1 /*
2 * Resizable, Scalable, Concurrent Hash Table
3 *
4 * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
5 * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
6 * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
7 *
8 * Code partially derived from nft_hash
9 * Rewritten with rehash code from br_multicast plus single list
10 * pointer as suggested by Josh Triplett
11 *
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License version 2 as
14 * published by the Free Software Foundation.
15 */
16
17 #include <linux/atomic.h>
18 #include <linux/kernel.h>
19 #include <linux/init.h>
20 #include <linux/log2.h>
21 #include <linux/sched.h>
22 #include <linux/rculist.h>
23 #include <linux/slab.h>
24 #include <linux/vmalloc.h>
25 #include <linux/mm.h>
26 #include <linux/jhash.h>
27 #include <linux/random.h>
28 #include <linux/rhashtable.h>
29 #include <linux/err.h>
30 #include <linux/export.h>
31
32 #define HASH_DEFAULT_SIZE 64UL
33 #define HASH_MIN_SIZE 4U
34 #define BUCKET_LOCKS_PER_CPU 32UL
35
36 union nested_table {
37 union nested_table __rcu *table;
38 struct rhash_head __rcu *bucket;
39 };
40
41 static u32 head_hashfn(struct rhashtable *ht,
42 const struct bucket_table *tbl,
43 const struct rhash_head *he)
44 {
45 return rht_head_hashfn(ht, tbl, he, ht->p);
46 }
47
48 #ifdef CONFIG_PROVE_LOCKING
49 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
50
51 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
52 {
53 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
54 }
55 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
56
57 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
58 {
59 spinlock_t *lock = rht_bucket_lock(tbl, hash);
60
61 return (debug_locks) ? lockdep_is_held(lock) : 1;
62 }
63 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
64 #else
65 #define ASSERT_RHT_MUTEX(HT)
66 #endif
67
68 static void nested_table_free(union nested_table *ntbl, unsigned int size)
69 {
70 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
71 const unsigned int len = 1 << shift;
72 unsigned int i;
73
74 ntbl = rcu_dereference_raw(ntbl->table);
75 if (!ntbl)
76 return;
77
78 if (size > len) {
79 size >>= shift;
80 for (i = 0; i < len; i++)
81 nested_table_free(ntbl + i, size);
82 }
83
84 kfree(ntbl);
85 }
86
87 static void nested_bucket_table_free(const struct bucket_table *tbl)
88 {
89 unsigned int size = tbl->size >> tbl->nest;
90 unsigned int len = 1 << tbl->nest;
91 union nested_table *ntbl;
92 unsigned int i;
93
94 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
95
96 for (i = 0; i < len; i++)
97 nested_table_free(ntbl + i, size);
98
99 kfree(ntbl);
100 }
101
102 static void bucket_table_free(const struct bucket_table *tbl)
103 {
104 if (tbl->nest)
105 nested_bucket_table_free(tbl);
106
107 free_bucket_spinlocks(tbl->locks);
108 kvfree(tbl);
109 }
110
111 static void bucket_table_free_rcu(struct rcu_head *head)
112 {
113 bucket_table_free(container_of(head, struct bucket_table, rcu));
114 }
115
116 static union nested_table *nested_table_alloc(struct rhashtable *ht,
117 union nested_table __rcu **prev,
118 unsigned int shifted,
119 unsigned int nhash)
120 {
121 union nested_table *ntbl;
122 int i;
123
124 ntbl = rcu_dereference(*prev);
125 if (ntbl)
126 return ntbl;
127
128 ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
129
130 if (ntbl && shifted) {
131 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0].bucket); i++)
132 INIT_RHT_NULLS_HEAD(ntbl[i].bucket, ht,
133 (i << shifted) | nhash);
134 }
135
136 rcu_assign_pointer(*prev, ntbl);
137
138 return ntbl;
139 }
140
141 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
142 size_t nbuckets,
143 gfp_t gfp)
144 {
145 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
146 struct bucket_table *tbl;
147 size_t size;
148
149 if (nbuckets < (1 << (shift + 1)))
150 return NULL;
151
152 size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
153
154 tbl = kzalloc(size, gfp);
155 if (!tbl)
156 return NULL;
157
158 if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
159 0, 0)) {
160 kfree(tbl);
161 return NULL;
162 }
163
164 tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
165
166 return tbl;
167 }
168
169 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
170 size_t nbuckets,
171 gfp_t gfp)
172 {
173 struct bucket_table *tbl = NULL;
174 size_t size, max_locks;
175 int i;
176
177 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
178 if (gfp != GFP_KERNEL)
179 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
180 else
181 tbl = kvzalloc(size, gfp);
182
183 size = nbuckets;
184
185 if (tbl == NULL && gfp != GFP_KERNEL) {
186 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
187 nbuckets = 0;
188 }
189 if (tbl == NULL)
190 return NULL;
191
192 tbl->size = size;
193
194 max_locks = size >> 1;
195 if (tbl->nest)
196 max_locks = min_t(size_t, max_locks, 1U << tbl->nest);
197
198 if (alloc_bucket_spinlocks(&tbl->locks, &tbl->locks_mask, max_locks,
199 ht->p.locks_mul, gfp) < 0) {
200 bucket_table_free(tbl);
201 return NULL;
202 }
203
204 INIT_LIST_HEAD(&tbl->walkers);
205
206 tbl->hash_rnd = get_random_u32();
207
208 for (i = 0; i < nbuckets; i++)
209 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
210
211 return tbl;
212 }
213
214 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
215 struct bucket_table *tbl)
216 {
217 struct bucket_table *new_tbl;
218
219 do {
220 new_tbl = tbl;
221 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
222 } while (tbl);
223
224 return new_tbl;
225 }
226
227 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
228 {
229 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
230 struct bucket_table *new_tbl = rhashtable_last_table(ht,
231 rht_dereference_rcu(old_tbl->future_tbl, ht));
232 struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
233 int err = -EAGAIN;
234 struct rhash_head *head, *next, *entry;
235 spinlock_t *new_bucket_lock;
236 unsigned int new_hash;
237
238 if (new_tbl->nest)
239 goto out;
240
241 err = -ENOENT;
242
243 rht_for_each(entry, old_tbl, old_hash) {
244 err = 0;
245 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
246
247 if (rht_is_a_nulls(next))
248 break;
249
250 pprev = &entry->next;
251 }
252
253 if (err)
254 goto out;
255
256 new_hash = head_hashfn(ht, new_tbl, entry);
257
258 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
259
260 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
261 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
262 new_tbl, new_hash);
263
264 RCU_INIT_POINTER(entry->next, head);
265
266 rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
267 spin_unlock(new_bucket_lock);
268
269 rcu_assign_pointer(*pprev, next);
270
271 out:
272 return err;
273 }
274
275 static int rhashtable_rehash_chain(struct rhashtable *ht,
276 unsigned int old_hash)
277 {
278 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
279 spinlock_t *old_bucket_lock;
280 int err;
281
282 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
283
284 spin_lock_bh(old_bucket_lock);
285 while (!(err = rhashtable_rehash_one(ht, old_hash)))
286 ;
287
288 if (err == -ENOENT) {
289 old_tbl->rehash++;
290 err = 0;
291 }
292 spin_unlock_bh(old_bucket_lock);
293
294 return err;
295 }
296
297 static int rhashtable_rehash_attach(struct rhashtable *ht,
298 struct bucket_table *old_tbl,
299 struct bucket_table *new_tbl)
300 {
301 /* Protect future_tbl using the first bucket lock. */
302 spin_lock_bh(old_tbl->locks);
303
304 /* Did somebody beat us to it? */
305 if (rcu_access_pointer(old_tbl->future_tbl)) {
306 spin_unlock_bh(old_tbl->locks);
307 return -EEXIST;
308 }
309
310 /* Make insertions go into the new, empty table right away. Deletions
311 * and lookups will be attempted in both tables until we synchronize.
312 */
313 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
314
315 spin_unlock_bh(old_tbl->locks);
316
317 return 0;
318 }
319
320 static int rhashtable_rehash_table(struct rhashtable *ht)
321 {
322 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
323 struct bucket_table *new_tbl;
324 struct rhashtable_walker *walker;
325 unsigned int old_hash;
326 int err;
327
328 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
329 if (!new_tbl)
330 return 0;
331
332 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
333 err = rhashtable_rehash_chain(ht, old_hash);
334 if (err)
335 return err;
336 }
337
338 /* Publish the new table pointer. */
339 rcu_assign_pointer(ht->tbl, new_tbl);
340
341 spin_lock(&ht->lock);
342 list_for_each_entry(walker, &old_tbl->walkers, list)
343 walker->tbl = NULL;
344 spin_unlock(&ht->lock);
345
346 /* Wait for readers. All new readers will see the new
347 * table, and thus no references to the old table will
348 * remain.
349 */
350 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
351
352 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
353 }
354
355 static int rhashtable_rehash_alloc(struct rhashtable *ht,
356 struct bucket_table *old_tbl,
357 unsigned int size)
358 {
359 struct bucket_table *new_tbl;
360 int err;
361
362 ASSERT_RHT_MUTEX(ht);
363
364 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
365 if (new_tbl == NULL)
366 return -ENOMEM;
367
368 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
369 if (err)
370 bucket_table_free(new_tbl);
371
372 return err;
373 }
374
375 /**
376 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
377 * @ht: the hash table to shrink
378 *
379 * This function shrinks the hash table to fit, i.e., the smallest
380 * size would not cause it to expand right away automatically.
381 *
382 * The caller must ensure that no concurrent resizing occurs by holding
383 * ht->mutex.
384 *
385 * The caller must ensure that no concurrent table mutations take place.
386 * It is however valid to have concurrent lookups if they are RCU protected.
387 *
388 * It is valid to have concurrent insertions and deletions protected by per
389 * bucket locks or concurrent RCU protected lookups and traversals.
390 */
391 static int rhashtable_shrink(struct rhashtable *ht)
392 {
393 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
394 unsigned int nelems = atomic_read(&ht->nelems);
395 unsigned int size = 0;
396
397 if (nelems)
398 size = roundup_pow_of_two(nelems * 3 / 2);
399 if (size < ht->p.min_size)
400 size = ht->p.min_size;
401
402 if (old_tbl->size <= size)
403 return 0;
404
405 if (rht_dereference(old_tbl->future_tbl, ht))
406 return -EEXIST;
407
408 return rhashtable_rehash_alloc(ht, old_tbl, size);
409 }
410
411 static void rht_deferred_worker(struct work_struct *work)
412 {
413 struct rhashtable *ht;
414 struct bucket_table *tbl;
415 int err = 0;
416
417 ht = container_of(work, struct rhashtable, run_work);
418 mutex_lock(&ht->mutex);
419
420 tbl = rht_dereference(ht->tbl, ht);
421 tbl = rhashtable_last_table(ht, tbl);
422
423 if (rht_grow_above_75(ht, tbl))
424 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
425 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
426 err = rhashtable_shrink(ht);
427 else if (tbl->nest)
428 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
429
430 if (!err)
431 err = rhashtable_rehash_table(ht);
432
433 mutex_unlock(&ht->mutex);
434
435 if (err)
436 schedule_work(&ht->run_work);
437 }
438
439 static int rhashtable_insert_rehash(struct rhashtable *ht,
440 struct bucket_table *tbl)
441 {
442 struct bucket_table *old_tbl;
443 struct bucket_table *new_tbl;
444 unsigned int size;
445 int err;
446
447 old_tbl = rht_dereference_rcu(ht->tbl, ht);
448
449 size = tbl->size;
450
451 err = -EBUSY;
452
453 if (rht_grow_above_75(ht, tbl))
454 size *= 2;
455 /* Do not schedule more than one rehash */
456 else if (old_tbl != tbl)
457 goto fail;
458
459 err = -ENOMEM;
460
461 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
462 if (new_tbl == NULL)
463 goto fail;
464
465 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
466 if (err) {
467 bucket_table_free(new_tbl);
468 if (err == -EEXIST)
469 err = 0;
470 } else
471 schedule_work(&ht->run_work);
472
473 return err;
474
475 fail:
476 /* Do not fail the insert if someone else did a rehash. */
477 if (likely(rcu_dereference_raw(tbl->future_tbl)))
478 return 0;
479
480 /* Schedule async rehash to retry allocation in process context. */
481 if (err == -ENOMEM)
482 schedule_work(&ht->run_work);
483
484 return err;
485 }
486
487 static void *rhashtable_lookup_one(struct rhashtable *ht,
488 struct bucket_table *tbl, unsigned int hash,
489 const void *key, struct rhash_head *obj)
490 {
491 struct rhashtable_compare_arg arg = {
492 .ht = ht,
493 .key = key,
494 };
495 struct rhash_head __rcu **pprev;
496 struct rhash_head *head;
497 int elasticity;
498
499 elasticity = RHT_ELASTICITY;
500 pprev = rht_bucket_var(tbl, hash);
501 rht_for_each_continue(head, *pprev, tbl, hash) {
502 struct rhlist_head *list;
503 struct rhlist_head *plist;
504
505 elasticity--;
506 if (!key ||
507 (ht->p.obj_cmpfn ?
508 ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
509 rhashtable_compare(&arg, rht_obj(ht, head)))) {
510 pprev = &head->next;
511 continue;
512 }
513
514 if (!ht->rhlist)
515 return rht_obj(ht, head);
516
517 list = container_of(obj, struct rhlist_head, rhead);
518 plist = container_of(head, struct rhlist_head, rhead);
519
520 RCU_INIT_POINTER(list->next, plist);
521 head = rht_dereference_bucket(head->next, tbl, hash);
522 RCU_INIT_POINTER(list->rhead.next, head);
523 rcu_assign_pointer(*pprev, obj);
524
525 return NULL;
526 }
527
528 if (elasticity <= 0)
529 return ERR_PTR(-EAGAIN);
530
531 return ERR_PTR(-ENOENT);
532 }
533
534 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
535 struct bucket_table *tbl,
536 unsigned int hash,
537 struct rhash_head *obj,
538 void *data)
539 {
540 struct rhash_head __rcu **pprev;
541 struct bucket_table *new_tbl;
542 struct rhash_head *head;
543
544 if (!IS_ERR_OR_NULL(data))
545 return ERR_PTR(-EEXIST);
546
547 if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
548 return ERR_CAST(data);
549
550 new_tbl = rcu_dereference(tbl->future_tbl);
551 if (new_tbl)
552 return new_tbl;
553
554 if (PTR_ERR(data) != -ENOENT)
555 return ERR_CAST(data);
556
557 if (unlikely(rht_grow_above_max(ht, tbl)))
558 return ERR_PTR(-E2BIG);
559
560 if (unlikely(rht_grow_above_100(ht, tbl)))
561 return ERR_PTR(-EAGAIN);
562
563 pprev = rht_bucket_insert(ht, tbl, hash);
564 if (!pprev)
565 return ERR_PTR(-ENOMEM);
566
567 head = rht_dereference_bucket(*pprev, tbl, hash);
568
569 RCU_INIT_POINTER(obj->next, head);
570 if (ht->rhlist) {
571 struct rhlist_head *list;
572
573 list = container_of(obj, struct rhlist_head, rhead);
574 RCU_INIT_POINTER(list->next, NULL);
575 }
576
577 rcu_assign_pointer(*pprev, obj);
578
579 atomic_inc(&ht->nelems);
580 if (rht_grow_above_75(ht, tbl))
581 schedule_work(&ht->run_work);
582
583 return NULL;
584 }
585
586 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
587 struct rhash_head *obj)
588 {
589 struct bucket_table *new_tbl;
590 struct bucket_table *tbl;
591 unsigned int hash;
592 spinlock_t *lock;
593 void *data;
594
595 tbl = rcu_dereference(ht->tbl);
596
597 /* All insertions must grab the oldest table containing
598 * the hashed bucket that is yet to be rehashed.
599 */
600 for (;;) {
601 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
602 lock = rht_bucket_lock(tbl, hash);
603 spin_lock_bh(lock);
604
605 if (tbl->rehash <= hash)
606 break;
607
608 spin_unlock_bh(lock);
609 tbl = rcu_dereference(tbl->future_tbl);
610 }
611
612 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
613 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
614 if (PTR_ERR(new_tbl) != -EEXIST)
615 data = ERR_CAST(new_tbl);
616
617 while (!IS_ERR_OR_NULL(new_tbl)) {
618 tbl = new_tbl;
619 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
620 spin_lock_nested(rht_bucket_lock(tbl, hash),
621 SINGLE_DEPTH_NESTING);
622
623 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
624 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
625 if (PTR_ERR(new_tbl) != -EEXIST)
626 data = ERR_CAST(new_tbl);
627
628 spin_unlock(rht_bucket_lock(tbl, hash));
629 }
630
631 spin_unlock_bh(lock);
632
633 if (PTR_ERR(data) == -EAGAIN)
634 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
635 -EAGAIN);
636
637 return data;
638 }
639
640 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
641 struct rhash_head *obj)
642 {
643 void *data;
644
645 do {
646 rcu_read_lock();
647 data = rhashtable_try_insert(ht, key, obj);
648 rcu_read_unlock();
649 } while (PTR_ERR(data) == -EAGAIN);
650
651 return data;
652 }
653 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
654
655 /**
656 * rhashtable_walk_enter - Initialise an iterator
657 * @ht: Table to walk over
658 * @iter: Hash table Iterator
659 *
660 * This function prepares a hash table walk.
661 *
662 * Note that if you restart a walk after rhashtable_walk_stop you
663 * may see the same object twice. Also, you may miss objects if
664 * there are removals in between rhashtable_walk_stop and the next
665 * call to rhashtable_walk_start.
666 *
667 * For a completely stable walk you should construct your own data
668 * structure outside the hash table.
669 *
670 * This function may sleep so you must not call it from interrupt
671 * context or with spin locks held.
672 *
673 * You must call rhashtable_walk_exit after this function returns.
674 */
675 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
676 {
677 iter->ht = ht;
678 iter->p = NULL;
679 iter->slot = 0;
680 iter->skip = 0;
681 iter->end_of_table = 0;
682
683 spin_lock(&ht->lock);
684 iter->walker.tbl =
685 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
686 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
687 spin_unlock(&ht->lock);
688 }
689 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
690
691 /**
692 * rhashtable_walk_exit - Free an iterator
693 * @iter: Hash table Iterator
694 *
695 * This function frees resources allocated by rhashtable_walk_init.
696 */
697 void rhashtable_walk_exit(struct rhashtable_iter *iter)
698 {
699 spin_lock(&iter->ht->lock);
700 if (iter->walker.tbl)
701 list_del(&iter->walker.list);
702 spin_unlock(&iter->ht->lock);
703 }
704 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
705
706 /**
707 * rhashtable_walk_start_check - Start a hash table walk
708 * @iter: Hash table iterator
709 *
710 * Start a hash table walk at the current iterator position. Note that we take
711 * the RCU lock in all cases including when we return an error. So you must
712 * always call rhashtable_walk_stop to clean up.
713 *
714 * Returns zero if successful.
715 *
716 * Returns -EAGAIN if resize event occured. Note that the iterator
717 * will rewind back to the beginning and you may use it immediately
718 * by calling rhashtable_walk_next.
719 *
720 * rhashtable_walk_start is defined as an inline variant that returns
721 * void. This is preferred in cases where the caller would ignore
722 * resize events and always continue.
723 */
724 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
725 __acquires(RCU)
726 {
727 struct rhashtable *ht = iter->ht;
728
729 rcu_read_lock();
730
731 spin_lock(&ht->lock);
732 if (iter->walker.tbl)
733 list_del(&iter->walker.list);
734 spin_unlock(&ht->lock);
735
736 if (!iter->walker.tbl && !iter->end_of_table) {
737 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
738 return -EAGAIN;
739 }
740
741 return 0;
742 }
743 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
744
745 /**
746 * __rhashtable_walk_find_next - Find the next element in a table (or the first
747 * one in case of a new walk).
748 *
749 * @iter: Hash table iterator
750 *
751 * Returns the found object or NULL when the end of the table is reached.
752 *
753 * Returns -EAGAIN if resize event occurred.
754 */
755 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
756 {
757 struct bucket_table *tbl = iter->walker.tbl;
758 struct rhlist_head *list = iter->list;
759 struct rhashtable *ht = iter->ht;
760 struct rhash_head *p = iter->p;
761 bool rhlist = ht->rhlist;
762
763 if (!tbl)
764 return NULL;
765
766 for (; iter->slot < tbl->size; iter->slot++) {
767 int skip = iter->skip;
768
769 rht_for_each_rcu(p, tbl, iter->slot) {
770 if (rhlist) {
771 list = container_of(p, struct rhlist_head,
772 rhead);
773 do {
774 if (!skip)
775 goto next;
776 skip--;
777 list = rcu_dereference(list->next);
778 } while (list);
779
780 continue;
781 }
782 if (!skip)
783 break;
784 skip--;
785 }
786
787 next:
788 if (!rht_is_a_nulls(p)) {
789 iter->skip++;
790 iter->p = p;
791 iter->list = list;
792 return rht_obj(ht, rhlist ? &list->rhead : p);
793 }
794
795 iter->skip = 0;
796 }
797
798 iter->p = NULL;
799
800 /* Ensure we see any new tables. */
801 smp_rmb();
802
803 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
804 if (iter->walker.tbl) {
805 iter->slot = 0;
806 iter->skip = 0;
807 return ERR_PTR(-EAGAIN);
808 } else {
809 iter->end_of_table = true;
810 }
811
812 return NULL;
813 }
814
815 /**
816 * rhashtable_walk_next - Return the next object and advance the iterator
817 * @iter: Hash table iterator
818 *
819 * Note that you must call rhashtable_walk_stop when you are finished
820 * with the walk.
821 *
822 * Returns the next object or NULL when the end of the table is reached.
823 *
824 * Returns -EAGAIN if resize event occurred. Note that the iterator
825 * will rewind back to the beginning and you may continue to use it.
826 */
827 void *rhashtable_walk_next(struct rhashtable_iter *iter)
828 {
829 struct rhlist_head *list = iter->list;
830 struct rhashtable *ht = iter->ht;
831 struct rhash_head *p = iter->p;
832 bool rhlist = ht->rhlist;
833
834 if (p) {
835 if (!rhlist || !(list = rcu_dereference(list->next))) {
836 p = rcu_dereference(p->next);
837 list = container_of(p, struct rhlist_head, rhead);
838 }
839 if (!rht_is_a_nulls(p)) {
840 iter->skip++;
841 iter->p = p;
842 iter->list = list;
843 return rht_obj(ht, rhlist ? &list->rhead : p);
844 }
845
846 /* At the end of this slot, switch to next one and then find
847 * next entry from that point.
848 */
849 iter->skip = 0;
850 iter->slot++;
851 }
852
853 return __rhashtable_walk_find_next(iter);
854 }
855 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
856
857 /**
858 * rhashtable_walk_peek - Return the next object but don't advance the iterator
859 * @iter: Hash table iterator
860 *
861 * Returns the next object or NULL when the end of the table is reached.
862 *
863 * Returns -EAGAIN if resize event occurred. Note that the iterator
864 * will rewind back to the beginning and you may continue to use it.
865 */
866 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
867 {
868 struct rhlist_head *list = iter->list;
869 struct rhashtable *ht = iter->ht;
870 struct rhash_head *p = iter->p;
871
872 if (p)
873 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
874
875 /* No object found in current iter, find next one in the table. */
876
877 if (iter->skip) {
878 /* A nonzero skip value points to the next entry in the table
879 * beyond that last one that was found. Decrement skip so
880 * we find the current value. __rhashtable_walk_find_next
881 * will restore the original value of skip assuming that
882 * the table hasn't changed.
883 */
884 iter->skip--;
885 }
886
887 return __rhashtable_walk_find_next(iter);
888 }
889 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
890
891 /**
892 * rhashtable_walk_stop - Finish a hash table walk
893 * @iter: Hash table iterator
894 *
895 * Finish a hash table walk. Does not reset the iterator to the start of the
896 * hash table.
897 */
898 void rhashtable_walk_stop(struct rhashtable_iter *iter)
899 __releases(RCU)
900 {
901 struct rhashtable *ht;
902 struct bucket_table *tbl = iter->walker.tbl;
903
904 if (!tbl)
905 goto out;
906
907 ht = iter->ht;
908
909 spin_lock(&ht->lock);
910 if (tbl->rehash < tbl->size)
911 list_add(&iter->walker.list, &tbl->walkers);
912 else
913 iter->walker.tbl = NULL;
914 spin_unlock(&ht->lock);
915
916 iter->p = NULL;
917
918 out:
919 rcu_read_unlock();
920 }
921 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
922
923 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
924 {
925 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
926 (unsigned long)params->min_size);
927 }
928
929 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
930 {
931 return jhash2(key, length, seed);
932 }
933
934 /**
935 * rhashtable_init - initialize a new hash table
936 * @ht: hash table to be initialized
937 * @params: configuration parameters
938 *
939 * Initializes a new hash table based on the provided configuration
940 * parameters. A table can be configured either with a variable or
941 * fixed length key:
942 *
943 * Configuration Example 1: Fixed length keys
944 * struct test_obj {
945 * int key;
946 * void * my_member;
947 * struct rhash_head node;
948 * };
949 *
950 * struct rhashtable_params params = {
951 * .head_offset = offsetof(struct test_obj, node),
952 * .key_offset = offsetof(struct test_obj, key),
953 * .key_len = sizeof(int),
954 * .hashfn = jhash,
955 * .nulls_base = (1U << RHT_BASE_SHIFT),
956 * };
957 *
958 * Configuration Example 2: Variable length keys
959 * struct test_obj {
960 * [...]
961 * struct rhash_head node;
962 * };
963 *
964 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
965 * {
966 * struct test_obj *obj = data;
967 *
968 * return [... hash ...];
969 * }
970 *
971 * struct rhashtable_params params = {
972 * .head_offset = offsetof(struct test_obj, node),
973 * .hashfn = jhash,
974 * .obj_hashfn = my_hash_fn,
975 * };
976 */
977 int rhashtable_init(struct rhashtable *ht,
978 const struct rhashtable_params *params)
979 {
980 struct bucket_table *tbl;
981 size_t size;
982
983 size = HASH_DEFAULT_SIZE;
984
985 if ((!params->key_len && !params->obj_hashfn) ||
986 (params->obj_hashfn && !params->obj_cmpfn))
987 return -EINVAL;
988
989 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
990 return -EINVAL;
991
992 memset(ht, 0, sizeof(*ht));
993 mutex_init(&ht->mutex);
994 spin_lock_init(&ht->lock);
995 memcpy(&ht->p, params, sizeof(*params));
996
997 if (params->min_size)
998 ht->p.min_size = roundup_pow_of_two(params->min_size);
999
1000 /* Cap total entries at 2^31 to avoid nelems overflow. */
1001 ht->max_elems = 1u << 31;
1002
1003 if (params->max_size) {
1004 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1005 if (ht->p.max_size < ht->max_elems / 2)
1006 ht->max_elems = ht->p.max_size * 2;
1007 }
1008
1009 ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1010
1011 if (params->nelem_hint)
1012 size = rounded_hashtable_size(&ht->p);
1013
1014 if (params->locks_mul)
1015 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1016 else
1017 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1018
1019 ht->key_len = ht->p.key_len;
1020 if (!params->hashfn) {
1021 ht->p.hashfn = jhash;
1022
1023 if (!(ht->key_len & (sizeof(u32) - 1))) {
1024 ht->key_len /= sizeof(u32);
1025 ht->p.hashfn = rhashtable_jhash2;
1026 }
1027 }
1028
1029 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1030 if (tbl == NULL)
1031 return -ENOMEM;
1032
1033 atomic_set(&ht->nelems, 0);
1034
1035 RCU_INIT_POINTER(ht->tbl, tbl);
1036
1037 INIT_WORK(&ht->run_work, rht_deferred_worker);
1038
1039 return 0;
1040 }
1041 EXPORT_SYMBOL_GPL(rhashtable_init);
1042
1043 /**
1044 * rhltable_init - initialize a new hash list table
1045 * @hlt: hash list table to be initialized
1046 * @params: configuration parameters
1047 *
1048 * Initializes a new hash list table.
1049 *
1050 * See documentation for rhashtable_init.
1051 */
1052 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1053 {
1054 int err;
1055
1056 /* No rhlist NULLs marking for now. */
1057 if (params->nulls_base)
1058 return -EINVAL;
1059
1060 err = rhashtable_init(&hlt->ht, params);
1061 hlt->ht.rhlist = true;
1062 return err;
1063 }
1064 EXPORT_SYMBOL_GPL(rhltable_init);
1065
1066 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1067 void (*free_fn)(void *ptr, void *arg),
1068 void *arg)
1069 {
1070 struct rhlist_head *list;
1071
1072 if (!ht->rhlist) {
1073 free_fn(rht_obj(ht, obj), arg);
1074 return;
1075 }
1076
1077 list = container_of(obj, struct rhlist_head, rhead);
1078 do {
1079 obj = &list->rhead;
1080 list = rht_dereference(list->next, ht);
1081 free_fn(rht_obj(ht, obj), arg);
1082 } while (list);
1083 }
1084
1085 /**
1086 * rhashtable_free_and_destroy - free elements and destroy hash table
1087 * @ht: the hash table to destroy
1088 * @free_fn: callback to release resources of element
1089 * @arg: pointer passed to free_fn
1090 *
1091 * Stops an eventual async resize. If defined, invokes free_fn for each
1092 * element to releasal resources. Please note that RCU protected
1093 * readers may still be accessing the elements. Releasing of resources
1094 * must occur in a compatible manner. Then frees the bucket array.
1095 *
1096 * This function will eventually sleep to wait for an async resize
1097 * to complete. The caller is responsible that no further write operations
1098 * occurs in parallel.
1099 */
1100 void rhashtable_free_and_destroy(struct rhashtable *ht,
1101 void (*free_fn)(void *ptr, void *arg),
1102 void *arg)
1103 {
1104 struct bucket_table *tbl;
1105 unsigned int i;
1106
1107 cancel_work_sync(&ht->run_work);
1108
1109 mutex_lock(&ht->mutex);
1110 tbl = rht_dereference(ht->tbl, ht);
1111 if (free_fn) {
1112 for (i = 0; i < tbl->size; i++) {
1113 struct rhash_head *pos, *next;
1114
1115 for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1116 next = !rht_is_a_nulls(pos) ?
1117 rht_dereference(pos->next, ht) : NULL;
1118 !rht_is_a_nulls(pos);
1119 pos = next,
1120 next = !rht_is_a_nulls(pos) ?
1121 rht_dereference(pos->next, ht) : NULL)
1122 rhashtable_free_one(ht, pos, free_fn, arg);
1123 }
1124 }
1125
1126 bucket_table_free(tbl);
1127 mutex_unlock(&ht->mutex);
1128 }
1129 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1130
1131 void rhashtable_destroy(struct rhashtable *ht)
1132 {
1133 return rhashtable_free_and_destroy(ht, NULL, NULL);
1134 }
1135 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1136
1137 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1138 unsigned int hash)
1139 {
1140 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1141 static struct rhash_head __rcu *rhnull =
1142 (struct rhash_head __rcu *)NULLS_MARKER(0);
1143 unsigned int index = hash & ((1 << tbl->nest) - 1);
1144 unsigned int size = tbl->size >> tbl->nest;
1145 unsigned int subhash = hash;
1146 union nested_table *ntbl;
1147
1148 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1149 ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1150 subhash >>= tbl->nest;
1151
1152 while (ntbl && size > (1 << shift)) {
1153 index = subhash & ((1 << shift) - 1);
1154 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1155 tbl, hash);
1156 size >>= shift;
1157 subhash >>= shift;
1158 }
1159
1160 if (!ntbl)
1161 return &rhnull;
1162
1163 return &ntbl[subhash].bucket;
1164
1165 }
1166 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1167
1168 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1169 struct bucket_table *tbl,
1170 unsigned int hash)
1171 {
1172 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1173 unsigned int index = hash & ((1 << tbl->nest) - 1);
1174 unsigned int size = tbl->size >> tbl->nest;
1175 union nested_table *ntbl;
1176 unsigned int shifted;
1177 unsigned int nhash;
1178
1179 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1180 hash >>= tbl->nest;
1181 nhash = index;
1182 shifted = tbl->nest;
1183 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1184 size <= (1 << shift) ? shifted : 0, nhash);
1185
1186 while (ntbl && size > (1 << shift)) {
1187 index = hash & ((1 << shift) - 1);
1188 size >>= shift;
1189 hash >>= shift;
1190 nhash |= index << shifted;
1191 shifted += shift;
1192 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1193 size <= (1 << shift) ? shifted : 0,
1194 nhash);
1195 }
1196
1197 if (!ntbl)
1198 return NULL;
1199
1200 return &ntbl[hash].bucket;
1201
1202 }
1203 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);
1204
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.