1 /*
2 * Copyright (c) 2014, 2015, 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/g1SATBCardTableModRefBS.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(max_size / _nlists),
127 _cached(PaddedArray<G1StringDedupEntryList, mtGC>::create_unfreeable((uint)_nlists)),
128 _overflowed(PaddedArray<G1StringDedupEntryList, mtGC>::create_unfreeable((uint)_nlists)) {
129 }
130
131 G1StringDedupEntryCache::~G1StringDedupEntryCache() {
132 ShouldNotReachHere();
133 }
134
135 void G1StringDedupEntryCache::set_max_size(size_t size) {
136 _max_list_length = size / _nlists;
137 }
138
139 G1StringDedupEntry* G1StringDedupEntryCache::alloc() {
140 for (size_t i = 0; i < _nlists; i++) {
141 G1StringDedupEntry* entry = _cached[i].remove();
142 if (entry != NULL) {
143 return entry;
144 }
145 }
146 return new G1StringDedupEntry();
147 }
148
149 void G1StringDedupEntryCache::free(G1StringDedupEntry* entry, uint worker_id) {
150 assert(entry->obj() != NULL, "Double free");
151 assert(worker_id < _nlists, "Invalid worker id");
152
153 entry->set_obj(NULL);
154 entry->set_hash(0);
155
156 if (_cached[worker_id].length() < _max_list_length) {
157 // Cache is not full
158 _cached[worker_id].add(entry);
159 } else {
160 // Cache is full, add to overflow list for later deletion
161 _overflowed[worker_id].add(entry);
162 }
163 }
164
165 size_t G1StringDedupEntryCache::size() {
166 size_t size = 0;
167 for (size_t i = 0; i < _nlists; i++) {
168 size += _cached[i].length();
169 }
170 return size;
171 }
172
173 void G1StringDedupEntryCache::delete_overflowed() {
174 double start = os::elapsedTime();
175 uintx count = 0;
176
177 for (size_t i = 0; i < _nlists; i++) {
178 G1StringDedupEntry* entry;
179
180 {
181 // The overflow list can be modified during safepoints, therefore
182 // we temporarily join the suspendible thread set while removing
183 // all entries from the list.
184 SuspendibleThreadSetJoiner sts_join;
185 entry = _overflowed[i].remove_all();
186 }
187
188 // Delete all entries
189 while (entry != NULL) {
190 G1StringDedupEntry* next = entry->next();
191 delete entry;
192 entry = next;
193 count++;
194 }
195 }
196
197 double end = os::elapsedTime();
198 log_trace(gc, stringdedup)("Deleted " UINTX_FORMAT " entries, " G1_STRDEDUP_TIME_FORMAT, count, end - start);
199 }
200
201 G1StringDedupTable* G1StringDedupTable::_table = NULL;
202 G1StringDedupEntryCache* G1StringDedupTable::_entry_cache = NULL;
203
204 const size_t G1StringDedupTable::_min_size = (1 << 10); // 1024
205 const size_t G1StringDedupTable::_max_size = (1 << 24); // 16777216
206 const double G1StringDedupTable::_grow_load_factor = 2.0; // Grow table at 200% load
207 const double G1StringDedupTable::_shrink_load_factor = _grow_load_factor / 3.0; // Shrink table at 67% load
208 const double G1StringDedupTable::_max_cache_factor = 0.1; // Cache a maximum of 10% of the table size
209 const uintx G1StringDedupTable::_rehash_multiple = 60; // Hash bucket has 60 times more collisions than expected
210 const uintx G1StringDedupTable::_rehash_threshold = (uintx)(_rehash_multiple * _grow_load_factor);
211
212 uintx G1StringDedupTable::_entries_added = 0;
213 uintx G1StringDedupTable::_entries_removed = 0;
214 uintx G1StringDedupTable::_resize_count = 0;
215 uintx G1StringDedupTable::_rehash_count = 0;
216
217 G1StringDedupTable::G1StringDedupTable(size_t size, jint hash_seed) :
218 _size(size),
219 _entries(0),
220 _grow_threshold((uintx)(size * _grow_load_factor)),
221 _shrink_threshold((uintx)(size * _shrink_load_factor)),
222 _rehash_needed(false),
223 _hash_seed(hash_seed) {
224 assert(is_power_of_2(size), "Table size must be a power of 2");
225 _buckets = NEW_C_HEAP_ARRAY(G1StringDedupEntry*, _size, mtGC);
226 memset(_buckets, 0, _size * sizeof(G1StringDedupEntry*));
227 }
228
229 G1StringDedupTable::~G1StringDedupTable() {
230 FREE_C_HEAP_ARRAY(G1StringDedupEntry*, _buckets);
231 }
232
233 void G1StringDedupTable::create() {
234 assert(_table == NULL, "One string deduplication table allowed");
235 _entry_cache = new G1StringDedupEntryCache(_min_size * _max_cache_factor);
236 _table = new G1StringDedupTable(_min_size);
237 }
238
239 void G1StringDedupTable::add(typeArrayOop value, bool latin1, unsigned int hash, G1StringDedupEntry** list) {
240 G1StringDedupEntry* entry = _entry_cache->alloc();
241 entry->set_obj(value);
242 entry->set_hash(hash);
243 entry->set_latin1(latin1);
244 entry->set_next(*list);
245 *list = entry;
246 _entries++;
247 }
248
249 void G1StringDedupTable::remove(G1StringDedupEntry** pentry, uint worker_id) {
250 G1StringDedupEntry* entry = *pentry;
251 *pentry = entry->next();
252 _entry_cache->free(entry, worker_id);
253 }
254
255 void G1StringDedupTable::transfer(G1StringDedupEntry** pentry, G1StringDedupTable* dest) {
256 G1StringDedupEntry* entry = *pentry;
257 *pentry = entry->next();
258 unsigned int hash = entry->hash();
259 size_t index = dest->hash_to_index(hash);
260 G1StringDedupEntry** list = dest->bucket(index);
261 entry->set_next(*list);
262 *list = entry;
263 }
264
265 bool G1StringDedupTable::equals(typeArrayOop value1, typeArrayOop value2) {
266 return (value1 == value2 ||
267 (value1->length() == value2->length() &&
268 (!memcmp(value1->base(T_BYTE),
269 value2->base(T_BYTE),
270 value1->length() * sizeof(jbyte)))));
271 }
272
273 typeArrayOop G1StringDedupTable::lookup(typeArrayOop value, bool latin1, unsigned int hash,
274 G1StringDedupEntry** list, uintx &count) {
275 for (G1StringDedupEntry* entry = *list; entry != NULL; entry = entry->next()) {
276 if (entry->hash() == hash && entry->latin1() == latin1) {
277 typeArrayOop existing_value = entry->obj();
278 if (equals(value, existing_value)) {
279 // Match found
280 return existing_value;
281 }
282 }
283 count++;
284 }
285
286 // Not found
287 return NULL;
288 }
289
290 typeArrayOop G1StringDedupTable::lookup_or_add_inner(typeArrayOop value, bool latin1, unsigned int hash) {
291 size_t index = hash_to_index(hash);
292 G1StringDedupEntry** list = bucket(index);
293 uintx count = 0;
294
295 // Lookup in list
296 typeArrayOop existing_value = lookup(value, latin1, hash, list, count);
297
298 // Check if rehash is needed
299 if (count > _rehash_threshold) {
300 _rehash_needed = true;
301 }
302
303 if (existing_value == NULL) {
304 // Not found, add new entry
305 add(value, latin1, hash, list);
306
307 // Update statistics
308 _entries_added++;
309 }
310
311 return existing_value;
312 }
313
314 unsigned int G1StringDedupTable::hash_code(typeArrayOop value, bool latin1) {
315 unsigned int hash;
316 int length = value->length();
317 if (latin1) {
318 const jbyte* data = (jbyte*)value->base(T_BYTE);
319 if (use_java_hash()) {
320 hash = java_lang_String::hash_code(data, length);
321 } else {
322 hash = AltHashing::murmur3_32(_table->_hash_seed, data, length);
323 }
324 } else {
325 length /= sizeof(jchar) / sizeof(jbyte); // Convert number of bytes to number of chars
326 const jchar* data = (jchar*)value->base(T_CHAR);
327 if (use_java_hash()) {
328 hash = java_lang_String::hash_code(data, length);
329 } else {
330 hash = AltHashing::murmur3_32(_table->_hash_seed, data, length);
331 }
332 }
333
334 return hash;
335 }
336
337 void G1StringDedupTable::deduplicate(oop java_string, G1StringDedupStat& stat) {
338 assert(java_lang_String::is_instance(java_string), "Must be a string");
339 NoSafepointVerifier nsv;
340
341 stat.inc_inspected();
342
343 typeArrayOop value = java_lang_String::value(java_string);
344 if (value == NULL) {
345 // String has no value
346 stat.inc_skipped();
347 return;
348 }
349
350 bool latin1 = java_lang_String::is_latin1(java_string);
351 unsigned int hash = 0;
352
353 if (use_java_hash()) {
354 // Get hash code from cache
355 hash = java_lang_String::hash(java_string);
356 }
357
358 if (hash == 0) {
359 // Compute hash
360 hash = hash_code(value, latin1);
361 stat.inc_hashed();
362
363 if (use_java_hash() && hash != 0) {
364 // Store hash code in cache
365 java_lang_String::set_hash(java_string, hash);
366 }
367 }
368
369 typeArrayOop existing_value = lookup_or_add(value, latin1, hash);
370 if (existing_value == value) {
371 // Same value, already known
372 stat.inc_known();
373 return;
374 }
375
376 // Get size of value array
377 uintx size_in_bytes = value->size() * HeapWordSize;
378 stat.inc_new(size_in_bytes);
379
380 if (existing_value != NULL) {
381 // Enqueue the reference to make sure it is kept alive. Concurrent mark might
382 // otherwise declare it dead if there are no other strong references to this object.
383 G1SATBCardTableModRefBS::enqueue(existing_value);
384
385 // Existing value found, deduplicate string
386 java_lang_String::set_value(java_string, existing_value);
387
388 if (G1CollectedHeap::heap()->is_in_young(value)) {
389 stat.inc_deduped_young(size_in_bytes);
390 } else {
391 stat.inc_deduped_old(size_in_bytes);
392 }
393 }
394 }
395
396 G1StringDedupTable* G1StringDedupTable::prepare_resize() {
397 size_t size = _table->_size;
398
399 // Check if the hashtable needs to be resized
400 if (_table->_entries > _table->_grow_threshold) {
401 // Grow table, double the size
402 size *= 2;
403 if (size > _max_size) {
404 // Too big, don't resize
405 return NULL;
406 }
407 } else if (_table->_entries < _table->_shrink_threshold) {
408 // Shrink table, half the size
409 size /= 2;
410 if (size < _min_size) {
411 // Too small, don't resize
412 return NULL;
413 }
414 } else if (StringDeduplicationResizeALot) {
415 // Force grow
416 size *= 2;
417 if (size > _max_size) {
418 // Too big, force shrink instead
419 size /= 4;
420 }
421 } else {
422 // Resize not needed
423 return NULL;
424 }
425
426 // Update statistics
427 _resize_count++;
428
429 // Update max cache size
430 _entry_cache->set_max_size(size * _max_cache_factor);
431
432 // Allocate the new table. The new table will be populated by workers
433 // calling unlink_or_oops_do() and finally installed by finish_resize().
434 return new G1StringDedupTable(size, _table->_hash_seed);
435 }
436
437 void G1StringDedupTable::finish_resize(G1StringDedupTable* resized_table) {
438 assert(resized_table != NULL, "Invalid table");
439
440 resized_table->_entries = _table->_entries;
441
442 // Free old table
443 delete _table;
444
445 // Install new table
446 _table = resized_table;
447 }
448
449 void G1StringDedupTable::unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl, uint worker_id) {
450 // The table is divided into partitions to allow lock-less parallel processing by
451 // multiple worker threads. A worker thread first claims a partition, which ensures
452 // exclusive access to that part of the table, then continues to process it. To allow
453 // shrinking of the table in parallel we also need to make sure that the same worker
454 // thread processes all partitions where entries will hash to the same destination
455 // partition. Since the table size is always a power of two and we always shrink by
456 // dividing the table in half, we know that for a given partition there is only one
457 // other partition whoes entries will hash to the same destination partition. That
458 // other partition is always the sibling partition in the second half of the table.
459 // For example, if the table is divided into 8 partitions, the sibling of partition 0
460 // is partition 4, the sibling of partition 1 is partition 5, etc.
461 size_t table_half = _table->_size / 2;
462
463 // Let each partition be one page worth of buckets
464 size_t partition_size = MIN2(table_half, os::vm_page_size() / sizeof(G1StringDedupEntry*));
465 assert(table_half % partition_size == 0, "Invalid partition size");
466
467 // Number of entries removed during the scan
468 uintx removed = 0;
469
470 for (;;) {
471 // Grab next partition to scan
472 size_t partition_begin = cl->claim_table_partition(partition_size);
473 size_t partition_end = partition_begin + partition_size;
474 if (partition_begin >= table_half) {
475 // End of table
476 break;
477 }
478
479 // Scan the partition followed by the sibling partition in the second half of the table
480 removed += unlink_or_oops_do(cl, partition_begin, partition_end, worker_id);
481 removed += unlink_or_oops_do(cl, table_half + partition_begin, table_half + partition_end, worker_id);
482 }
483
484 // Delayed update to avoid contention on the table lock
485 if (removed > 0) {
486 MutexLockerEx ml(StringDedupTable_lock, Mutex::_no_safepoint_check_flag);
487 _table->_entries -= removed;
488 _entries_removed += removed;
489 }
490 }
491
492 uintx G1StringDedupTable::unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl,
493 size_t partition_begin,
494 size_t partition_end,
495 uint worker_id) {
496 uintx removed = 0;
497 for (size_t bucket = partition_begin; bucket < partition_end; bucket++) {
498 G1StringDedupEntry** entry = _table->bucket(bucket);
499 while (*entry != NULL) {
500 oop* p = (oop*)(*entry)->obj_addr();
501 if (cl->is_alive(*p)) {
502 cl->keep_alive(p);
503 if (cl->is_resizing()) {
504 // We are resizing the table, transfer entry to the new table
505 _table->transfer(entry, cl->resized_table());
506 } else {
507 if (cl->is_rehashing()) {
508 // We are rehashing the table, rehash the entry but keep it
509 // in the table. We can't transfer entries into the new table
510 // at this point since we don't have exclusive access to all
511 // destination partitions. finish_rehash() will do a single
512 // threaded transfer of all entries.
513 typeArrayOop value = (typeArrayOop)*p;
514 bool latin1 = (*entry)->latin1();
515 unsigned int hash = hash_code(value, latin1);
516 (*entry)->set_hash(hash);
517 }
518
519 // Move to next entry
520 entry = (*entry)->next_addr();
521 }
522 } else {
523 // Not alive, remove entry from table
524 _table->remove(entry, worker_id);
525 removed++;
526 }
527 }
528 }
529
530 return removed;
531 }
532
533 G1StringDedupTable* G1StringDedupTable::prepare_rehash() {
534 if (!_table->_rehash_needed && !StringDeduplicationRehashALot) {
535 // Rehash not needed
536 return NULL;
537 }
538
539 // Update statistics
540 _rehash_count++;
541
542 // Compute new hash seed
543 _table->_hash_seed = AltHashing::compute_seed();
544
545 // Allocate the new table, same size and hash seed
546 return new G1StringDedupTable(_table->_size, _table->_hash_seed);
547 }
548
549 void G1StringDedupTable::finish_rehash(G1StringDedupTable* rehashed_table) {
550 assert(rehashed_table != NULL, "Invalid table");
551
552 // Move all newly rehashed entries into the correct buckets in the new table
553 for (size_t bucket = 0; bucket < _table->_size; bucket++) {
554 G1StringDedupEntry** entry = _table->bucket(bucket);
555 while (*entry != NULL) {
556 _table->transfer(entry, rehashed_table);
557 }
558 }
559
560 rehashed_table->_entries = _table->_entries;
561
562 // Free old table
563 delete _table;
564
565 // Install new table
566 _table = rehashed_table;
567 }
568
569 void G1StringDedupTable::verify() {
570 for (size_t bucket = 0; bucket < _table->_size; bucket++) {
571 // Verify entries
572 G1StringDedupEntry** entry = _table->bucket(bucket);
573 while (*entry != NULL) {
574 typeArrayOop value = (*entry)->obj();
575 guarantee(value != NULL, "Object must not be NULL");
576 guarantee(G1CollectedHeap::heap()->is_in_reserved(value), "Object must be on the heap");
577 guarantee(!value->is_forwarded(), "Object must not be forwarded");
578 guarantee(value->is_typeArray(), "Object must be a typeArrayOop");
579 bool latin1 = (*entry)->latin1();
580 unsigned int hash = hash_code(value, latin1);
581 guarantee((*entry)->hash() == hash, "Table entry has inorrect hash");
582 guarantee(_table->hash_to_index(hash) == bucket, "Table entry has incorrect index");
583 entry = (*entry)->next_addr();
584 }
585
586 // Verify that we do not have entries with identical oops or identical arrays.
587 // We only need to compare entries in the same bucket. If the same oop or an
588 // identical array has been inserted more than once into different/incorrect
589 // buckets the verification step above will catch that.
590 G1StringDedupEntry** entry1 = _table->bucket(bucket);
591 while (*entry1 != NULL) {
592 typeArrayOop value1 = (*entry1)->obj();
593 bool latin1_1 = (*entry1)->latin1();
594 G1StringDedupEntry** entry2 = (*entry1)->next_addr();
595 while (*entry2 != NULL) {
596 typeArrayOop value2 = (*entry2)->obj();
597 bool latin1_2 = (*entry2)->latin1();
598 guarantee(latin1_1 != latin1_2 || !equals(value1, value2), "Table entries must not have identical arrays");
599 entry2 = (*entry2)->next_addr();
600 }
601 entry1 = (*entry1)->next_addr();
602 }
603 }
604 }
605
606 void G1StringDedupTable::clean_entry_cache() {
607 _entry_cache->delete_overflowed();
608 }
609
610 void G1StringDedupTable::print_statistics() {
611 Log(gc, stringdedup) log;
612 log.debug(" [Table]");
613 log.debug(" [Memory Usage: " G1_STRDEDUP_BYTES_FORMAT_NS "]",
614 G1_STRDEDUP_BYTES_PARAM(_table->_size * sizeof(G1StringDedupEntry*) + (_table->_entries + _entry_cache->size()) * sizeof(G1StringDedupEntry)));
615 log.debug(" [Size: " SIZE_FORMAT ", Min: " SIZE_FORMAT ", Max: " SIZE_FORMAT "]", _table->_size, _min_size, _max_size);
616 log.debug(" [Entries: " UINTX_FORMAT ", Load: " G1_STRDEDUP_PERCENT_FORMAT_NS ", Cached: " UINTX_FORMAT ", Added: " UINTX_FORMAT ", Removed: " UINTX_FORMAT "]",
617 _table->_entries, (double)_table->_entries / (double)_table->_size * 100.0, _entry_cache->size(), _entries_added, _entries_removed);
618 log.debug(" [Resize Count: " UINTX_FORMAT ", Shrink Threshold: " UINTX_FORMAT "(" G1_STRDEDUP_PERCENT_FORMAT_NS "), Grow Threshold: " UINTX_FORMAT "(" G1_STRDEDUP_PERCENT_FORMAT_NS ")]",
619 _resize_count, _table->_shrink_threshold, _shrink_load_factor * 100.0, _table->_grow_threshold, _grow_load_factor * 100.0);
620 log.debug(" [Rehash Count: " UINTX_FORMAT ", Rehash Threshold: " UINTX_FORMAT ", Hash Seed: 0x%x]", _rehash_count, _rehash_threshold, _table->_hash_seed);
621 log.debug(" [Age Threshold: " UINTX_FORMAT "]", StringDeduplicationAgeThreshold);
622 }