xref: /linux/lib/rhashtable.c (revision 3ce095c16263630dde46d6051854073edaacf3d7)
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   128UL
34 
35 static 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 
45 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
46 {
47 	return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
48 }
49 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
50 
51 int 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 }
57 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
58 #else
59 #define ASSERT_RHT_MUTEX(HT)
60 #endif
61 
62 
63 static 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, 32UL);
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 #ifdef CONFIG_NUMA
81 		if (size * sizeof(spinlock_t) > PAGE_SIZE &&
82 		    gfp == GFP_KERNEL)
83 			tbl->locks = vmalloc(size * sizeof(spinlock_t));
84 		else
85 #endif
86 		tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
87 					   gfp);
88 		if (!tbl->locks)
89 			return -ENOMEM;
90 		for (i = 0; i < size; i++)
91 			spin_lock_init(&tbl->locks[i]);
92 	}
93 	tbl->locks_mask = size - 1;
94 
95 	return 0;
96 }
97 
98 static void bucket_table_free(const struct bucket_table *tbl)
99 {
100 	if (tbl)
101 		kvfree(tbl->locks);
102 
103 	kvfree(tbl);
104 }
105 
106 static void bucket_table_free_rcu(struct rcu_head *head)
107 {
108 	bucket_table_free(container_of(head, struct bucket_table, rcu));
109 }
110 
111 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
112 					       size_t nbuckets,
113 					       gfp_t gfp)
114 {
115 	struct bucket_table *tbl = NULL;
116 	size_t size;
117 	int i;
118 
119 	size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
120 	if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
121 	    gfp != GFP_KERNEL)
122 		tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
123 	if (tbl == NULL && gfp == GFP_KERNEL)
124 		tbl = vzalloc(size);
125 	if (tbl == NULL)
126 		return NULL;
127 
128 	tbl->size = nbuckets;
129 
130 	if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
131 		bucket_table_free(tbl);
132 		return NULL;
133 	}
134 
135 	INIT_LIST_HEAD(&tbl->walkers);
136 
137 	get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
138 
139 	for (i = 0; i < nbuckets; i++)
140 		INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
141 
142 	return tbl;
143 }
144 
145 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
146 						  struct bucket_table *tbl)
147 {
148 	struct bucket_table *new_tbl;
149 
150 	do {
151 		new_tbl = tbl;
152 		tbl = rht_dereference_rcu(tbl->future_tbl, ht);
153 	} while (tbl);
154 
155 	return new_tbl;
156 }
157 
158 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
159 {
160 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
161 	struct bucket_table *new_tbl = rhashtable_last_table(ht,
162 		rht_dereference_rcu(old_tbl->future_tbl, ht));
163 	struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
164 	int err = -ENOENT;
165 	struct rhash_head *head, *next, *entry;
166 	spinlock_t *new_bucket_lock;
167 	unsigned int new_hash;
168 
169 	rht_for_each(entry, old_tbl, old_hash) {
170 		err = 0;
171 		next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
172 
173 		if (rht_is_a_nulls(next))
174 			break;
175 
176 		pprev = &entry->next;
177 	}
178 
179 	if (err)
180 		goto out;
181 
182 	new_hash = head_hashfn(ht, new_tbl, entry);
183 
184 	new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
185 
186 	spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
187 	head = rht_dereference_bucket(new_tbl->buckets[new_hash],
188 				      new_tbl, new_hash);
189 
190 	if (rht_is_a_nulls(head))
191 		INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash);
192 	else
193 		RCU_INIT_POINTER(entry->next, head);
194 
195 	rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
196 	spin_unlock(new_bucket_lock);
197 
198 	rcu_assign_pointer(*pprev, next);
199 
200 out:
201 	return err;
202 }
203 
204 static void rhashtable_rehash_chain(struct rhashtable *ht,
205 				    unsigned int old_hash)
206 {
207 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
208 	spinlock_t *old_bucket_lock;
209 
210 	old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
211 
212 	spin_lock_bh(old_bucket_lock);
213 	while (!rhashtable_rehash_one(ht, old_hash))
214 		;
215 	old_tbl->rehash++;
216 	spin_unlock_bh(old_bucket_lock);
217 }
218 
219 static int rhashtable_rehash_attach(struct rhashtable *ht,
220 				    struct bucket_table *old_tbl,
221 				    struct bucket_table *new_tbl)
222 {
223 	/* Protect future_tbl using the first bucket lock. */
224 	spin_lock_bh(old_tbl->locks);
225 
226 	/* Did somebody beat us to it? */
227 	if (rcu_access_pointer(old_tbl->future_tbl)) {
228 		spin_unlock_bh(old_tbl->locks);
229 		return -EEXIST;
230 	}
231 
232 	/* Make insertions go into the new, empty table right away. Deletions
233 	 * and lookups will be attempted in both tables until we synchronize.
234 	 */
235 	rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
236 
237 	/* Ensure the new table is visible to readers. */
238 	smp_wmb();
239 
240 	spin_unlock_bh(old_tbl->locks);
241 
242 	return 0;
243 }
244 
245 static int rhashtable_rehash_table(struct rhashtable *ht)
246 {
247 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
248 	struct bucket_table *new_tbl;
249 	struct rhashtable_walker *walker;
250 	unsigned int old_hash;
251 
252 	new_tbl = rht_dereference(old_tbl->future_tbl, ht);
253 	if (!new_tbl)
254 		return 0;
255 
256 	for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
257 		rhashtable_rehash_chain(ht, old_hash);
258 
259 	/* Publish the new table pointer. */
260 	rcu_assign_pointer(ht->tbl, new_tbl);
261 
262 	spin_lock(&ht->lock);
263 	list_for_each_entry(walker, &old_tbl->walkers, list)
264 		walker->tbl = NULL;
265 	spin_unlock(&ht->lock);
266 
267 	/* Wait for readers. All new readers will see the new
268 	 * table, and thus no references to the old table will
269 	 * remain.
270 	 */
271 	call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
272 
273 	return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
274 }
275 
276 /**
277  * rhashtable_expand - Expand hash table while allowing concurrent lookups
278  * @ht:		the hash table to expand
279  *
280  * A secondary bucket array is allocated and the hash entries are migrated.
281  *
282  * This function may only be called in a context where it is safe to call
283  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
284  *
285  * The caller must ensure that no concurrent resizing occurs by holding
286  * ht->mutex.
287  *
288  * It is valid to have concurrent insertions and deletions protected by per
289  * bucket locks or concurrent RCU protected lookups and traversals.
290  */
291 static int rhashtable_expand(struct rhashtable *ht)
292 {
293 	struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
294 	int err;
295 
296 	ASSERT_RHT_MUTEX(ht);
297 
298 	old_tbl = rhashtable_last_table(ht, old_tbl);
299 
300 	new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
301 	if (new_tbl == NULL)
302 		return -ENOMEM;
303 
304 	err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
305 	if (err)
306 		bucket_table_free(new_tbl);
307 
308 	return err;
309 }
310 
311 /**
312  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
313  * @ht:		the hash table to shrink
314  *
315  * This function shrinks the hash table to fit, i.e., the smallest
316  * size would not cause it to expand right away automatically.
317  *
318  * The caller must ensure that no concurrent resizing occurs by holding
319  * ht->mutex.
320  *
321  * The caller must ensure that no concurrent table mutations take place.
322  * It is however valid to have concurrent lookups if they are RCU protected.
323  *
324  * It is valid to have concurrent insertions and deletions protected by per
325  * bucket locks or concurrent RCU protected lookups and traversals.
326  */
327 static int rhashtable_shrink(struct rhashtable *ht)
328 {
329 	struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
330 	unsigned int size;
331 	int err;
332 
333 	ASSERT_RHT_MUTEX(ht);
334 
335 	size = roundup_pow_of_two(atomic_read(&ht->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 
356 static 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 
381 static bool rhashtable_check_elasticity(struct rhashtable *ht,
382 					struct bucket_table *tbl,
383 					unsigned int hash)
384 {
385 	unsigned int elasticity = ht->elasticity;
386 	struct rhash_head *head;
387 
388 	rht_for_each(head, tbl, hash)
389 		if (!--elasticity)
390 			return true;
391 
392 	return false;
393 }
394 
395 int rhashtable_insert_rehash(struct rhashtable *ht)
396 {
397 	struct bucket_table *old_tbl;
398 	struct bucket_table *new_tbl;
399 	struct bucket_table *tbl;
400 	unsigned int size;
401 	int err;
402 
403 	old_tbl = rht_dereference_rcu(ht->tbl, ht);
404 	tbl = rhashtable_last_table(ht, old_tbl);
405 
406 	size = tbl->size;
407 
408 	if (rht_grow_above_75(ht, tbl))
409 		size *= 2;
410 	/* Do not schedule more than one rehash */
411 	else if (old_tbl != tbl)
412 		return -EBUSY;
413 
414 	new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
415 	if (new_tbl == NULL) {
416 		/* Schedule async resize/rehash to try allocation
417 		 * non-atomic context.
418 		 */
419 		schedule_work(&ht->run_work);
420 		return -ENOMEM;
421 	}
422 
423 	err = rhashtable_rehash_attach(ht, tbl, new_tbl);
424 	if (err) {
425 		bucket_table_free(new_tbl);
426 		if (err == -EEXIST)
427 			err = 0;
428 	} else
429 		schedule_work(&ht->run_work);
430 
431 	return err;
432 }
433 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
434 
435 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
436 			   struct rhash_head *obj,
437 			   struct bucket_table *tbl)
438 {
439 	struct rhash_head *head;
440 	unsigned int hash;
441 	int err;
442 
443 	tbl = rhashtable_last_table(ht, tbl);
444 	hash = head_hashfn(ht, tbl, obj);
445 	spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
446 
447 	err = -EEXIST;
448 	if (key && rhashtable_lookup_fast(ht, key, ht->p))
449 		goto exit;
450 
451 	err = -E2BIG;
452 	if (unlikely(rht_grow_above_max(ht, tbl)))
453 		goto exit;
454 
455 	err = -EAGAIN;
456 	if (rhashtable_check_elasticity(ht, tbl, hash) ||
457 	    rht_grow_above_100(ht, tbl))
458 		goto exit;
459 
460 	err = 0;
461 
462 	head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
463 
464 	RCU_INIT_POINTER(obj->next, head);
465 
466 	rcu_assign_pointer(tbl->buckets[hash], obj);
467 
468 	atomic_inc(&ht->nelems);
469 
470 exit:
471 	spin_unlock(rht_bucket_lock(tbl, hash));
472 
473 	return err;
474 }
475 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
476 
477 /**
478  * rhashtable_walk_init - Initialise an iterator
479  * @ht:		Table to walk over
480  * @iter:	Hash table Iterator
481  *
482  * This function prepares a hash table walk.
483  *
484  * Note that if you restart a walk after rhashtable_walk_stop you
485  * may see the same object twice.  Also, you may miss objects if
486  * there are removals in between rhashtable_walk_stop and the next
487  * call to rhashtable_walk_start.
488  *
489  * For a completely stable walk you should construct your own data
490  * structure outside the hash table.
491  *
492  * This function may sleep so you must not call it from interrupt
493  * context or with spin locks held.
494  *
495  * You must call rhashtable_walk_exit if this function returns
496  * successfully.
497  */
498 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
499 {
500 	iter->ht = ht;
501 	iter->p = NULL;
502 	iter->slot = 0;
503 	iter->skip = 0;
504 
505 	iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
506 	if (!iter->walker)
507 		return -ENOMEM;
508 
509 	mutex_lock(&ht->mutex);
510 	iter->walker->tbl = rht_dereference(ht->tbl, ht);
511 	list_add(&iter->walker->list, &iter->walker->tbl->walkers);
512 	mutex_unlock(&ht->mutex);
513 
514 	return 0;
515 }
516 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
517 
518 /**
519  * rhashtable_walk_exit - Free an iterator
520  * @iter:	Hash table Iterator
521  *
522  * This function frees resources allocated by rhashtable_walk_init.
523  */
524 void rhashtable_walk_exit(struct rhashtable_iter *iter)
525 {
526 	mutex_lock(&iter->ht->mutex);
527 	if (iter->walker->tbl)
528 		list_del(&iter->walker->list);
529 	mutex_unlock(&iter->ht->mutex);
530 	kfree(iter->walker);
531 }
532 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
533 
534 /**
535  * rhashtable_walk_start - Start a hash table walk
536  * @iter:	Hash table iterator
537  *
538  * Start a hash table walk.  Note that we take the RCU lock in all
539  * cases including when we return an error.  So you must always call
540  * rhashtable_walk_stop to clean up.
541  *
542  * Returns zero if successful.
543  *
544  * Returns -EAGAIN if resize event occured.  Note that the iterator
545  * will rewind back to the beginning and you may use it immediately
546  * by calling rhashtable_walk_next.
547  */
548 int rhashtable_walk_start(struct rhashtable_iter *iter)
549 	__acquires(RCU)
550 {
551 	struct rhashtable *ht = iter->ht;
552 
553 	mutex_lock(&ht->mutex);
554 
555 	if (iter->walker->tbl)
556 		list_del(&iter->walker->list);
557 
558 	rcu_read_lock();
559 
560 	mutex_unlock(&ht->mutex);
561 
562 	if (!iter->walker->tbl) {
563 		iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
564 		return -EAGAIN;
565 	}
566 
567 	return 0;
568 }
569 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
570 
571 /**
572  * rhashtable_walk_next - Return the next object and advance the iterator
573  * @iter:	Hash table iterator
574  *
575  * Note that you must call rhashtable_walk_stop when you are finished
576  * with the walk.
577  *
578  * Returns the next object or NULL when the end of the table is reached.
579  *
580  * Returns -EAGAIN if resize event occured.  Note that the iterator
581  * will rewind back to the beginning and you may continue to use it.
582  */
583 void *rhashtable_walk_next(struct rhashtable_iter *iter)
584 {
585 	struct bucket_table *tbl = iter->walker->tbl;
586 	struct rhashtable *ht = iter->ht;
587 	struct rhash_head *p = iter->p;
588 	void *obj = NULL;
589 
590 	if (p) {
591 		p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
592 		goto next;
593 	}
594 
595 	for (; iter->slot < tbl->size; iter->slot++) {
596 		int skip = iter->skip;
597 
598 		rht_for_each_rcu(p, tbl, iter->slot) {
599 			if (!skip)
600 				break;
601 			skip--;
602 		}
603 
604 next:
605 		if (!rht_is_a_nulls(p)) {
606 			iter->skip++;
607 			iter->p = p;
608 			obj = rht_obj(ht, p);
609 			goto out;
610 		}
611 
612 		iter->skip = 0;
613 	}
614 
615 	/* Ensure we see any new tables. */
616 	smp_rmb();
617 
618 	iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
619 	if (iter->walker->tbl) {
620 		iter->slot = 0;
621 		iter->skip = 0;
622 		return ERR_PTR(-EAGAIN);
623 	}
624 
625 	iter->p = NULL;
626 
627 out:
628 
629 	return obj;
630 }
631 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
632 
633 /**
634  * rhashtable_walk_stop - Finish a hash table walk
635  * @iter:	Hash table iterator
636  *
637  * Finish a hash table walk.
638  */
639 void rhashtable_walk_stop(struct rhashtable_iter *iter)
640 	__releases(RCU)
641 {
642 	struct rhashtable *ht;
643 	struct bucket_table *tbl = iter->walker->tbl;
644 
645 	if (!tbl)
646 		goto out;
647 
648 	ht = iter->ht;
649 
650 	spin_lock(&ht->lock);
651 	if (tbl->rehash < tbl->size)
652 		list_add(&iter->walker->list, &tbl->walkers);
653 	else
654 		iter->walker->tbl = NULL;
655 	spin_unlock(&ht->lock);
656 
657 	iter->p = NULL;
658 
659 out:
660 	rcu_read_unlock();
661 }
662 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
663 
664 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
665 {
666 	return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
667 		   (unsigned long)params->min_size);
668 }
669 
670 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
671 {
672 	return jhash2(key, length, seed);
673 }
674 
675 /**
676  * rhashtable_init - initialize a new hash table
677  * @ht:		hash table to be initialized
678  * @params:	configuration parameters
679  *
680  * Initializes a new hash table based on the provided configuration
681  * parameters. A table can be configured either with a variable or
682  * fixed length key:
683  *
684  * Configuration Example 1: Fixed length keys
685  * struct test_obj {
686  *	int			key;
687  *	void *			my_member;
688  *	struct rhash_head	node;
689  * };
690  *
691  * struct rhashtable_params params = {
692  *	.head_offset = offsetof(struct test_obj, node),
693  *	.key_offset = offsetof(struct test_obj, key),
694  *	.key_len = sizeof(int),
695  *	.hashfn = jhash,
696  *	.nulls_base = (1U << RHT_BASE_SHIFT),
697  * };
698  *
699  * Configuration Example 2: Variable length keys
700  * struct test_obj {
701  *	[...]
702  *	struct rhash_head	node;
703  * };
704  *
705  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
706  * {
707  *	struct test_obj *obj = data;
708  *
709  *	return [... hash ...];
710  * }
711  *
712  * struct rhashtable_params params = {
713  *	.head_offset = offsetof(struct test_obj, node),
714  *	.hashfn = jhash,
715  *	.obj_hashfn = my_hash_fn,
716  * };
717  */
718 int rhashtable_init(struct rhashtable *ht,
719 		    const struct rhashtable_params *params)
720 {
721 	struct bucket_table *tbl;
722 	size_t size;
723 
724 	size = HASH_DEFAULT_SIZE;
725 
726 	if ((!params->key_len && !params->obj_hashfn) ||
727 	    (params->obj_hashfn && !params->obj_cmpfn))
728 		return -EINVAL;
729 
730 	if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
731 		return -EINVAL;
732 
733 	if (params->nelem_hint)
734 		size = rounded_hashtable_size(params);
735 
736 	memset(ht, 0, sizeof(*ht));
737 	mutex_init(&ht->mutex);
738 	spin_lock_init(&ht->lock);
739 	memcpy(&ht->p, params, sizeof(*params));
740 
741 	if (params->min_size)
742 		ht->p.min_size = roundup_pow_of_two(params->min_size);
743 
744 	if (params->max_size)
745 		ht->p.max_size = rounddown_pow_of_two(params->max_size);
746 
747 	if (params->insecure_max_entries)
748 		ht->p.insecure_max_entries =
749 			rounddown_pow_of_two(params->insecure_max_entries);
750 	else
751 		ht->p.insecure_max_entries = ht->p.max_size * 2;
752 
753 	ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
754 
755 	/* The maximum (not average) chain length grows with the
756 	 * size of the hash table, at a rate of (log N)/(log log N).
757 	 * The value of 16 is selected so that even if the hash
758 	 * table grew to 2^32 you would not expect the maximum
759 	 * chain length to exceed it unless we are under attack
760 	 * (or extremely unlucky).
761 	 *
762 	 * As this limit is only to detect attacks, we don't need
763 	 * to set it to a lower value as you'd need the chain
764 	 * length to vastly exceed 16 to have any real effect
765 	 * on the system.
766 	 */
767 	if (!params->insecure_elasticity)
768 		ht->elasticity = 16;
769 
770 	if (params->locks_mul)
771 		ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
772 	else
773 		ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
774 
775 	ht->key_len = ht->p.key_len;
776 	if (!params->hashfn) {
777 		ht->p.hashfn = jhash;
778 
779 		if (!(ht->key_len & (sizeof(u32) - 1))) {
780 			ht->key_len /= sizeof(u32);
781 			ht->p.hashfn = rhashtable_jhash2;
782 		}
783 	}
784 
785 	tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
786 	if (tbl == NULL)
787 		return -ENOMEM;
788 
789 	atomic_set(&ht->nelems, 0);
790 
791 	RCU_INIT_POINTER(ht->tbl, tbl);
792 
793 	INIT_WORK(&ht->run_work, rht_deferred_worker);
794 
795 	return 0;
796 }
797 EXPORT_SYMBOL_GPL(rhashtable_init);
798 
799 /**
800  * rhashtable_free_and_destroy - free elements and destroy hash table
801  * @ht:		the hash table to destroy
802  * @free_fn:	callback to release resources of element
803  * @arg:	pointer passed to free_fn
804  *
805  * Stops an eventual async resize. If defined, invokes free_fn for each
806  * element to releasal resources. Please note that RCU protected
807  * readers may still be accessing the elements. Releasing of resources
808  * must occur in a compatible manner. Then frees the bucket array.
809  *
810  * This function will eventually sleep to wait for an async resize
811  * to complete. The caller is responsible that no further write operations
812  * occurs in parallel.
813  */
814 void rhashtable_free_and_destroy(struct rhashtable *ht,
815 				 void (*free_fn)(void *ptr, void *arg),
816 				 void *arg)
817 {
818 	const struct bucket_table *tbl;
819 	unsigned int i;
820 
821 	cancel_work_sync(&ht->run_work);
822 
823 	mutex_lock(&ht->mutex);
824 	tbl = rht_dereference(ht->tbl, ht);
825 	if (free_fn) {
826 		for (i = 0; i < tbl->size; i++) {
827 			struct rhash_head *pos, *next;
828 
829 			for (pos = rht_dereference(tbl->buckets[i], ht),
830 			     next = !rht_is_a_nulls(pos) ?
831 					rht_dereference(pos->next, ht) : NULL;
832 			     !rht_is_a_nulls(pos);
833 			     pos = next,
834 			     next = !rht_is_a_nulls(pos) ?
835 					rht_dereference(pos->next, ht) : NULL)
836 				free_fn(rht_obj(ht, pos), arg);
837 		}
838 	}
839 
840 	bucket_table_free(tbl);
841 	mutex_unlock(&ht->mutex);
842 }
843 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
844 
845 void rhashtable_destroy(struct rhashtable *ht)
846 {
847 	return rhashtable_free_and_destroy(ht, NULL, NULL);
848 }
849 EXPORT_SYMBOL_GPL(rhashtable_destroy);
850