or_oops_do(cl, table_half + partition_begin, table_half + partition_end, worker_id);
484 }
485
486 // Delayed update to avoid contention on the table lock
487 if (removed > 0) {
488 MutexLockerEx ml(StringDedupTable_lock, Mutex::_no_safepoint_check_flag);
489 _table->_entries -= removed;
490 _entries_removed += removed;
491 }
492 }
493
494 uintx G1StringDedupTable::unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl,
495 size_t partition_begin,
496 size_t partition_end,
497 uint worker_id) {
498 uintx removed = 0;
499 for (size_t bucket = partition_begin; bucket < partition_end; bucket++) {
500 G1StringDedupEntry** entry = _table->bucket(bucket);
501 while (*entry != NULL) {
502 oop* p = (oop*)(*entry)->obj_addr();
503 if (cl->is_alive(*p)) {
504 cl->keep_alive(p);
505 if (cl->is_resizing()) {
506 // We are resizing the table, transfer entry to the new table
507 _table->transfer(entry, cl->resized_table());
508 } else {
509 if (cl->is_rehashing()) {
510 // We are rehashing the table, rehash the entry but keep it
511 // in the table. We can't transfer entries into the new table
512 // at this point since we don't have exclusive access to all
513 // destination partitions. finish_rehash() will do a single
514 // threaded transfer of all entries.
515 typeArrayOop value = (typeArrayOop)*p;
516 bool latin1 = (*entry)->latin1();
517 unsigned int hash = hash_code(value, latin1);
518 (*entry)->set_hash(hash);
519 }
520
521 // Move to next entry
522 entry = (*entry)->next_addr();
523 }
524 } else {
525 // Not alive, remove entry from table
526 _table->remove(entry, worker_id);
527 removed++;
528 }
529 }
530 }
531
532 return removed;
533 }
534
535 G1StringDedupTable* G1StringDedupTable::prepare_rehash() {
536 if (!_table->_rehash_needed && !StringDeduplicationRehashALot) {
537 // Rehash not needed
538 return NULL;
539 }
540
541 // Update statistics
542 _rehash_count++;
543
544 // Compute new hash seed
545 _table->_hash_seed = AltHashing::compute_seed();
546
547 // Allocate the new table, same size and hash seed
548 return new G1StringDedupTable(_table->_size, _table->_hash_seed);
549 }
550
551 void G1StringDedupTable::finish_rehash(G1StringDedupTable* rehashed_table) {
552 assert(rehashed_table != NULL, "Invalid table");
553
554 // Move all newly rehashed entries into the correct buckets in the new table
555 for (size_t bucket = 0; bucket < _table->_size; bucket++) {
556 G1StringDedupEntry** entry = _table->bucket(bucket);
557 while (*entry != NULL) {
558 _table->transfer(entry, rehashed_table);
559 }
560 }
561
562 rehashed_table->_entries = _table->_entries;
563
564 // Free old table
565 delete _table;
566
567 // Install new table
568 _table = rehashed_table;
569 }
570
571 void G1StringDedupTable::verify() {
572 for (size_t bucket = 0; bucket < _table->_size; bucket++) {
573 // Verify entries
574 G1StringDedupEntry** entry = _table->bucket(bucket);
575 while (*entry != NULL) {
576 typeArrayOop value = (*entry)->obj();
577 guarantee(value != NULL, "Object must not be NULL");
578 guarantee(G1CollectedHeap::heap()->is_in_reserved(value), "Object must be on the heap");
579 guarantee(!value->is_forwarded(), "Object must not be forwarded");
580 guarantee(value->is_typeArray(), "Object must be a typeArrayOop");
581 bool latin1 = (*entry)->latin1();
582 unsigned int hash = hash_code(value, latin1);
583 guarantee((*entry)->hash() == hash, "Table entry has inorrect hash");
584 guarantee(_table->hash_to_index(hash) == bucket, "Table entry has incorrect index");
585 entry = (*entry)->next_addr();
586 }
587
588 // Verify that we do not have entries with identical oops or identical arrays.
589 // We only need to compare entries in the same bucket. If the same oop or an
590 // identical array has been inserted more than once into different/incorrect
591 // buckets the verification step above will catch that.
592 G1StringDedupEntry** entry1 = _table->bucket(bucket);
593 while (*entry1 != NULL) {
594 typeArrayOop value1 = (*entry1)->obj();
595 bool latin1_1 = (*entry1)->latin1();
596 G1StringDedupEntry** entry2 = (*entry1)->next_addr();
597 while (*entry2 != NULL) {
598 typeArrayOop value2 = (*entry2)->obj();
599 bool latin1_2 = (*entry2)->latin1();
600 guarantee(latin1_1 != latin1_2 || !equals(value1, value2), "Table entries must not have identical arrays");
601 entry2 = (*entry2)->next_addr();
602 }
603 entry1 = (*entry1)->next_addr();
604 }
605 }
606 }
607
608 void G1StringDedupTable::clean_entry_cache() {
609 _entry_cache->delete_overflowed();
610 }
611
612 void G1StringDedupTable::print_statistics() {
613 Log(gc, stringdedup) log;
614 log.debug(" Table");
615 log.debug(" Memory Usage: " G1_STRDEDUP_BYTES_FORMAT_NS,
616 G1_STRDEDUP_BYTES_PARAM(_table->_size * sizeof(G1StringDedupEntry*) + (_table->_entries + _entry_cache->size()) * sizeof(G1StringDedupEntry)));
617 log.debug(" Size: " SIZE_FORMAT ", Min: " SIZE_FORMAT ", Max: " SIZE_FORMAT, _table->_size, _min_size, _max_size);
618 log.debug(" Entries: " UINTX_FORMAT ", Load: " G1_STRDEDUP_PERCENT_FORMAT_NS ", Cached: " UINTX_FORMAT ", Added: " UINTX_FORMAT ", Removed: " UINTX_FORMAT,
619 _table->_entries, (double)_table->_entries / (double)_table->_size * 100.0, _entry_cache->size(), _entries_added, _entries_removed);
620 log.debug(" Resize Count: " UINTX_FORMAT ", Shrink Threshold: " UINTX_FORMAT "(" G1_STRDEDUP_PERCENT_FORMAT_NS "), Grow Threshold: " UINTX_FORMAT "(" G1_STRDEDUP_PERCENT_FORMAT_NS ")",
621 _resize_count, _table->_shrink_threshold, _shrink_load_factor * 100.0, _table->_grow_threshold, _grow_load_factor * 100.0);
622 log.debug(" Rehash Count: " UINTX_FORMAT ", Rehash Threshold: " UINTX_FORMAT ", Hash Seed: 0x%x", _rehash_count, _rehash_threshold, _table->_hash_seed);
623 log.debug(" Age Threshold: " UINTX_FORMAT, StringDeduplicationAgeThreshold);
624 }