Loading...
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/slab.h>
23#include <linux/vmalloc.h>
24#include <linux/mm.h>
25#include <linux/jhash.h>
26#include <linux/random.h>
27#include <linux/rhashtable.h>
28#include <linux/err.h>
29#include <linux/export.h>
30
31#define HASH_DEFAULT_SIZE 64UL
32#define HASH_MIN_SIZE 4U
33#define BUCKET_LOCKS_PER_CPU 32UL
34
35static u32 head_hashfn(struct rhashtable *ht,
36 const struct bucket_table *tbl,
37 const struct rhash_head *he)
38{
39 return rht_head_hashfn(ht, tbl, he, ht->p);
40}
41
42#ifdef CONFIG_PROVE_LOCKING
43#define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
44
45int lockdep_rht_mutex_is_held(struct rhashtable *ht)
46{
47 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
48}
49EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
50
51int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
52{
53 spinlock_t *lock = rht_bucket_lock(tbl, hash);
54
55 return (debug_locks) ? lockdep_is_held(lock) : 1;
56}
57EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
58#else
59#define ASSERT_RHT_MUTEX(HT)
60#endif
61
62
63static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
64 gfp_t gfp)
65{
66 unsigned int i, size;
67#if defined(CONFIG_PROVE_LOCKING)
68 unsigned int nr_pcpus = 2;
69#else
70 unsigned int nr_pcpus = num_possible_cpus();
71#endif
72
73 nr_pcpus = min_t(unsigned int, nr_pcpus, 64UL);
74 size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
75
76 /* Never allocate more than 0.5 locks per bucket */
77 size = min_t(unsigned int, size, tbl->size >> 1);
78
79 if (sizeof(spinlock_t) != 0) {
80 tbl->locks = NULL;
81#ifdef CONFIG_NUMA
82 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
83 gfp == GFP_KERNEL)
84 tbl->locks = vmalloc(size * sizeof(spinlock_t));
85#endif
86 if (gfp != GFP_KERNEL)
87 gfp |= __GFP_NOWARN | __GFP_NORETRY;
88
89 if (!tbl->locks)
90 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
91 gfp);
92 if (!tbl->locks)
93 return -ENOMEM;
94 for (i = 0; i < size; i++)
95 spin_lock_init(&tbl->locks[i]);
96 }
97 tbl->locks_mask = size - 1;
98
99 return 0;
100}
101
102static void bucket_table_free(const struct bucket_table *tbl)
103{
104 if (tbl)
105 kvfree(tbl->locks);
106
107 kvfree(tbl);
108}
109
110static void bucket_table_free_rcu(struct rcu_head *head)
111{
112 bucket_table_free(container_of(head, struct bucket_table, rcu));
113}
114
115static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
116 size_t nbuckets,
117 gfp_t gfp)
118{
119 struct bucket_table *tbl = NULL;
120 size_t size;
121 int i;
122
123 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
124 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
125 gfp != GFP_KERNEL)
126 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
127 if (tbl == NULL && gfp == GFP_KERNEL)
128 tbl = vzalloc(size);
129 if (tbl == NULL)
130 return NULL;
131
132 tbl->size = nbuckets;
133
134 if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
135 bucket_table_free(tbl);
136 return NULL;
137 }
138
139 INIT_LIST_HEAD(&tbl->walkers);
140
141 get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
142
143 for (i = 0; i < nbuckets; i++)
144 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
145
146 return tbl;
147}
148
149static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
150 struct bucket_table *tbl)
151{
152 struct bucket_table *new_tbl;
153
154 do {
155 new_tbl = tbl;
156 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
157 } while (tbl);
158
159 return new_tbl;
160}
161
162static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
163{
164 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
165 struct bucket_table *new_tbl = rhashtable_last_table(ht,
166 rht_dereference_rcu(old_tbl->future_tbl, ht));
167 struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
168 int err = -ENOENT;
169 struct rhash_head *head, *next, *entry;
170 spinlock_t *new_bucket_lock;
171 unsigned int new_hash;
172
173 rht_for_each(entry, old_tbl, old_hash) {
174 err = 0;
175 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
176
177 if (rht_is_a_nulls(next))
178 break;
179
180 pprev = &entry->next;
181 }
182
183 if (err)
184 goto out;
185
186 new_hash = head_hashfn(ht, new_tbl, entry);
187
188 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
189
190 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
191 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
192 new_tbl, new_hash);
193
194 RCU_INIT_POINTER(entry->next, head);
195
196 rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
197 spin_unlock(new_bucket_lock);
198
199 rcu_assign_pointer(*pprev, next);
200
201out:
202 return err;
203}
204
205static void rhashtable_rehash_chain(struct rhashtable *ht,
206 unsigned int old_hash)
207{
208 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
209 spinlock_t *old_bucket_lock;
210
211 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
212
213 spin_lock_bh(old_bucket_lock);
214 while (!rhashtable_rehash_one(ht, old_hash))
215 ;
216 old_tbl->rehash++;
217 spin_unlock_bh(old_bucket_lock);
218}
219
220static int rhashtable_rehash_attach(struct rhashtable *ht,
221 struct bucket_table *old_tbl,
222 struct bucket_table *new_tbl)
223{
224 /* Protect future_tbl using the first bucket lock. */
225 spin_lock_bh(old_tbl->locks);
226
227 /* Did somebody beat us to it? */
228 if (rcu_access_pointer(old_tbl->future_tbl)) {
229 spin_unlock_bh(old_tbl->locks);
230 return -EEXIST;
231 }
232
233 /* Make insertions go into the new, empty table right away. Deletions
234 * and lookups will be attempted in both tables until we synchronize.
235 */
236 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
237
238 spin_unlock_bh(old_tbl->locks);
239
240 return 0;
241}
242
243static int rhashtable_rehash_table(struct rhashtable *ht)
244{
245 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
246 struct bucket_table *new_tbl;
247 struct rhashtable_walker *walker;
248 unsigned int old_hash;
249
250 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
251 if (!new_tbl)
252 return 0;
253
254 for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
255 rhashtable_rehash_chain(ht, old_hash);
256
257 /* Publish the new table pointer. */
258 rcu_assign_pointer(ht->tbl, new_tbl);
259
260 spin_lock(&ht->lock);
261 list_for_each_entry(walker, &old_tbl->walkers, list)
262 walker->tbl = NULL;
263 spin_unlock(&ht->lock);
264
265 /* Wait for readers. All new readers will see the new
266 * table, and thus no references to the old table will
267 * remain.
268 */
269 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
270
271 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
272}
273
274/**
275 * rhashtable_expand - Expand hash table while allowing concurrent lookups
276 * @ht: the hash table to expand
277 *
278 * A secondary bucket array is allocated and the hash entries are migrated.
279 *
280 * This function may only be called in a context where it is safe to call
281 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
282 *
283 * The caller must ensure that no concurrent resizing occurs by holding
284 * ht->mutex.
285 *
286 * It is valid to have concurrent insertions and deletions protected by per
287 * bucket locks or concurrent RCU protected lookups and traversals.
288 */
289static int rhashtable_expand(struct rhashtable *ht)
290{
291 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
292 int err;
293
294 ASSERT_RHT_MUTEX(ht);
295
296 old_tbl = rhashtable_last_table(ht, old_tbl);
297
298 new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
299 if (new_tbl == NULL)
300 return -ENOMEM;
301
302 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
303 if (err)
304 bucket_table_free(new_tbl);
305
306 return err;
307}
308
309/**
310 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
311 * @ht: the hash table to shrink
312 *
313 * This function shrinks the hash table to fit, i.e., the smallest
314 * size would not cause it to expand right away automatically.
315 *
316 * The caller must ensure that no concurrent resizing occurs by holding
317 * ht->mutex.
318 *
319 * The caller must ensure that no concurrent table mutations take place.
320 * It is however valid to have concurrent lookups if they are RCU protected.
321 *
322 * It is valid to have concurrent insertions and deletions protected by per
323 * bucket locks or concurrent RCU protected lookups and traversals.
324 */
325static int rhashtable_shrink(struct rhashtable *ht)
326{
327 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
328 unsigned int nelems = atomic_read(&ht->nelems);
329 unsigned int size = 0;
330 int err;
331
332 ASSERT_RHT_MUTEX(ht);
333
334 if (nelems)
335 size = roundup_pow_of_two(nelems * 3 / 2);
336 if (size < ht->p.min_size)
337 size = ht->p.min_size;
338
339 if (old_tbl->size <= size)
340 return 0;
341
342 if (rht_dereference(old_tbl->future_tbl, ht))
343 return -EEXIST;
344
345 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
346 if (new_tbl == NULL)
347 return -ENOMEM;
348
349 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
350 if (err)
351 bucket_table_free(new_tbl);
352
353 return err;
354}
355
356static void rht_deferred_worker(struct work_struct *work)
357{
358 struct rhashtable *ht;
359 struct bucket_table *tbl;
360 int err = 0;
361
362 ht = container_of(work, struct rhashtable, run_work);
363 mutex_lock(&ht->mutex);
364
365 tbl = rht_dereference(ht->tbl, ht);
366 tbl = rhashtable_last_table(ht, tbl);
367
368 if (rht_grow_above_75(ht, tbl))
369 rhashtable_expand(ht);
370 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
371 rhashtable_shrink(ht);
372
373 err = rhashtable_rehash_table(ht);
374
375 mutex_unlock(&ht->mutex);
376
377 if (err)
378 schedule_work(&ht->run_work);
379}
380
381static int rhashtable_insert_rehash(struct rhashtable *ht,
382 struct bucket_table *tbl)
383{
384 struct bucket_table *old_tbl;
385 struct bucket_table *new_tbl;
386 unsigned int size;
387 int err;
388
389 old_tbl = rht_dereference_rcu(ht->tbl, ht);
390
391 size = tbl->size;
392
393 err = -EBUSY;
394
395 if (rht_grow_above_75(ht, tbl))
396 size *= 2;
397 /* Do not schedule more than one rehash */
398 else if (old_tbl != tbl)
399 goto fail;
400
401 err = -ENOMEM;
402
403 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
404 if (new_tbl == NULL)
405 goto fail;
406
407 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
408 if (err) {
409 bucket_table_free(new_tbl);
410 if (err == -EEXIST)
411 err = 0;
412 } else
413 schedule_work(&ht->run_work);
414
415 return err;
416
417fail:
418 /* Do not fail the insert if someone else did a rehash. */
419 if (likely(rcu_dereference_raw(tbl->future_tbl)))
420 return 0;
421
422 /* Schedule async rehash to retry allocation in process context. */
423 if (err == -ENOMEM)
424 schedule_work(&ht->run_work);
425
426 return err;
427}
428
429static void *rhashtable_lookup_one(struct rhashtable *ht,
430 struct bucket_table *tbl, unsigned int hash,
431 const void *key, struct rhash_head *obj)
432{
433 struct rhashtable_compare_arg arg = {
434 .ht = ht,
435 .key = key,
436 };
437 struct rhash_head __rcu **pprev;
438 struct rhash_head *head;
439 int elasticity;
440
441 elasticity = ht->elasticity;
442 pprev = &tbl->buckets[hash];
443 rht_for_each(head, tbl, hash) {
444 struct rhlist_head *list;
445 struct rhlist_head *plist;
446
447 elasticity--;
448 if (!key ||
449 (ht->p.obj_cmpfn ?
450 ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
451 rhashtable_compare(&arg, rht_obj(ht, head))))
452 continue;
453
454 if (!ht->rhlist)
455 return rht_obj(ht, head);
456
457 list = container_of(obj, struct rhlist_head, rhead);
458 plist = container_of(head, struct rhlist_head, rhead);
459
460 RCU_INIT_POINTER(list->next, plist);
461 head = rht_dereference_bucket(head->next, tbl, hash);
462 RCU_INIT_POINTER(list->rhead.next, head);
463 rcu_assign_pointer(*pprev, obj);
464
465 return NULL;
466 }
467
468 if (elasticity <= 0)
469 return ERR_PTR(-EAGAIN);
470
471 return ERR_PTR(-ENOENT);
472}
473
474static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
475 struct bucket_table *tbl,
476 unsigned int hash,
477 struct rhash_head *obj,
478 void *data)
479{
480 struct bucket_table *new_tbl;
481 struct rhash_head *head;
482
483 if (!IS_ERR_OR_NULL(data))
484 return ERR_PTR(-EEXIST);
485
486 if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
487 return ERR_CAST(data);
488
489 new_tbl = rcu_dereference(tbl->future_tbl);
490 if (new_tbl)
491 return new_tbl;
492
493 if (PTR_ERR(data) != -ENOENT)
494 return ERR_CAST(data);
495
496 if (unlikely(rht_grow_above_max(ht, tbl)))
497 return ERR_PTR(-E2BIG);
498
499 if (unlikely(rht_grow_above_100(ht, tbl)))
500 return ERR_PTR(-EAGAIN);
501
502 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
503
504 RCU_INIT_POINTER(obj->next, head);
505 if (ht->rhlist) {
506 struct rhlist_head *list;
507
508 list = container_of(obj, struct rhlist_head, rhead);
509 RCU_INIT_POINTER(list->next, NULL);
510 }
511
512 rcu_assign_pointer(tbl->buckets[hash], obj);
513
514 atomic_inc(&ht->nelems);
515 if (rht_grow_above_75(ht, tbl))
516 schedule_work(&ht->run_work);
517
518 return NULL;
519}
520
521static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
522 struct rhash_head *obj)
523{
524 struct bucket_table *new_tbl;
525 struct bucket_table *tbl;
526 unsigned int hash;
527 spinlock_t *lock;
528 void *data;
529
530 tbl = rcu_dereference(ht->tbl);
531
532 /* All insertions must grab the oldest table containing
533 * the hashed bucket that is yet to be rehashed.
534 */
535 for (;;) {
536 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
537 lock = rht_bucket_lock(tbl, hash);
538 spin_lock_bh(lock);
539
540 if (tbl->rehash <= hash)
541 break;
542
543 spin_unlock_bh(lock);
544 tbl = rcu_dereference(tbl->future_tbl);
545 }
546
547 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
548 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
549 if (PTR_ERR(new_tbl) != -EEXIST)
550 data = ERR_CAST(new_tbl);
551
552 while (!IS_ERR_OR_NULL(new_tbl)) {
553 tbl = new_tbl;
554 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
555 spin_lock_nested(rht_bucket_lock(tbl, hash),
556 SINGLE_DEPTH_NESTING);
557
558 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
559 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
560 if (PTR_ERR(new_tbl) != -EEXIST)
561 data = ERR_CAST(new_tbl);
562
563 spin_unlock(rht_bucket_lock(tbl, hash));
564 }
565
566 spin_unlock_bh(lock);
567
568 if (PTR_ERR(data) == -EAGAIN)
569 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
570 -EAGAIN);
571
572 return data;
573}
574
575void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
576 struct rhash_head *obj)
577{
578 void *data;
579
580 do {
581 rcu_read_lock();
582 data = rhashtable_try_insert(ht, key, obj);
583 rcu_read_unlock();
584 } while (PTR_ERR(data) == -EAGAIN);
585
586 return data;
587}
588EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
589
590/**
591 * rhashtable_walk_enter - Initialise an iterator
592 * @ht: Table to walk over
593 * @iter: Hash table Iterator
594 *
595 * This function prepares a hash table walk.
596 *
597 * Note that if you restart a walk after rhashtable_walk_stop you
598 * may see the same object twice. Also, you may miss objects if
599 * there are removals in between rhashtable_walk_stop and the next
600 * call to rhashtable_walk_start.
601 *
602 * For a completely stable walk you should construct your own data
603 * structure outside the hash table.
604 *
605 * This function may sleep so you must not call it from interrupt
606 * context or with spin locks held.
607 *
608 * You must call rhashtable_walk_exit after this function returns.
609 */
610void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
611{
612 iter->ht = ht;
613 iter->p = NULL;
614 iter->slot = 0;
615 iter->skip = 0;
616
617 spin_lock(&ht->lock);
618 iter->walker.tbl =
619 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
620 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
621 spin_unlock(&ht->lock);
622}
623EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
624
625/**
626 * rhashtable_walk_exit - Free an iterator
627 * @iter: Hash table Iterator
628 *
629 * This function frees resources allocated by rhashtable_walk_init.
630 */
631void rhashtable_walk_exit(struct rhashtable_iter *iter)
632{
633 spin_lock(&iter->ht->lock);
634 if (iter->walker.tbl)
635 list_del(&iter->walker.list);
636 spin_unlock(&iter->ht->lock);
637}
638EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
639
640/**
641 * rhashtable_walk_start - Start a hash table walk
642 * @iter: Hash table iterator
643 *
644 * Start a hash table walk. Note that we take the RCU lock in all
645 * cases including when we return an error. So you must always call
646 * rhashtable_walk_stop to clean up.
647 *
648 * Returns zero if successful.
649 *
650 * Returns -EAGAIN if resize event occured. Note that the iterator
651 * will rewind back to the beginning and you may use it immediately
652 * by calling rhashtable_walk_next.
653 */
654int rhashtable_walk_start(struct rhashtable_iter *iter)
655 __acquires(RCU)
656{
657 struct rhashtable *ht = iter->ht;
658
659 rcu_read_lock();
660
661 spin_lock(&ht->lock);
662 if (iter->walker.tbl)
663 list_del(&iter->walker.list);
664 spin_unlock(&ht->lock);
665
666 if (!iter->walker.tbl) {
667 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
668 return -EAGAIN;
669 }
670
671 return 0;
672}
673EXPORT_SYMBOL_GPL(rhashtable_walk_start);
674
675/**
676 * rhashtable_walk_next - Return the next object and advance the iterator
677 * @iter: Hash table iterator
678 *
679 * Note that you must call rhashtable_walk_stop when you are finished
680 * with the walk.
681 *
682 * Returns the next object or NULL when the end of the table is reached.
683 *
684 * Returns -EAGAIN if resize event occured. Note that the iterator
685 * will rewind back to the beginning and you may continue to use it.
686 */
687void *rhashtable_walk_next(struct rhashtable_iter *iter)
688{
689 struct bucket_table *tbl = iter->walker.tbl;
690 struct rhlist_head *list = iter->list;
691 struct rhashtable *ht = iter->ht;
692 struct rhash_head *p = iter->p;
693 bool rhlist = ht->rhlist;
694
695 if (p) {
696 if (!rhlist || !(list = rcu_dereference(list->next))) {
697 p = rcu_dereference(p->next);
698 list = container_of(p, struct rhlist_head, rhead);
699 }
700 goto next;
701 }
702
703 for (; iter->slot < tbl->size; iter->slot++) {
704 int skip = iter->skip;
705
706 rht_for_each_rcu(p, tbl, iter->slot) {
707 if (rhlist) {
708 list = container_of(p, struct rhlist_head,
709 rhead);
710 do {
711 if (!skip)
712 goto next;
713 skip--;
714 list = rcu_dereference(list->next);
715 } while (list);
716
717 continue;
718 }
719 if (!skip)
720 break;
721 skip--;
722 }
723
724next:
725 if (!rht_is_a_nulls(p)) {
726 iter->skip++;
727 iter->p = p;
728 iter->list = list;
729 return rht_obj(ht, rhlist ? &list->rhead : p);
730 }
731
732 iter->skip = 0;
733 }
734
735 iter->p = NULL;
736
737 /* Ensure we see any new tables. */
738 smp_rmb();
739
740 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
741 if (iter->walker.tbl) {
742 iter->slot = 0;
743 iter->skip = 0;
744 return ERR_PTR(-EAGAIN);
745 }
746
747 return NULL;
748}
749EXPORT_SYMBOL_GPL(rhashtable_walk_next);
750
751/**
752 * rhashtable_walk_stop - Finish a hash table walk
753 * @iter: Hash table iterator
754 *
755 * Finish a hash table walk.
756 */
757void rhashtable_walk_stop(struct rhashtable_iter *iter)
758 __releases(RCU)
759{
760 struct rhashtable *ht;
761 struct bucket_table *tbl = iter->walker.tbl;
762
763 if (!tbl)
764 goto out;
765
766 ht = iter->ht;
767
768 spin_lock(&ht->lock);
769 if (tbl->rehash < tbl->size)
770 list_add(&iter->walker.list, &tbl->walkers);
771 else
772 iter->walker.tbl = NULL;
773 spin_unlock(&ht->lock);
774
775 iter->p = NULL;
776
777out:
778 rcu_read_unlock();
779}
780EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
781
782static size_t rounded_hashtable_size(const struct rhashtable_params *params)
783{
784 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
785 (unsigned long)params->min_size);
786}
787
788static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
789{
790 return jhash2(key, length, seed);
791}
792
793/**
794 * rhashtable_init - initialize a new hash table
795 * @ht: hash table to be initialized
796 * @params: configuration parameters
797 *
798 * Initializes a new hash table based on the provided configuration
799 * parameters. A table can be configured either with a variable or
800 * fixed length key:
801 *
802 * Configuration Example 1: Fixed length keys
803 * struct test_obj {
804 * int key;
805 * void * my_member;
806 * struct rhash_head node;
807 * };
808 *
809 * struct rhashtable_params params = {
810 * .head_offset = offsetof(struct test_obj, node),
811 * .key_offset = offsetof(struct test_obj, key),
812 * .key_len = sizeof(int),
813 * .hashfn = jhash,
814 * .nulls_base = (1U << RHT_BASE_SHIFT),
815 * };
816 *
817 * Configuration Example 2: Variable length keys
818 * struct test_obj {
819 * [...]
820 * struct rhash_head node;
821 * };
822 *
823 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
824 * {
825 * struct test_obj *obj = data;
826 *
827 * return [... hash ...];
828 * }
829 *
830 * struct rhashtable_params params = {
831 * .head_offset = offsetof(struct test_obj, node),
832 * .hashfn = jhash,
833 * .obj_hashfn = my_hash_fn,
834 * };
835 */
836int rhashtable_init(struct rhashtable *ht,
837 const struct rhashtable_params *params)
838{
839 struct bucket_table *tbl;
840 size_t size;
841
842 size = HASH_DEFAULT_SIZE;
843
844 if ((!params->key_len && !params->obj_hashfn) ||
845 (params->obj_hashfn && !params->obj_cmpfn))
846 return -EINVAL;
847
848 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
849 return -EINVAL;
850
851 memset(ht, 0, sizeof(*ht));
852 mutex_init(&ht->mutex);
853 spin_lock_init(&ht->lock);
854 memcpy(&ht->p, params, sizeof(*params));
855
856 if (params->min_size)
857 ht->p.min_size = roundup_pow_of_two(params->min_size);
858
859 if (params->max_size)
860 ht->p.max_size = rounddown_pow_of_two(params->max_size);
861
862 if (params->insecure_max_entries)
863 ht->p.insecure_max_entries =
864 rounddown_pow_of_two(params->insecure_max_entries);
865 else
866 ht->p.insecure_max_entries = ht->p.max_size * 2;
867
868 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
869
870 if (params->nelem_hint)
871 size = rounded_hashtable_size(&ht->p);
872
873 /* The maximum (not average) chain length grows with the
874 * size of the hash table, at a rate of (log N)/(log log N).
875 * The value of 16 is selected so that even if the hash
876 * table grew to 2^32 you would not expect the maximum
877 * chain length to exceed it unless we are under attack
878 * (or extremely unlucky).
879 *
880 * As this limit is only to detect attacks, we don't need
881 * to set it to a lower value as you'd need the chain
882 * length to vastly exceed 16 to have any real effect
883 * on the system.
884 */
885 if (!params->insecure_elasticity)
886 ht->elasticity = 16;
887
888 if (params->locks_mul)
889 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
890 else
891 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
892
893 ht->key_len = ht->p.key_len;
894 if (!params->hashfn) {
895 ht->p.hashfn = jhash;
896
897 if (!(ht->key_len & (sizeof(u32) - 1))) {
898 ht->key_len /= sizeof(u32);
899 ht->p.hashfn = rhashtable_jhash2;
900 }
901 }
902
903 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
904 if (tbl == NULL)
905 return -ENOMEM;
906
907 atomic_set(&ht->nelems, 0);
908
909 RCU_INIT_POINTER(ht->tbl, tbl);
910
911 INIT_WORK(&ht->run_work, rht_deferred_worker);
912
913 return 0;
914}
915EXPORT_SYMBOL_GPL(rhashtable_init);
916
917/**
918 * rhltable_init - initialize a new hash list table
919 * @hlt: hash list table to be initialized
920 * @params: configuration parameters
921 *
922 * Initializes a new hash list table.
923 *
924 * See documentation for rhashtable_init.
925 */
926int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
927{
928 int err;
929
930 /* No rhlist NULLs marking for now. */
931 if (params->nulls_base)
932 return -EINVAL;
933
934 err = rhashtable_init(&hlt->ht, params);
935 hlt->ht.rhlist = true;
936 return err;
937}
938EXPORT_SYMBOL_GPL(rhltable_init);
939
940static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
941 void (*free_fn)(void *ptr, void *arg),
942 void *arg)
943{
944 struct rhlist_head *list;
945
946 if (!ht->rhlist) {
947 free_fn(rht_obj(ht, obj), arg);
948 return;
949 }
950
951 list = container_of(obj, struct rhlist_head, rhead);
952 do {
953 obj = &list->rhead;
954 list = rht_dereference(list->next, ht);
955 free_fn(rht_obj(ht, obj), arg);
956 } while (list);
957}
958
959/**
960 * rhashtable_free_and_destroy - free elements and destroy hash table
961 * @ht: the hash table to destroy
962 * @free_fn: callback to release resources of element
963 * @arg: pointer passed to free_fn
964 *
965 * Stops an eventual async resize. If defined, invokes free_fn for each
966 * element to releasal resources. Please note that RCU protected
967 * readers may still be accessing the elements. Releasing of resources
968 * must occur in a compatible manner. Then frees the bucket array.
969 *
970 * This function will eventually sleep to wait for an async resize
971 * to complete. The caller is responsible that no further write operations
972 * occurs in parallel.
973 */
974void rhashtable_free_and_destroy(struct rhashtable *ht,
975 void (*free_fn)(void *ptr, void *arg),
976 void *arg)
977{
978 const struct bucket_table *tbl;
979 unsigned int i;
980
981 cancel_work_sync(&ht->run_work);
982
983 mutex_lock(&ht->mutex);
984 tbl = rht_dereference(ht->tbl, ht);
985 if (free_fn) {
986 for (i = 0; i < tbl->size; i++) {
987 struct rhash_head *pos, *next;
988
989 for (pos = rht_dereference(tbl->buckets[i], ht),
990 next = !rht_is_a_nulls(pos) ?
991 rht_dereference(pos->next, ht) : NULL;
992 !rht_is_a_nulls(pos);
993 pos = next,
994 next = !rht_is_a_nulls(pos) ?
995 rht_dereference(pos->next, ht) : NULL)
996 rhashtable_free_one(ht, pos, free_fn, arg);
997 }
998 }
999
1000 bucket_table_free(tbl);
1001 mutex_unlock(&ht->mutex);
1002}
1003EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1004
1005void rhashtable_destroy(struct rhashtable *ht)
1006{
1007 return rhashtable_free_and_destroy(ht, NULL, NULL);
1008}
1009EXPORT_SYMBOL_GPL(rhashtable_destroy);
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
36union nested_table {
37 union nested_table __rcu *table;
38 struct rhash_head __rcu *bucket;
39};
40
41static 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
51int lockdep_rht_mutex_is_held(struct rhashtable *ht)
52{
53 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
54}
55EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
56
57int 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}
63EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
64#else
65#define ASSERT_RHT_MUTEX(HT)
66#endif
67
68static 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
87static 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
102static 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
111static void bucket_table_free_rcu(struct rcu_head *head)
112{
113 bucket_table_free(container_of(head, struct bucket_table, rcu));
114}
115
116static 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
141static 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
169static 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
214static 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
227static 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
271out:
272 return err;
273}
274
275static 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
297static 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
320static 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 cond_resched();
337 }
338
339 /* Publish the new table pointer. */
340 rcu_assign_pointer(ht->tbl, new_tbl);
341
342 spin_lock(&ht->lock);
343 list_for_each_entry(walker, &old_tbl->walkers, list)
344 walker->tbl = NULL;
345 spin_unlock(&ht->lock);
346
347 /* Wait for readers. All new readers will see the new
348 * table, and thus no references to the old table will
349 * remain.
350 */
351 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
352
353 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
354}
355
356static int rhashtable_rehash_alloc(struct rhashtable *ht,
357 struct bucket_table *old_tbl,
358 unsigned int size)
359{
360 struct bucket_table *new_tbl;
361 int err;
362
363 ASSERT_RHT_MUTEX(ht);
364
365 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
366 if (new_tbl == NULL)
367 return -ENOMEM;
368
369 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
370 if (err)
371 bucket_table_free(new_tbl);
372
373 return err;
374}
375
376/**
377 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
378 * @ht: the hash table to shrink
379 *
380 * This function shrinks the hash table to fit, i.e., the smallest
381 * size would not cause it to expand right away automatically.
382 *
383 * The caller must ensure that no concurrent resizing occurs by holding
384 * ht->mutex.
385 *
386 * The caller must ensure that no concurrent table mutations take place.
387 * It is however valid to have concurrent lookups if they are RCU protected.
388 *
389 * It is valid to have concurrent insertions and deletions protected by per
390 * bucket locks or concurrent RCU protected lookups and traversals.
391 */
392static int rhashtable_shrink(struct rhashtable *ht)
393{
394 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
395 unsigned int nelems = atomic_read(&ht->nelems);
396 unsigned int size = 0;
397
398 if (nelems)
399 size = roundup_pow_of_two(nelems * 3 / 2);
400 if (size < ht->p.min_size)
401 size = ht->p.min_size;
402
403 if (old_tbl->size <= size)
404 return 0;
405
406 if (rht_dereference(old_tbl->future_tbl, ht))
407 return -EEXIST;
408
409 return rhashtable_rehash_alloc(ht, old_tbl, size);
410}
411
412static void rht_deferred_worker(struct work_struct *work)
413{
414 struct rhashtable *ht;
415 struct bucket_table *tbl;
416 int err = 0;
417
418 ht = container_of(work, struct rhashtable, run_work);
419 mutex_lock(&ht->mutex);
420
421 tbl = rht_dereference(ht->tbl, ht);
422 tbl = rhashtable_last_table(ht, tbl);
423
424 if (rht_grow_above_75(ht, tbl))
425 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
426 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
427 err = rhashtable_shrink(ht);
428 else if (tbl->nest)
429 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
430
431 if (!err)
432 err = rhashtable_rehash_table(ht);
433
434 mutex_unlock(&ht->mutex);
435
436 if (err)
437 schedule_work(&ht->run_work);
438}
439
440static int rhashtable_insert_rehash(struct rhashtable *ht,
441 struct bucket_table *tbl)
442{
443 struct bucket_table *old_tbl;
444 struct bucket_table *new_tbl;
445 unsigned int size;
446 int err;
447
448 old_tbl = rht_dereference_rcu(ht->tbl, ht);
449
450 size = tbl->size;
451
452 err = -EBUSY;
453
454 if (rht_grow_above_75(ht, tbl))
455 size *= 2;
456 /* Do not schedule more than one rehash */
457 else if (old_tbl != tbl)
458 goto fail;
459
460 err = -ENOMEM;
461
462 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
463 if (new_tbl == NULL)
464 goto fail;
465
466 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
467 if (err) {
468 bucket_table_free(new_tbl);
469 if (err == -EEXIST)
470 err = 0;
471 } else
472 schedule_work(&ht->run_work);
473
474 return err;
475
476fail:
477 /* Do not fail the insert if someone else did a rehash. */
478 if (likely(rcu_dereference_raw(tbl->future_tbl)))
479 return 0;
480
481 /* Schedule async rehash to retry allocation in process context. */
482 if (err == -ENOMEM)
483 schedule_work(&ht->run_work);
484
485 return err;
486}
487
488static void *rhashtable_lookup_one(struct rhashtable *ht,
489 struct bucket_table *tbl, unsigned int hash,
490 const void *key, struct rhash_head *obj)
491{
492 struct rhashtable_compare_arg arg = {
493 .ht = ht,
494 .key = key,
495 };
496 struct rhash_head __rcu **pprev;
497 struct rhash_head *head;
498 int elasticity;
499
500 elasticity = RHT_ELASTICITY;
501 pprev = rht_bucket_var(tbl, hash);
502 rht_for_each_continue(head, *pprev, tbl, hash) {
503 struct rhlist_head *list;
504 struct rhlist_head *plist;
505
506 elasticity--;
507 if (!key ||
508 (ht->p.obj_cmpfn ?
509 ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
510 rhashtable_compare(&arg, rht_obj(ht, head)))) {
511 pprev = &head->next;
512 continue;
513 }
514
515 if (!ht->rhlist)
516 return rht_obj(ht, head);
517
518 list = container_of(obj, struct rhlist_head, rhead);
519 plist = container_of(head, struct rhlist_head, rhead);
520
521 RCU_INIT_POINTER(list->next, plist);
522 head = rht_dereference_bucket(head->next, tbl, hash);
523 RCU_INIT_POINTER(list->rhead.next, head);
524 rcu_assign_pointer(*pprev, obj);
525
526 return NULL;
527 }
528
529 if (elasticity <= 0)
530 return ERR_PTR(-EAGAIN);
531
532 return ERR_PTR(-ENOENT);
533}
534
535static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
536 struct bucket_table *tbl,
537 unsigned int hash,
538 struct rhash_head *obj,
539 void *data)
540{
541 struct rhash_head __rcu **pprev;
542 struct bucket_table *new_tbl;
543 struct rhash_head *head;
544
545 if (!IS_ERR_OR_NULL(data))
546 return ERR_PTR(-EEXIST);
547
548 if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
549 return ERR_CAST(data);
550
551 new_tbl = rcu_dereference(tbl->future_tbl);
552 if (new_tbl)
553 return new_tbl;
554
555 if (PTR_ERR(data) != -ENOENT)
556 return ERR_CAST(data);
557
558 if (unlikely(rht_grow_above_max(ht, tbl)))
559 return ERR_PTR(-E2BIG);
560
561 if (unlikely(rht_grow_above_100(ht, tbl)))
562 return ERR_PTR(-EAGAIN);
563
564 pprev = rht_bucket_insert(ht, tbl, hash);
565 if (!pprev)
566 return ERR_PTR(-ENOMEM);
567
568 head = rht_dereference_bucket(*pprev, tbl, hash);
569
570 RCU_INIT_POINTER(obj->next, head);
571 if (ht->rhlist) {
572 struct rhlist_head *list;
573
574 list = container_of(obj, struct rhlist_head, rhead);
575 RCU_INIT_POINTER(list->next, NULL);
576 }
577
578 rcu_assign_pointer(*pprev, obj);
579
580 atomic_inc(&ht->nelems);
581 if (rht_grow_above_75(ht, tbl))
582 schedule_work(&ht->run_work);
583
584 return NULL;
585}
586
587static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
588 struct rhash_head *obj)
589{
590 struct bucket_table *new_tbl;
591 struct bucket_table *tbl;
592 unsigned int hash;
593 spinlock_t *lock;
594 void *data;
595
596 tbl = rcu_dereference(ht->tbl);
597
598 /* All insertions must grab the oldest table containing
599 * the hashed bucket that is yet to be rehashed.
600 */
601 for (;;) {
602 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
603 lock = rht_bucket_lock(tbl, hash);
604 spin_lock_bh(lock);
605
606 if (tbl->rehash <= hash)
607 break;
608
609 spin_unlock_bh(lock);
610 tbl = rcu_dereference(tbl->future_tbl);
611 }
612
613 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
614 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
615 if (PTR_ERR(new_tbl) != -EEXIST)
616 data = ERR_CAST(new_tbl);
617
618 while (!IS_ERR_OR_NULL(new_tbl)) {
619 tbl = new_tbl;
620 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
621 spin_lock_nested(rht_bucket_lock(tbl, hash),
622 SINGLE_DEPTH_NESTING);
623
624 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
625 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
626 if (PTR_ERR(new_tbl) != -EEXIST)
627 data = ERR_CAST(new_tbl);
628
629 spin_unlock(rht_bucket_lock(tbl, hash));
630 }
631
632 spin_unlock_bh(lock);
633
634 if (PTR_ERR(data) == -EAGAIN)
635 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
636 -EAGAIN);
637
638 return data;
639}
640
641void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
642 struct rhash_head *obj)
643{
644 void *data;
645
646 do {
647 rcu_read_lock();
648 data = rhashtable_try_insert(ht, key, obj);
649 rcu_read_unlock();
650 } while (PTR_ERR(data) == -EAGAIN);
651
652 return data;
653}
654EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
655
656/**
657 * rhashtable_walk_enter - Initialise an iterator
658 * @ht: Table to walk over
659 * @iter: Hash table Iterator
660 *
661 * This function prepares a hash table walk.
662 *
663 * Note that if you restart a walk after rhashtable_walk_stop you
664 * may see the same object twice. Also, you may miss objects if
665 * there are removals in between rhashtable_walk_stop and the next
666 * call to rhashtable_walk_start.
667 *
668 * For a completely stable walk you should construct your own data
669 * structure outside the hash table.
670 *
671 * This function may sleep so you must not call it from interrupt
672 * context or with spin locks held.
673 *
674 * You must call rhashtable_walk_exit after this function returns.
675 */
676void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
677{
678 iter->ht = ht;
679 iter->p = NULL;
680 iter->slot = 0;
681 iter->skip = 0;
682 iter->end_of_table = 0;
683
684 spin_lock(&ht->lock);
685 iter->walker.tbl =
686 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
687 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
688 spin_unlock(&ht->lock);
689}
690EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
691
692/**
693 * rhashtable_walk_exit - Free an iterator
694 * @iter: Hash table Iterator
695 *
696 * This function frees resources allocated by rhashtable_walk_init.
697 */
698void rhashtable_walk_exit(struct rhashtable_iter *iter)
699{
700 spin_lock(&iter->ht->lock);
701 if (iter->walker.tbl)
702 list_del(&iter->walker.list);
703 spin_unlock(&iter->ht->lock);
704}
705EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
706
707/**
708 * rhashtable_walk_start_check - Start a hash table walk
709 * @iter: Hash table iterator
710 *
711 * Start a hash table walk at the current iterator position. Note that we take
712 * the RCU lock in all cases including when we return an error. So you must
713 * always call rhashtable_walk_stop to clean up.
714 *
715 * Returns zero if successful.
716 *
717 * Returns -EAGAIN if resize event occured. Note that the iterator
718 * will rewind back to the beginning and you may use it immediately
719 * by calling rhashtable_walk_next.
720 *
721 * rhashtable_walk_start is defined as an inline variant that returns
722 * void. This is preferred in cases where the caller would ignore
723 * resize events and always continue.
724 */
725int rhashtable_walk_start_check(struct rhashtable_iter *iter)
726 __acquires(RCU)
727{
728 struct rhashtable *ht = iter->ht;
729
730 rcu_read_lock();
731
732 spin_lock(&ht->lock);
733 if (iter->walker.tbl)
734 list_del(&iter->walker.list);
735 spin_unlock(&ht->lock);
736
737 if (!iter->walker.tbl && !iter->end_of_table) {
738 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
739 return -EAGAIN;
740 }
741
742 return 0;
743}
744EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
745
746/**
747 * __rhashtable_walk_find_next - Find the next element in a table (or the first
748 * one in case of a new walk).
749 *
750 * @iter: Hash table iterator
751 *
752 * Returns the found object or NULL when the end of the table is reached.
753 *
754 * Returns -EAGAIN if resize event occurred.
755 */
756static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
757{
758 struct bucket_table *tbl = iter->walker.tbl;
759 struct rhlist_head *list = iter->list;
760 struct rhashtable *ht = iter->ht;
761 struct rhash_head *p = iter->p;
762 bool rhlist = ht->rhlist;
763
764 if (!tbl)
765 return NULL;
766
767 for (; iter->slot < tbl->size; iter->slot++) {
768 int skip = iter->skip;
769
770 rht_for_each_rcu(p, tbl, iter->slot) {
771 if (rhlist) {
772 list = container_of(p, struct rhlist_head,
773 rhead);
774 do {
775 if (!skip)
776 goto next;
777 skip--;
778 list = rcu_dereference(list->next);
779 } while (list);
780
781 continue;
782 }
783 if (!skip)
784 break;
785 skip--;
786 }
787
788next:
789 if (!rht_is_a_nulls(p)) {
790 iter->skip++;
791 iter->p = p;
792 iter->list = list;
793 return rht_obj(ht, rhlist ? &list->rhead : p);
794 }
795
796 iter->skip = 0;
797 }
798
799 iter->p = NULL;
800
801 /* Ensure we see any new tables. */
802 smp_rmb();
803
804 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
805 if (iter->walker.tbl) {
806 iter->slot = 0;
807 iter->skip = 0;
808 return ERR_PTR(-EAGAIN);
809 } else {
810 iter->end_of_table = true;
811 }
812
813 return NULL;
814}
815
816/**
817 * rhashtable_walk_next - Return the next object and advance the iterator
818 * @iter: Hash table iterator
819 *
820 * Note that you must call rhashtable_walk_stop when you are finished
821 * with the walk.
822 *
823 * Returns the next object or NULL when the end of the table is reached.
824 *
825 * Returns -EAGAIN if resize event occurred. Note that the iterator
826 * will rewind back to the beginning and you may continue to use it.
827 */
828void *rhashtable_walk_next(struct rhashtable_iter *iter)
829{
830 struct rhlist_head *list = iter->list;
831 struct rhashtable *ht = iter->ht;
832 struct rhash_head *p = iter->p;
833 bool rhlist = ht->rhlist;
834
835 if (p) {
836 if (!rhlist || !(list = rcu_dereference(list->next))) {
837 p = rcu_dereference(p->next);
838 list = container_of(p, struct rhlist_head, rhead);
839 }
840 if (!rht_is_a_nulls(p)) {
841 iter->skip++;
842 iter->p = p;
843 iter->list = list;
844 return rht_obj(ht, rhlist ? &list->rhead : p);
845 }
846
847 /* At the end of this slot, switch to next one and then find
848 * next entry from that point.
849 */
850 iter->skip = 0;
851 iter->slot++;
852 }
853
854 return __rhashtable_walk_find_next(iter);
855}
856EXPORT_SYMBOL_GPL(rhashtable_walk_next);
857
858/**
859 * rhashtable_walk_peek - Return the next object but don't advance the iterator
860 * @iter: Hash table iterator
861 *
862 * Returns the next object or NULL when the end of the table is reached.
863 *
864 * Returns -EAGAIN if resize event occurred. Note that the iterator
865 * will rewind back to the beginning and you may continue to use it.
866 */
867void *rhashtable_walk_peek(struct rhashtable_iter *iter)
868{
869 struct rhlist_head *list = iter->list;
870 struct rhashtable *ht = iter->ht;
871 struct rhash_head *p = iter->p;
872
873 if (p)
874 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
875
876 /* No object found in current iter, find next one in the table. */
877
878 if (iter->skip) {
879 /* A nonzero skip value points to the next entry in the table
880 * beyond that last one that was found. Decrement skip so
881 * we find the current value. __rhashtable_walk_find_next
882 * will restore the original value of skip assuming that
883 * the table hasn't changed.
884 */
885 iter->skip--;
886 }
887
888 return __rhashtable_walk_find_next(iter);
889}
890EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
891
892/**
893 * rhashtable_walk_stop - Finish a hash table walk
894 * @iter: Hash table iterator
895 *
896 * Finish a hash table walk. Does not reset the iterator to the start of the
897 * hash table.
898 */
899void rhashtable_walk_stop(struct rhashtable_iter *iter)
900 __releases(RCU)
901{
902 struct rhashtable *ht;
903 struct bucket_table *tbl = iter->walker.tbl;
904
905 if (!tbl)
906 goto out;
907
908 ht = iter->ht;
909
910 spin_lock(&ht->lock);
911 if (tbl->rehash < tbl->size)
912 list_add(&iter->walker.list, &tbl->walkers);
913 else
914 iter->walker.tbl = NULL;
915 spin_unlock(&ht->lock);
916
917 iter->p = NULL;
918
919out:
920 rcu_read_unlock();
921}
922EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
923
924static size_t rounded_hashtable_size(const struct rhashtable_params *params)
925{
926 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
927 (unsigned long)params->min_size);
928}
929
930static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
931{
932 return jhash2(key, length, seed);
933}
934
935/**
936 * rhashtable_init - initialize a new hash table
937 * @ht: hash table to be initialized
938 * @params: configuration parameters
939 *
940 * Initializes a new hash table based on the provided configuration
941 * parameters. A table can be configured either with a variable or
942 * fixed length key:
943 *
944 * Configuration Example 1: Fixed length keys
945 * struct test_obj {
946 * int key;
947 * void * my_member;
948 * struct rhash_head node;
949 * };
950 *
951 * struct rhashtable_params params = {
952 * .head_offset = offsetof(struct test_obj, node),
953 * .key_offset = offsetof(struct test_obj, key),
954 * .key_len = sizeof(int),
955 * .hashfn = jhash,
956 * .nulls_base = (1U << RHT_BASE_SHIFT),
957 * };
958 *
959 * Configuration Example 2: Variable length keys
960 * struct test_obj {
961 * [...]
962 * struct rhash_head node;
963 * };
964 *
965 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
966 * {
967 * struct test_obj *obj = data;
968 *
969 * return [... hash ...];
970 * }
971 *
972 * struct rhashtable_params params = {
973 * .head_offset = offsetof(struct test_obj, node),
974 * .hashfn = jhash,
975 * .obj_hashfn = my_hash_fn,
976 * };
977 */
978int rhashtable_init(struct rhashtable *ht,
979 const struct rhashtable_params *params)
980{
981 struct bucket_table *tbl;
982 size_t size;
983
984 size = HASH_DEFAULT_SIZE;
985
986 if ((!params->key_len && !params->obj_hashfn) ||
987 (params->obj_hashfn && !params->obj_cmpfn))
988 return -EINVAL;
989
990 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
991 return -EINVAL;
992
993 memset(ht, 0, sizeof(*ht));
994 mutex_init(&ht->mutex);
995 spin_lock_init(&ht->lock);
996 memcpy(&ht->p, params, sizeof(*params));
997
998 if (params->min_size)
999 ht->p.min_size = roundup_pow_of_two(params->min_size);
1000
1001 /* Cap total entries at 2^31 to avoid nelems overflow. */
1002 ht->max_elems = 1u << 31;
1003
1004 if (params->max_size) {
1005 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1006 if (ht->p.max_size < ht->max_elems / 2)
1007 ht->max_elems = ht->p.max_size * 2;
1008 }
1009
1010 ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1011
1012 if (params->nelem_hint)
1013 size = rounded_hashtable_size(&ht->p);
1014
1015 if (params->locks_mul)
1016 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1017 else
1018 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1019
1020 ht->key_len = ht->p.key_len;
1021 if (!params->hashfn) {
1022 ht->p.hashfn = jhash;
1023
1024 if (!(ht->key_len & (sizeof(u32) - 1))) {
1025 ht->key_len /= sizeof(u32);
1026 ht->p.hashfn = rhashtable_jhash2;
1027 }
1028 }
1029
1030 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1031 if (tbl == NULL)
1032 return -ENOMEM;
1033
1034 atomic_set(&ht->nelems, 0);
1035
1036 RCU_INIT_POINTER(ht->tbl, tbl);
1037
1038 INIT_WORK(&ht->run_work, rht_deferred_worker);
1039
1040 return 0;
1041}
1042EXPORT_SYMBOL_GPL(rhashtable_init);
1043
1044/**
1045 * rhltable_init - initialize a new hash list table
1046 * @hlt: hash list table to be initialized
1047 * @params: configuration parameters
1048 *
1049 * Initializes a new hash list table.
1050 *
1051 * See documentation for rhashtable_init.
1052 */
1053int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1054{
1055 int err;
1056
1057 /* No rhlist NULLs marking for now. */
1058 if (params->nulls_base)
1059 return -EINVAL;
1060
1061 err = rhashtable_init(&hlt->ht, params);
1062 hlt->ht.rhlist = true;
1063 return err;
1064}
1065EXPORT_SYMBOL_GPL(rhltable_init);
1066
1067static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1068 void (*free_fn)(void *ptr, void *arg),
1069 void *arg)
1070{
1071 struct rhlist_head *list;
1072
1073 if (!ht->rhlist) {
1074 free_fn(rht_obj(ht, obj), arg);
1075 return;
1076 }
1077
1078 list = container_of(obj, struct rhlist_head, rhead);
1079 do {
1080 obj = &list->rhead;
1081 list = rht_dereference(list->next, ht);
1082 free_fn(rht_obj(ht, obj), arg);
1083 } while (list);
1084}
1085
1086/**
1087 * rhashtable_free_and_destroy - free elements and destroy hash table
1088 * @ht: the hash table to destroy
1089 * @free_fn: callback to release resources of element
1090 * @arg: pointer passed to free_fn
1091 *
1092 * Stops an eventual async resize. If defined, invokes free_fn for each
1093 * element to releasal resources. Please note that RCU protected
1094 * readers may still be accessing the elements. Releasing of resources
1095 * must occur in a compatible manner. Then frees the bucket array.
1096 *
1097 * This function will eventually sleep to wait for an async resize
1098 * to complete. The caller is responsible that no further write operations
1099 * occurs in parallel.
1100 */
1101void rhashtable_free_and_destroy(struct rhashtable *ht,
1102 void (*free_fn)(void *ptr, void *arg),
1103 void *arg)
1104{
1105 struct bucket_table *tbl;
1106 unsigned int i;
1107
1108 cancel_work_sync(&ht->run_work);
1109
1110 mutex_lock(&ht->mutex);
1111 tbl = rht_dereference(ht->tbl, ht);
1112 if (free_fn) {
1113 for (i = 0; i < tbl->size; i++) {
1114 struct rhash_head *pos, *next;
1115
1116 cond_resched();
1117 for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1118 next = !rht_is_a_nulls(pos) ?
1119 rht_dereference(pos->next, ht) : NULL;
1120 !rht_is_a_nulls(pos);
1121 pos = next,
1122 next = !rht_is_a_nulls(pos) ?
1123 rht_dereference(pos->next, ht) : NULL)
1124 rhashtable_free_one(ht, pos, free_fn, arg);
1125 }
1126 }
1127
1128 bucket_table_free(tbl);
1129 mutex_unlock(&ht->mutex);
1130}
1131EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1132
1133void rhashtable_destroy(struct rhashtable *ht)
1134{
1135 return rhashtable_free_and_destroy(ht, NULL, NULL);
1136}
1137EXPORT_SYMBOL_GPL(rhashtable_destroy);
1138
1139struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1140 unsigned int hash)
1141{
1142 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1143 static struct rhash_head __rcu *rhnull =
1144 (struct rhash_head __rcu *)NULLS_MARKER(0);
1145 unsigned int index = hash & ((1 << tbl->nest) - 1);
1146 unsigned int size = tbl->size >> tbl->nest;
1147 unsigned int subhash = hash;
1148 union nested_table *ntbl;
1149
1150 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1151 ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1152 subhash >>= tbl->nest;
1153
1154 while (ntbl && size > (1 << shift)) {
1155 index = subhash & ((1 << shift) - 1);
1156 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1157 tbl, hash);
1158 size >>= shift;
1159 subhash >>= shift;
1160 }
1161
1162 if (!ntbl)
1163 return &rhnull;
1164
1165 return &ntbl[subhash].bucket;
1166
1167}
1168EXPORT_SYMBOL_GPL(rht_bucket_nested);
1169
1170struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1171 struct bucket_table *tbl,
1172 unsigned int hash)
1173{
1174 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1175 unsigned int index = hash & ((1 << tbl->nest) - 1);
1176 unsigned int size = tbl->size >> tbl->nest;
1177 union nested_table *ntbl;
1178 unsigned int shifted;
1179 unsigned int nhash;
1180
1181 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1182 hash >>= tbl->nest;
1183 nhash = index;
1184 shifted = tbl->nest;
1185 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1186 size <= (1 << shift) ? shifted : 0, nhash);
1187
1188 while (ntbl && size > (1 << shift)) {
1189 index = hash & ((1 << shift) - 1);
1190 size >>= shift;
1191 hash >>= shift;
1192 nhash |= index << shifted;
1193 shifted += shift;
1194 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1195 size <= (1 << shift) ? shifted : 0,
1196 nhash);
1197 }
1198
1199 if (!ntbl)
1200 return NULL;
1201
1202 return &ntbl[hash].bucket;
1203
1204}
1205EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);