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 // Freelist in the deduplication table entry cache. Links table
41 // entries together using their _next fields.
42 //
43 class G1StringDedupEntryFreeList : public CHeapObj<mtGC> {
44 private:
45 G1StringDedupEntry* _list;
46 size_t _length;
47
48 public:
49 G1StringDedupEntryFreeList() :
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 size_t length() {
70 return _length;
71 }
72 };
73
74 //
75 // Cache of deduplication table entries. This cache provides fast allocation and
76 // reuse of table entries to lower the pressure on the underlying allocator.
77 // But more importantly, it provides fast/deferred freeing of table entries. This
78 // is important because freeing of table entries is done during stop-the-world
79 // phases and it is not uncommon for large number of entries to be freed at once.
80 // Tables entries that are freed during these phases are placed onto a freelist in
81 // the cache. The deduplication thread, which executes in a concurrent phase, will
82 // later reuse or free the underlying memory for these entries.
83 //
84 // The cache allows for single-threaded allocations and multi-threaded frees.
85 // Allocations are synchronized by StringDedupTable_lock as part of a table
86 // modification.
87 //
88 class G1StringDedupEntryCache : public CHeapObj<mtGC> {
89 private:
90 // One freelist per GC worker to allow lock less freeing of
91 // entries while doing a parallel scan of the table. Using
92 // PaddedEnd to avoid false sharing.
93 PaddedEnd<G1StringDedupEntryFreeList>* _lists;
94 size_t _nlists;
95
96 public:
97 G1StringDedupEntryCache();
98 ~G1StringDedupEntryCache();
99
100 // Get a table entry from the cache freelist, or allocate a new
101 // entry if the cache is empty.
102 G1StringDedupEntry* alloc();
103
104 // Insert a table entry into the cache freelist.
105 void free(G1StringDedupEntry* entry, uint worker_id);
106
107 // Returns current number of entries in the cache.
108 size_t size();
109
110 // If the cache has grown above the given max size, trim it down
111 // and deallocate the memory occupied by trimmed of entries.
112 void trim(size_t max_size);
113 };
114
115 G1StringDedupEntryCache::G1StringDedupEntryCache() {
116 _nlists = ParallelGCThreads;
117 _lists = PaddedArray<G1StringDedupEntryFreeList, mtGC>::create_unfreeable((uint)_nlists);
118 }
119
120 G1StringDedupEntryCache::~G1StringDedupEntryCache() {
121 ShouldNotReachHere();
122 }
123
124 G1StringDedupEntry* G1StringDedupEntryCache::alloc() {
125 for (size_t i = 0; i < _nlists; i++) {
126 G1StringDedupEntry* entry = _lists[i].remove();
127 if (entry != NULL) {
128 return entry;
129 }
130 }
131 return new G1StringDedupEntry();
132 }
133
134 void G1StringDedupEntryCache::free(G1StringDedupEntry* entry, uint worker_id) {
135 assert(entry->obj() != NULL, "Double free");
136 assert(worker_id < _nlists, "Invalid worker id");
137 entry->set_obj(NULL);
138 entry->set_hash(0);
139 _lists[worker_id].add(entry);
140 }
141
142 size_t G1StringDedupEntryCache::size() {
143 size_t size = 0;
144 for (size_t i = 0; i < _nlists; i++) {
145 size += _lists[i].length();
146 }
147 return size;
148 }
149
150 void G1StringDedupEntryCache::trim(size_t max_size) {
151 size_t cache_size = 0;
152 for (size_t i = 0; i < _nlists; i++) {
153 G1StringDedupEntryFreeList* list = &_lists[i];
154 cache_size += list->length();
155 while (cache_size > max_size) {
156 G1StringDedupEntry* entry = list->remove();
157 assert(entry != NULL, "Should not be null");
158 cache_size--;
159 delete entry;
160 }
161 }
162 }
163
164 G1StringDedupTable* G1StringDedupTable::_table = NULL;
165 G1StringDedupEntryCache* G1StringDedupTable::_entry_cache = NULL;
166
167 const size_t G1StringDedupTable::_min_size = (1 << 10); // 1024
168 const size_t G1StringDedupTable::_max_size = (1 << 24); // 16777216
169 const double G1StringDedupTable::_grow_load_factor = 2.0; // Grow table at 200% load
170 const double G1StringDedupTable::_shrink_load_factor = _grow_load_factor / 3.0; // Shrink table at 67% load
171 const double G1StringDedupTable::_max_cache_factor = 0.1; // Cache a maximum of 10% of the table size
172 const uintx G1StringDedupTable::_rehash_multiple = 60; // Hash bucket has 60 times more collisions than expected
173 const uintx G1StringDedupTable::_rehash_threshold = (uintx)(_rehash_multiple * _grow_load_factor);
174
175 uintx G1StringDedupTable::_entries_added = 0;
176 uintx G1StringDedupTable::_entries_removed = 0;
177 uintx G1StringDedupTable::_resize_count = 0;
178 uintx G1StringDedupTable::_rehash_count = 0;
179
180 G1StringDedupTable::G1StringDedupTable(size_t size, jint hash_seed) :
181 _size(size),
182 _entries(0),
183 _grow_threshold((uintx)(size * _grow_load_factor)),
184 _shrink_threshold((uintx)(size * _shrink_load_factor)),
185 _rehash_needed(false),
186 _hash_seed(hash_seed) {
187 assert(is_power_of_2(size), "Table size must be a power of 2");
188 _buckets = NEW_C_HEAP_ARRAY(G1StringDedupEntry*, _size, mtGC);
189 memset(_buckets, 0, _size * sizeof(G1StringDedupEntry*));
190 }
191
192 G1StringDedupTable::~G1StringDedupTable() {
193 FREE_C_HEAP_ARRAY(G1StringDedupEntry*, _buckets);
194 }
195
196 void G1StringDedupTable::create() {
197 assert(_table == NULL, "One string deduplication table allowed");
198 _entry_cache = new G1StringDedupEntryCache();
199 _table = new G1StringDedupTable(_min_size);
200 }
201
202 void G1StringDedupTable::add(typeArrayOop value, unsigned int hash, G1StringDedupEntry** list) {
203 G1StringDedupEntry* entry = _entry_cache->alloc();
204 entry->set_obj(value);
205 entry->set_hash(hash);
206 entry->set_next(*list);
207 *list = entry;
208 _entries++;
209 }
210
211 void G1StringDedupTable::remove(G1StringDedupEntry** pentry, uint worker_id) {
212 G1StringDedupEntry* entry = *pentry;
213 *pentry = entry->next();
214 _entry_cache->free(entry, worker_id);
215 }
216
217 void G1StringDedupTable::transfer(G1StringDedupEntry** pentry, G1StringDedupTable* dest) {
218 G1StringDedupEntry* entry = *pentry;
219 *pentry = entry->next();
220 unsigned int hash = entry->hash();
221 size_t index = dest->hash_to_index(hash);
222 G1StringDedupEntry** list = dest->bucket(index);
223 entry->set_next(*list);
224 *list = entry;
225 }
226
227 bool G1StringDedupTable::equals(typeArrayOop value1, typeArrayOop value2) {
228 return (value1 == value2 ||
229 (value1->length() == value2->length() &&
230 (!memcmp(value1->base(T_CHAR),
231 value2->base(T_CHAR),
232 value1->length() * sizeof(jchar)))));
233 }
234
235 typeArrayOop G1StringDedupTable::lookup(typeArrayOop value, unsigned int hash,
236 G1StringDedupEntry** list, uintx &count) {
237 for (G1StringDedupEntry* entry = *list; entry != NULL; entry = entry->next()) {
238 if (entry->hash() == hash) {
239 typeArrayOop existing_value = entry->obj();
240 if (equals(value, existing_value)) {
241 // Match found
242 return existing_value;
243 }
244 }
245 count++;
246 }
247
248 // Not found
249 return NULL;
250 }
251
252 typeArrayOop G1StringDedupTable::lookup_or_add_inner(typeArrayOop value, unsigned int hash) {
253 size_t index = hash_to_index(hash);
254 G1StringDedupEntry** list = bucket(index);
255 uintx count = 0;
256
257 // Lookup in list
258 typeArrayOop existing_value = lookup(value, hash, list, count);
259
260 // Check if rehash is needed
261 if (count > _rehash_threshold) {
262 _rehash_needed = true;
263 }
264
265 if (existing_value == NULL) {
266 // Not found, add new entry
267 add(value, hash, list);
268
269 // Update statistics
270 _entries_added++;
271 }
272
273 return existing_value;
274 }
275
276 unsigned int G1StringDedupTable::hash_code(typeArrayOop value) {
277 unsigned int hash;
278 int length = value->length();
279 const jchar* data = (jchar*)value->base(T_CHAR);
280
281 if (use_java_hash()) {
282 hash = java_lang_String::hash_code(data, length);
283 } else {
284 hash = AltHashing::murmur3_32(_table->_hash_seed, data, length);
285 }
286
287 return hash;
288 }
289
290 void G1StringDedupTable::deduplicate(oop java_string, G1StringDedupStat& stat) {
291 assert(java_lang_String::is_instance(java_string), "Must be a string");
292 No_Safepoint_Verifier nsv;
293
294 stat.inc_inspected();
295
296 typeArrayOop value = java_lang_String::value(java_string);
297 if (value == NULL) {
298 // String has no value
299 stat.inc_skipped();
300 return;
301 }
302
303 unsigned int hash = 0;
304
305 if (use_java_hash()) {
306 // Get hash code from cache
307 hash = java_lang_String::hash(java_string);
308 }
309
310 if (hash == 0) {
311 // Compute hash
312 hash = hash_code(value);
313 stat.inc_hashed();
314
315 if (use_java_hash() && hash != 0) {
316 // Store hash code in cache
317 java_lang_String::set_hash(java_string, hash);
318 }
319 }
320
321 typeArrayOop existing_value = lookup_or_add(value, hash);
322 if (existing_value == value) {
323 // Same value, already known
324 stat.inc_known();
325 return;
326 }
327
328 // Get size of value array
329 uintx size_in_bytes = value->size() * HeapWordSize;
330 stat.inc_new(size_in_bytes);
331
332 if (existing_value != NULL) {
333 // Enqueue the reference to make sure it is kept alive. Concurrent mark might
334 // otherwise declare it dead if there are no other strong references to this object.
335 G1SATBCardTableModRefBS::enqueue(existing_value);
336
337 // Existing value found, deduplicate string
338 java_lang_String::set_value(java_string, existing_value);
339
340 if (G1CollectedHeap::heap()->is_in_young(value)) {
341 stat.inc_deduped_young(size_in_bytes);
342 } else {
343 stat.inc_deduped_old(size_in_bytes);
344 }
345 }
346 }
347
348 G1StringDedupTable* G1StringDedupTable::prepare_resize() {
349 size_t size = _table->_size;
350
351 // Check if the hashtable needs to be resized
352 if (_table->_entries > _table->_grow_threshold) {
353 // Grow table, double the size
354 size *= 2;
355 if (size > _max_size) {
356 // Too big, don't resize
357 return NULL;
358 }
359 } else if (_table->_entries < _table->_shrink_threshold) {
360 // Shrink table, half the size
361 size /= 2;
362 if (size < _min_size) {
363 // Too small, don't resize
364 return NULL;
365 }
366 } else if (StringDeduplicationResizeALot) {
367 // Force grow
368 size *= 2;
369 if (size > _max_size) {
370 // Too big, force shrink instead
371 size /= 4;
372 }
373 } else {
374 // Resize not needed
375 return NULL;
376 }
377
378 // Update statistics
379 _resize_count++;
380
381 // Allocate the new table. The new table will be populated by workers
382 // calling unlink_or_oops_do() and finally installed by finish_resize().
383 return new G1StringDedupTable(size, _table->_hash_seed);
384 }
385
386 void G1StringDedupTable::finish_resize(G1StringDedupTable* resized_table) {
387 assert(resized_table != NULL, "Invalid table");
388
389 resized_table->_entries = _table->_entries;
390
391 // Free old table
392 delete _table;
393
394 // Install new table
395 _table = resized_table;
396 }
397
398 void G1StringDedupTable::unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl, uint worker_id) {
399 // The table is divided into partitions to allow lock-less parallel processing by
400 // multiple worker threads. A worker thread first claims a partition, which ensures
401 // exclusive access to that part of the table, then continues to process it. To allow
402 // shrinking of the table in parallel we also need to make sure that the same worker
403 // thread processes all partitions where entries will hash to the same destination
404 // partition. Since the table size is always a power of two and we always shrink by
405 // dividing the table in half, we know that for a given partition there is only one
406 // other partition whoes entries will hash to the same destination partition. That
407 // other partition is always the sibling partition in the second half of the table.
408 // For example, if the table is divided into 8 partitions, the sibling of partition 0
409 // is partition 4, the sibling of partition 1 is partition 5, etc.
410 size_t table_half = _table->_size / 2;
411
412 // Let each partition be one page worth of buckets
413 size_t partition_size = MIN2(table_half, os::vm_page_size() / sizeof(G1StringDedupEntry*));
414 assert(table_half % partition_size == 0, "Invalid partition size");
415
416 // Number of entries removed during the scan
417 uintx removed = 0;
418
419 for (;;) {
420 // Grab next partition to scan
421 size_t partition_begin = cl->claim_table_partition(partition_size);
422 size_t partition_end = partition_begin + partition_size;
423 if (partition_begin >= table_half) {
424 // End of table
425 break;
426 }
427
428 // Scan the partition followed by the sibling partition in the second half of the table
429 removed += unlink_or_oops_do(cl, partition_begin, partition_end, worker_id);
430 removed += unlink_or_oops_do(cl, table_half + partition_begin, table_half + partition_end, worker_id);
431 }
432
433 // Delayed update avoid contention on the table lock
434 if (removed > 0) {
435 MutexLockerEx ml(StringDedupTable_lock, Mutex::_no_safepoint_check_flag);
436 _table->_entries -= removed;
437 _entries_removed += removed;
438 }
439 }
440
441 uintx G1StringDedupTable::unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl,
442 size_t partition_begin,
443 size_t partition_end,
444 uint worker_id) {
445 uintx removed = 0;
446 for (size_t bucket = partition_begin; bucket < partition_end; bucket++) {
447 G1StringDedupEntry** entry = _table->bucket(bucket);
448 while (*entry != NULL) {
449 oop* p = (oop*)(*entry)->obj_addr();
450 if (cl->is_alive(*p)) {
451 cl->keep_alive(p);
452 if (cl->is_resizing()) {
453 // We are resizing the table, transfer entry to the new table
454 _table->transfer(entry, cl->resized_table());
455 } else {
456 if (cl->is_rehashing()) {
457 // We are rehashing the table, rehash the entry but keep it
458 // in the table. We can't transfer entries into the new table
459 // at this point since we don't have exclusive access to all
460 // destination partitions. finish_rehash() will do a single
461 // threaded transfer of all entries.
462 typeArrayOop value = (typeArrayOop)*p;
463 unsigned int hash = hash_code(value);
464 (*entry)->set_hash(hash);
465 }
466
467 // Move to next entry
468 entry = (*entry)->next_addr();
469 }
470 } else {
471 // Not alive, remove entry from table
472 _table->remove(entry, worker_id);
473 removed++;
474 }
475 }
476 }
477
478 return removed;
479 }
480
481 G1StringDedupTable* G1StringDedupTable::prepare_rehash() {
482 if (!_table->_rehash_needed && !StringDeduplicationRehashALot) {
483 // Rehash not needed
484 return NULL;
485 }
486
487 // Update statistics
488 _rehash_count++;
489
490 // Compute new hash seed
491 _table->_hash_seed = AltHashing::compute_seed();
492
493 // Allocate the new table, same size and hash seed
494 return new G1StringDedupTable(_table->_size, _table->_hash_seed);
495 }
496
497 void G1StringDedupTable::finish_rehash(G1StringDedupTable* rehashed_table) {
498 assert(rehashed_table != NULL, "Invalid table");
499
500 // Move all newly rehashed entries into the correct buckets in the new table
501 for (size_t bucket = 0; bucket < _table->_size; bucket++) {
502 G1StringDedupEntry** entry = _table->bucket(bucket);
503 while (*entry != NULL) {
504 _table->transfer(entry, rehashed_table);
505 }
506 }
507
508 rehashed_table->_entries = _table->_entries;
509
510 // Free old table
511 delete _table;
512
513 // Install new table
514 _table = rehashed_table;
515 }
516
517 void G1StringDedupTable::verify() {
518 for (size_t bucket = 0; bucket < _table->_size; bucket++) {
519 // Verify entries
520 G1StringDedupEntry** entry = _table->bucket(bucket);
521 while (*entry != NULL) {
522 typeArrayOop value = (*entry)->obj();
523 guarantee(value != NULL, "Object must not be NULL");
524 guarantee(G1CollectedHeap::heap()->is_in_reserved(value), "Object must be on the heap");
525 guarantee(!value->is_forwarded(), "Object must not be forwarded");
526 guarantee(value->is_typeArray(), "Object must be a typeArrayOop");
527 unsigned int hash = hash_code(value);
528 guarantee((*entry)->hash() == hash, "Table entry has inorrect hash");
529 guarantee(_table->hash_to_index(hash) == bucket, "Table entry has incorrect index");
530 entry = (*entry)->next_addr();
531 }
532
533 // Verify that we do not have entries with identical oops or identical arrays.
534 // We only need to compare entries in the same bucket. If the same oop or an
535 // identical array has been inserted more than once into different/incorrect
536 // buckets the verification step above will catch that.
537 G1StringDedupEntry** entry1 = _table->bucket(bucket);
538 while (*entry1 != NULL) {
539 typeArrayOop value1 = (*entry1)->obj();
540 G1StringDedupEntry** entry2 = (*entry1)->next_addr();
541 while (*entry2 != NULL) {
542 typeArrayOop value2 = (*entry2)->obj();
543 guarantee(!equals(value1, value2), "Table entries must not have identical arrays");
544 entry2 = (*entry2)->next_addr();
545 }
546 entry1 = (*entry1)->next_addr();
547 }
548 }
549 }
550
551 void G1StringDedupTable::trim_entry_cache() {
552 MutexLockerEx ml(StringDedupTable_lock, Mutex::_no_safepoint_check_flag);
553 size_t max_cache_size = (size_t)(_table->_size * _max_cache_factor);
554 _entry_cache->trim(max_cache_size);
555 }
556
557 void G1StringDedupTable::print_statistics() {
558 log_trace(gc, stringdedup)(
559 " [Table]\n"
560 " [Memory Usage: " G1_STRDEDUP_BYTES_FORMAT_NS "]\n"
561 " [Size: " SIZE_FORMAT ", Min: " SIZE_FORMAT ", Max: " SIZE_FORMAT "]\n"
562 " [Entries: " UINTX_FORMAT ", Load: " G1_STRDEDUP_PERCENT_FORMAT_NS ", Cached: " UINTX_FORMAT ", Added: " UINTX_FORMAT ", Removed: " UINTX_FORMAT "]\n"
563 " [Resize Count: " UINTX_FORMAT ", Shrink Threshold: " UINTX_FORMAT "(" G1_STRDEDUP_PERCENT_FORMAT_NS "), Grow Threshold: " UINTX_FORMAT "(" G1_STRDEDUP_PERCENT_FORMAT_NS ")]\n"
564 " [Rehash Count: " UINTX_FORMAT ", Rehash Threshold: " UINTX_FORMAT ", Hash Seed: 0x%x]\n"
565 " [Age Threshold: " UINTX_FORMAT "]",
566 G1_STRDEDUP_BYTES_PARAM(_table->_size * sizeof(G1StringDedupEntry*) + (_table->_entries + _entry_cache->size()) * sizeof(G1StringDedupEntry)),
567 _table->_size, _min_size, _max_size,
568 _table->_entries, (double)_table->_entries / (double)_table->_size * 100.0, _entry_cache->size(), _entries_added, _entries_removed,
569 _resize_count, _table->_shrink_threshold, _shrink_load_factor * 100.0, _table->_grow_threshold, _grow_load_factor * 100.0,
570 _rehash_count, _rehash_threshold, _table->_hash_seed,
571 StringDeduplicationAgeThreshold);
572 }