xref: /linux/lib/rhashtable.c (revision 7a08cb9b4bb92fb86f5fe8a3aa0ac08a9b3d783b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Resizable, Scalable, Concurrent Hash Table
4  *
5  * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
6  * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
7  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
8  *
9  * Code partially derived from nft_hash
10  * Rewritten with rehash code from br_multicast plus single list
11  * pointer as suggested by Josh Triplett
12  */
13 
14 #include <linux/atomic.h>
15 #include <linux/kernel.h>
16 #include <linux/init.h>
17 #include <linux/log2.h>
18 #include <linux/sched.h>
19 #include <linux/rculist.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
22 #include <linux/mm.h>
23 #include <linux/jhash.h>
24 #include <linux/random.h>
25 #include <linux/rhashtable.h>
26 #include <linux/err.h>
27 #include <linux/export.h>
28 
29 #define HASH_DEFAULT_SIZE	64UL
30 #define HASH_MIN_SIZE		4U
31 
32 union nested_table {
33 	union nested_table __rcu *table;
34 	struct rhash_lock_head __rcu *bucket;
35 };
36 
37 static u32 head_hashfn(struct rhashtable *ht,
38 		       const struct bucket_table *tbl,
39 		       const struct rhash_head *he)
40 {
41 	return rht_head_hashfn(ht, tbl, he, ht->p);
42 }
43 
44 #ifdef CONFIG_PROVE_LOCKING
45 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
46 
47 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
48 {
49 	return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
50 }
51 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
52 
53 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
54 {
55 	if (!debug_locks)
56 		return 1;
57 	if (unlikely(tbl->nest))
58 		return 1;
59 	return bit_spin_is_locked(0, (unsigned long *)&tbl->buckets[hash]);
60 }
61 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
62 #else
63 #define ASSERT_RHT_MUTEX(HT)
64 #endif
65 
66 static inline union nested_table *nested_table_top(
67 	const struct bucket_table *tbl)
68 {
69 	/* The top-level bucket entry does not need RCU protection
70 	 * because it's set at the same time as tbl->nest.
71 	 */
72 	return (void *)rcu_dereference_protected(tbl->buckets[0], 1);
73 }
74 
75 static void nested_table_free(union nested_table *ntbl, unsigned int size)
76 {
77 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
78 	const unsigned int len = 1 << shift;
79 	unsigned int i;
80 
81 	ntbl = rcu_dereference_protected(ntbl->table, 1);
82 	if (!ntbl)
83 		return;
84 
85 	if (size > len) {
86 		size >>= shift;
87 		for (i = 0; i < len; i++)
88 			nested_table_free(ntbl + i, size);
89 	}
90 
91 	kfree(ntbl);
92 }
93 
94 static void nested_bucket_table_free(const struct bucket_table *tbl)
95 {
96 	unsigned int size = tbl->size >> tbl->nest;
97 	unsigned int len = 1 << tbl->nest;
98 	union nested_table *ntbl;
99 	unsigned int i;
100 
101 	ntbl = nested_table_top(tbl);
102 
103 	for (i = 0; i < len; i++)
104 		nested_table_free(ntbl + i, size);
105 
106 	kfree(ntbl);
107 }
108 
109 static void bucket_table_free(const struct bucket_table *tbl)
110 {
111 	if (tbl->nest)
112 		nested_bucket_table_free(tbl);
113 
114 	kvfree(tbl);
115 }
116 
117 static void bucket_table_free_rcu(struct rcu_head *head)
118 {
119 	bucket_table_free(container_of(head, struct bucket_table, rcu));
120 }
121 
122 static union nested_table *nested_table_alloc(struct rhashtable *ht,
123 					      union nested_table __rcu **prev,
124 					      bool leaf)
125 {
126 	union nested_table *ntbl;
127 	int i;
128 
129 	ntbl = rcu_dereference(*prev);
130 	if (ntbl)
131 		return ntbl;
132 
133 	ntbl = alloc_hooks_tag(ht->alloc_tag,
134 			kmalloc_noprof(PAGE_SIZE, GFP_ATOMIC|__GFP_ZERO));
135 
136 	if (ntbl && leaf) {
137 		for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
138 			INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
139 	}
140 
141 	if (cmpxchg((union nested_table **)prev, NULL, ntbl) == NULL)
142 		return ntbl;
143 	/* Raced with another thread. */
144 	kfree(ntbl);
145 	return rcu_dereference(*prev);
146 }
147 
148 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
149 						      size_t nbuckets,
150 						      gfp_t gfp)
151 {
152 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
153 	struct bucket_table *tbl;
154 	size_t size;
155 
156 	if (nbuckets < (1 << (shift + 1)))
157 		return NULL;
158 
159 	size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
160 
161 	tbl = alloc_hooks_tag(ht->alloc_tag,
162 			kmalloc_noprof(size, gfp|__GFP_ZERO));
163 	if (!tbl)
164 		return NULL;
165 
166 	if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
167 				false)) {
168 		kfree(tbl);
169 		return NULL;
170 	}
171 
172 	tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
173 
174 	return tbl;
175 }
176 
177 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
178 					       size_t nbuckets,
179 					       gfp_t gfp)
180 {
181 	struct bucket_table *tbl = NULL;
182 	size_t size;
183 	int i;
184 	static struct lock_class_key __key;
185 
186 	tbl = alloc_hooks_tag(ht->alloc_tag,
187 			kvmalloc_node_noprof(struct_size(tbl, buckets, nbuckets),
188 					     gfp|__GFP_ZERO, NUMA_NO_NODE));
189 
190 	size = nbuckets;
191 
192 	if (tbl == NULL && !gfpflags_allow_blocking(gfp)) {
193 		tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
194 		nbuckets = 0;
195 	}
196 
197 	if (tbl == NULL)
198 		return NULL;
199 
200 	lockdep_init_map(&tbl->dep_map, "rhashtable_bucket", &__key, 0);
201 
202 	tbl->size = size;
203 
204 	rcu_head_init(&tbl->rcu);
205 	INIT_LIST_HEAD(&tbl->walkers);
206 
207 	tbl->hash_rnd = get_random_u32();
208 
209 	for (i = 0; i < nbuckets; i++)
210 		INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
211 
212 	return tbl;
213 }
214 
215 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
216 						  struct bucket_table *tbl)
217 {
218 	struct bucket_table *new_tbl;
219 
220 	do {
221 		new_tbl = tbl;
222 		tbl = rht_dereference_rcu(tbl->future_tbl, ht);
223 	} while (tbl);
224 
225 	return new_tbl;
226 }
227 
228 static int rhashtable_rehash_one(struct rhashtable *ht,
229 				 struct rhash_lock_head __rcu **bkt,
230 				 unsigned int old_hash)
231 {
232 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
233 	struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
234 	int err = -EAGAIN;
235 	struct rhash_head *head, *next, *entry;
236 	struct rhash_head __rcu **pprev = NULL;
237 	unsigned int new_hash;
238 	unsigned long flags;
239 
240 	if (new_tbl->nest)
241 		goto out;
242 
243 	err = -ENOENT;
244 
245 	rht_for_each_from(entry, rht_ptr(bkt, old_tbl, old_hash),
246 			  old_tbl, old_hash) {
247 		err = 0;
248 		next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
249 
250 		if (rht_is_a_nulls(next))
251 			break;
252 
253 		pprev = &entry->next;
254 	}
255 
256 	if (err)
257 		goto out;
258 
259 	new_hash = head_hashfn(ht, new_tbl, entry);
260 
261 	flags = rht_lock_nested(new_tbl, &new_tbl->buckets[new_hash],
262 				SINGLE_DEPTH_NESTING);
263 
264 	head = rht_ptr(new_tbl->buckets + new_hash, new_tbl, new_hash);
265 
266 	RCU_INIT_POINTER(entry->next, head);
267 
268 	rht_assign_unlock(new_tbl, &new_tbl->buckets[new_hash], entry, flags);
269 
270 	if (pprev)
271 		rcu_assign_pointer(*pprev, next);
272 	else
273 		/* Need to preserved the bit lock. */
274 		rht_assign_locked(bkt, next);
275 
276 out:
277 	return err;
278 }
279 
280 static int rhashtable_rehash_chain(struct rhashtable *ht,
281 				    unsigned int old_hash)
282 {
283 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
284 	struct rhash_lock_head __rcu **bkt = rht_bucket_var(old_tbl, old_hash);
285 	unsigned long flags;
286 	int err;
287 
288 	if (!bkt)
289 		return 0;
290 	flags = rht_lock(old_tbl, bkt);
291 
292 	while (!(err = rhashtable_rehash_one(ht, bkt, old_hash)))
293 		;
294 
295 	if (err == -ENOENT)
296 		err = 0;
297 	rht_unlock(old_tbl, bkt, flags);
298 
299 	return err;
300 }
301 
302 static int rhashtable_rehash_attach(struct rhashtable *ht,
303 				    struct bucket_table *old_tbl,
304 				    struct bucket_table *new_tbl)
305 {
306 	/* Make insertions go into the new, empty table right away. Deletions
307 	 * and lookups will be attempted in both tables until we synchronize.
308 	 * As cmpxchg() provides strong barriers, we do not need
309 	 * rcu_assign_pointer().
310 	 */
311 
312 	if (cmpxchg((struct bucket_table **)&old_tbl->future_tbl, NULL,
313 		    new_tbl) != NULL)
314 		return -EEXIST;
315 
316 	return 0;
317 }
318 
319 static int rhashtable_rehash_table(struct rhashtable *ht)
320 {
321 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
322 	struct bucket_table *new_tbl;
323 	struct rhashtable_walker *walker;
324 	unsigned int old_hash;
325 	int err;
326 
327 	new_tbl = rht_dereference(old_tbl->future_tbl, ht);
328 	if (!new_tbl)
329 		return 0;
330 
331 	for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
332 		err = rhashtable_rehash_chain(ht, old_hash);
333 		if (err)
334 			return err;
335 		cond_resched();
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 
345 	/* Wait for readers. All new readers will see the new
346 	 * table, and thus no references to the old table will
347 	 * remain.
348 	 * We do this inside the locked region so that
349 	 * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
350 	 * to check if it should not re-link the table.
351 	 */
352 	call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
353 	spin_unlock(&ht->lock);
354 
355 	return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
356 }
357 
358 static int rhashtable_rehash_alloc(struct rhashtable *ht,
359 				   struct bucket_table *old_tbl,
360 				   unsigned int size)
361 {
362 	struct bucket_table *new_tbl;
363 	int err;
364 
365 	ASSERT_RHT_MUTEX(ht);
366 
367 	new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
368 	if (new_tbl == NULL)
369 		return -ENOMEM;
370 
371 	err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
372 	if (err)
373 		bucket_table_free(new_tbl);
374 
375 	return err;
376 }
377 
378 /**
379  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
380  * @ht:		the hash table to shrink
381  *
382  * This function shrinks the hash table to fit, i.e., the smallest
383  * size would not cause it to expand right away automatically.
384  *
385  * The caller must ensure that no concurrent resizing occurs by holding
386  * ht->mutex.
387  *
388  * The caller must ensure that no concurrent table mutations take place.
389  * It is however valid to have concurrent lookups if they are RCU protected.
390  *
391  * It is valid to have concurrent insertions and deletions protected by per
392  * bucket locks or concurrent RCU protected lookups and traversals.
393  */
394 static int rhashtable_shrink(struct rhashtable *ht)
395 {
396 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
397 	unsigned int nelems = atomic_read(&ht->nelems);
398 	unsigned int size = 0;
399 
400 	if (nelems)
401 		size = roundup_pow_of_two(nelems * 3 / 2);
402 	if (size < ht->p.min_size)
403 		size = ht->p.min_size;
404 
405 	if (old_tbl->size <= size)
406 		return 0;
407 
408 	if (rht_dereference(old_tbl->future_tbl, ht))
409 		return -EEXIST;
410 
411 	return rhashtable_rehash_alloc(ht, old_tbl, size);
412 }
413 
414 static void rht_deferred_worker(struct work_struct *work)
415 {
416 	struct rhashtable *ht;
417 	struct bucket_table *tbl;
418 	int err = 0;
419 
420 	ht = container_of(work, struct rhashtable, run_work);
421 	mutex_lock(&ht->mutex);
422 
423 	tbl = rht_dereference(ht->tbl, ht);
424 	tbl = rhashtable_last_table(ht, tbl);
425 
426 	if (rht_grow_above_75(ht, tbl))
427 		err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
428 	else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
429 		err = rhashtable_shrink(ht);
430 	else if (tbl->nest)
431 		err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
432 
433 	if (!err || err == -EEXIST) {
434 		int nerr;
435 
436 		nerr = rhashtable_rehash_table(ht);
437 		err = err ?: nerr;
438 	}
439 
440 	mutex_unlock(&ht->mutex);
441 
442 	if (err)
443 		schedule_work(&ht->run_work);
444 }
445 
446 static int rhashtable_insert_rehash(struct rhashtable *ht,
447 				    struct bucket_table *tbl)
448 {
449 	struct bucket_table *old_tbl;
450 	struct bucket_table *new_tbl;
451 	unsigned int size;
452 	int err;
453 
454 	old_tbl = rht_dereference_rcu(ht->tbl, ht);
455 
456 	size = tbl->size;
457 
458 	err = -EBUSY;
459 
460 	if (rht_grow_above_75(ht, tbl))
461 		size *= 2;
462 	/* Do not schedule more than one rehash */
463 	else if (old_tbl != tbl)
464 		goto fail;
465 
466 	err = -ENOMEM;
467 
468 	new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
469 	if (new_tbl == NULL)
470 		goto fail;
471 
472 	err = rhashtable_rehash_attach(ht, tbl, new_tbl);
473 	if (err) {
474 		bucket_table_free(new_tbl);
475 		if (err == -EEXIST)
476 			err = 0;
477 	} else
478 		schedule_work(&ht->run_work);
479 
480 	return err;
481 
482 fail:
483 	/* Do not fail the insert if someone else did a rehash. */
484 	if (likely(rcu_access_pointer(tbl->future_tbl)))
485 		return 0;
486 
487 	/* Schedule async rehash to retry allocation in process context. */
488 	if (err == -ENOMEM)
489 		schedule_work(&ht->run_work);
490 
491 	return err;
492 }
493 
494 static void *rhashtable_lookup_one(struct rhashtable *ht,
495 				   struct rhash_lock_head __rcu **bkt,
496 				   struct bucket_table *tbl, unsigned int hash,
497 				   const void *key, struct rhash_head *obj)
498 {
499 	struct rhashtable_compare_arg arg = {
500 		.ht = ht,
501 		.key = key,
502 	};
503 	struct rhash_head __rcu **pprev = NULL;
504 	struct rhash_head *head;
505 	int elasticity;
506 
507 	elasticity = RHT_ELASTICITY;
508 	rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
509 		struct rhlist_head *list;
510 		struct rhlist_head *plist;
511 
512 		elasticity--;
513 		if (!key ||
514 		    (ht->p.obj_cmpfn ?
515 		     ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
516 		     rhashtable_compare(&arg, rht_obj(ht, head)))) {
517 			pprev = &head->next;
518 			continue;
519 		}
520 
521 		if (!ht->rhlist)
522 			return rht_obj(ht, head);
523 
524 		list = container_of(obj, struct rhlist_head, rhead);
525 		plist = container_of(head, struct rhlist_head, rhead);
526 
527 		RCU_INIT_POINTER(list->next, plist);
528 		head = rht_dereference_bucket(head->next, tbl, hash);
529 		RCU_INIT_POINTER(list->rhead.next, head);
530 		if (pprev)
531 			rcu_assign_pointer(*pprev, obj);
532 		else
533 			/* Need to preserve the bit lock */
534 			rht_assign_locked(bkt, obj);
535 
536 		return NULL;
537 	}
538 
539 	if (elasticity <= 0)
540 		return ERR_PTR(-EAGAIN);
541 
542 	return ERR_PTR(-ENOENT);
543 }
544 
545 static struct bucket_table *rhashtable_insert_one(
546 	struct rhashtable *ht, struct rhash_lock_head __rcu **bkt,
547 	struct bucket_table *tbl, unsigned int hash, struct rhash_head *obj,
548 	void *data)
549 {
550 	struct bucket_table *new_tbl;
551 	struct rhash_head *head;
552 
553 	if (!IS_ERR_OR_NULL(data))
554 		return ERR_PTR(-EEXIST);
555 
556 	if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
557 		return ERR_CAST(data);
558 
559 	new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
560 	if (new_tbl)
561 		return new_tbl;
562 
563 	if (PTR_ERR(data) != -ENOENT)
564 		return ERR_CAST(data);
565 
566 	if (unlikely(rht_grow_above_max(ht, tbl)))
567 		return ERR_PTR(-E2BIG);
568 
569 	if (unlikely(rht_grow_above_100(ht, tbl)))
570 		return ERR_PTR(-EAGAIN);
571 
572 	head = rht_ptr(bkt, tbl, hash);
573 
574 	RCU_INIT_POINTER(obj->next, head);
575 	if (ht->rhlist) {
576 		struct rhlist_head *list;
577 
578 		list = container_of(obj, struct rhlist_head, rhead);
579 		RCU_INIT_POINTER(list->next, NULL);
580 	}
581 
582 	/* bkt is always the head of the list, so it holds
583 	 * the lock, which we need to preserve
584 	 */
585 	rht_assign_locked(bkt, obj);
586 
587 	atomic_inc(&ht->nelems);
588 	if (rht_grow_above_75(ht, tbl))
589 		schedule_work(&ht->run_work);
590 
591 	return NULL;
592 }
593 
594 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
595 				   struct rhash_head *obj)
596 {
597 	struct bucket_table *new_tbl;
598 	struct bucket_table *tbl;
599 	struct rhash_lock_head __rcu **bkt;
600 	unsigned long flags;
601 	unsigned int hash;
602 	void *data;
603 
604 	new_tbl = rcu_dereference(ht->tbl);
605 
606 	do {
607 		tbl = new_tbl;
608 		hash = rht_head_hashfn(ht, tbl, obj, ht->p);
609 		if (rcu_access_pointer(tbl->future_tbl))
610 			/* Failure is OK */
611 			bkt = rht_bucket_var(tbl, hash);
612 		else
613 			bkt = rht_bucket_insert(ht, tbl, hash);
614 		if (bkt == NULL) {
615 			new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
616 			data = ERR_PTR(-EAGAIN);
617 		} else {
618 			flags = rht_lock(tbl, bkt);
619 			data = rhashtable_lookup_one(ht, bkt, tbl,
620 						     hash, key, obj);
621 			new_tbl = rhashtable_insert_one(ht, bkt, tbl,
622 							hash, obj, data);
623 			if (PTR_ERR(new_tbl) != -EEXIST)
624 				data = ERR_CAST(new_tbl);
625 
626 			rht_unlock(tbl, bkt, flags);
627 		}
628 	} while (!IS_ERR_OR_NULL(new_tbl));
629 
630 	if (PTR_ERR(data) == -EAGAIN)
631 		data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
632 			       -EAGAIN);
633 
634 	return data;
635 }
636 
637 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
638 			     struct rhash_head *obj)
639 {
640 	void *data;
641 
642 	do {
643 		rcu_read_lock();
644 		data = rhashtable_try_insert(ht, key, obj);
645 		rcu_read_unlock();
646 	} while (PTR_ERR(data) == -EAGAIN);
647 
648 	return data;
649 }
650 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
651 
652 /**
653  * rhashtable_walk_enter - Initialise an iterator
654  * @ht:		Table to walk over
655  * @iter:	Hash table Iterator
656  *
657  * This function prepares a hash table walk.
658  *
659  * Note that if you restart a walk after rhashtable_walk_stop you
660  * may see the same object twice.  Also, you may miss objects if
661  * there are removals in between rhashtable_walk_stop and the next
662  * call to rhashtable_walk_start.
663  *
664  * For a completely stable walk you should construct your own data
665  * structure outside the hash table.
666  *
667  * This function may be called from any process context, including
668  * non-preemptable context, but cannot be called from softirq or
669  * hardirq context.
670  *
671  * You must call rhashtable_walk_exit after this function returns.
672  */
673 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
674 {
675 	iter->ht = ht;
676 	iter->p = NULL;
677 	iter->slot = 0;
678 	iter->skip = 0;
679 	iter->end_of_table = 0;
680 
681 	spin_lock(&ht->lock);
682 	iter->walker.tbl =
683 		rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
684 	list_add(&iter->walker.list, &iter->walker.tbl->walkers);
685 	spin_unlock(&ht->lock);
686 }
687 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
688 
689 /**
690  * rhashtable_walk_exit - Free an iterator
691  * @iter:	Hash table Iterator
692  *
693  * This function frees resources allocated by rhashtable_walk_enter.
694  */
695 void rhashtable_walk_exit(struct rhashtable_iter *iter)
696 {
697 	spin_lock(&iter->ht->lock);
698 	if (iter->walker.tbl)
699 		list_del(&iter->walker.list);
700 	spin_unlock(&iter->ht->lock);
701 }
702 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
703 
704 /**
705  * rhashtable_walk_start_check - Start a hash table walk
706  * @iter:	Hash table iterator
707  *
708  * Start a hash table walk at the current iterator position.  Note that we take
709  * the RCU lock in all cases including when we return an error.  So you must
710  * always call rhashtable_walk_stop to clean up.
711  *
712  * Returns zero if successful.
713  *
714  * Returns -EAGAIN if resize event occurred.  Note that the iterator
715  * will rewind back to the beginning and you may use it immediately
716  * by calling rhashtable_walk_next.
717  *
718  * rhashtable_walk_start is defined as an inline variant that returns
719  * void. This is preferred in cases where the caller would ignore
720  * resize events and always continue.
721  */
722 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
723 	__acquires(RCU)
724 {
725 	struct rhashtable *ht = iter->ht;
726 	bool rhlist = ht->rhlist;
727 
728 	rcu_read_lock();
729 
730 	spin_lock(&ht->lock);
731 	if (iter->walker.tbl)
732 		list_del(&iter->walker.list);
733 	spin_unlock(&ht->lock);
734 
735 	if (iter->end_of_table)
736 		return 0;
737 	if (!iter->walker.tbl) {
738 		iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
739 		iter->slot = 0;
740 		iter->skip = 0;
741 		return -EAGAIN;
742 	}
743 
744 	if (iter->p && !rhlist) {
745 		/*
746 		 * We need to validate that 'p' is still in the table, and
747 		 * if so, update 'skip'
748 		 */
749 		struct rhash_head *p;
750 		int skip = 0;
751 		rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
752 			skip++;
753 			if (p == iter->p) {
754 				iter->skip = skip;
755 				goto found;
756 			}
757 		}
758 		iter->p = NULL;
759 	} else if (iter->p && rhlist) {
760 		/* Need to validate that 'list' is still in the table, and
761 		 * if so, update 'skip' and 'p'.
762 		 */
763 		struct rhash_head *p;
764 		struct rhlist_head *list;
765 		int skip = 0;
766 		rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
767 			for (list = container_of(p, struct rhlist_head, rhead);
768 			     list;
769 			     list = rcu_dereference(list->next)) {
770 				skip++;
771 				if (list == iter->list) {
772 					iter->p = p;
773 					iter->skip = skip;
774 					goto found;
775 				}
776 			}
777 		}
778 		iter->p = NULL;
779 	}
780 found:
781 	return 0;
782 }
783 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
784 
785 /**
786  * __rhashtable_walk_find_next - Find the next element in a table (or the first
787  * one in case of a new walk).
788  *
789  * @iter:	Hash table iterator
790  *
791  * Returns the found object or NULL when the end of the table is reached.
792  *
793  * Returns -EAGAIN if resize event occurred.
794  */
795 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
796 {
797 	struct bucket_table *tbl = iter->walker.tbl;
798 	struct rhlist_head *list = iter->list;
799 	struct rhashtable *ht = iter->ht;
800 	struct rhash_head *p = iter->p;
801 	bool rhlist = ht->rhlist;
802 
803 	if (!tbl)
804 		return NULL;
805 
806 	for (; iter->slot < tbl->size; iter->slot++) {
807 		int skip = iter->skip;
808 
809 		rht_for_each_rcu(p, tbl, iter->slot) {
810 			if (rhlist) {
811 				list = container_of(p, struct rhlist_head,
812 						    rhead);
813 				do {
814 					if (!skip)
815 						goto next;
816 					skip--;
817 					list = rcu_dereference(list->next);
818 				} while (list);
819 
820 				continue;
821 			}
822 			if (!skip)
823 				break;
824 			skip--;
825 		}
826 
827 next:
828 		if (!rht_is_a_nulls(p)) {
829 			iter->skip++;
830 			iter->p = p;
831 			iter->list = list;
832 			return rht_obj(ht, rhlist ? &list->rhead : p);
833 		}
834 
835 		iter->skip = 0;
836 	}
837 
838 	iter->p = NULL;
839 
840 	/* Ensure we see any new tables. */
841 	smp_rmb();
842 
843 	iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
844 	if (iter->walker.tbl) {
845 		iter->slot = 0;
846 		iter->skip = 0;
847 		return ERR_PTR(-EAGAIN);
848 	} else {
849 		iter->end_of_table = true;
850 	}
851 
852 	return NULL;
853 }
854 
855 /**
856  * rhashtable_walk_next - Return the next object and advance the iterator
857  * @iter:	Hash table iterator
858  *
859  * Note that you must call rhashtable_walk_stop when you are finished
860  * with the walk.
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  */
867 void *rhashtable_walk_next(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 	bool rhlist = ht->rhlist;
873 
874 	if (p) {
875 		if (!rhlist || !(list = rcu_dereference(list->next))) {
876 			p = rcu_dereference(p->next);
877 			list = container_of(p, struct rhlist_head, rhead);
878 		}
879 		if (!rht_is_a_nulls(p)) {
880 			iter->skip++;
881 			iter->p = p;
882 			iter->list = list;
883 			return rht_obj(ht, rhlist ? &list->rhead : p);
884 		}
885 
886 		/* At the end of this slot, switch to next one and then find
887 		 * next entry from that point.
888 		 */
889 		iter->skip = 0;
890 		iter->slot++;
891 	}
892 
893 	return __rhashtable_walk_find_next(iter);
894 }
895 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
896 
897 /**
898  * rhashtable_walk_peek - Return the next object but don't advance the iterator
899  * @iter:	Hash table iterator
900  *
901  * Returns the next object or NULL when the end of the table is reached.
902  *
903  * Returns -EAGAIN if resize event occurred.  Note that the iterator
904  * will rewind back to the beginning and you may continue to use it.
905  */
906 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
907 {
908 	struct rhlist_head *list = iter->list;
909 	struct rhashtable *ht = iter->ht;
910 	struct rhash_head *p = iter->p;
911 
912 	if (p)
913 		return rht_obj(ht, ht->rhlist ? &list->rhead : p);
914 
915 	/* No object found in current iter, find next one in the table. */
916 
917 	if (iter->skip) {
918 		/* A nonzero skip value points to the next entry in the table
919 		 * beyond that last one that was found. Decrement skip so
920 		 * we find the current value. __rhashtable_walk_find_next
921 		 * will restore the original value of skip assuming that
922 		 * the table hasn't changed.
923 		 */
924 		iter->skip--;
925 	}
926 
927 	return __rhashtable_walk_find_next(iter);
928 }
929 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
930 
931 /**
932  * rhashtable_walk_stop - Finish a hash table walk
933  * @iter:	Hash table iterator
934  *
935  * Finish a hash table walk.  Does not reset the iterator to the start of the
936  * hash table.
937  */
938 void rhashtable_walk_stop(struct rhashtable_iter *iter)
939 	__releases(RCU)
940 {
941 	struct rhashtable *ht;
942 	struct bucket_table *tbl = iter->walker.tbl;
943 
944 	if (!tbl)
945 		goto out;
946 
947 	ht = iter->ht;
948 
949 	spin_lock(&ht->lock);
950 	if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
951 		/* This bucket table is being freed, don't re-link it. */
952 		iter->walker.tbl = NULL;
953 	else
954 		list_add(&iter->walker.list, &tbl->walkers);
955 	spin_unlock(&ht->lock);
956 
957 out:
958 	rcu_read_unlock();
959 }
960 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
961 
962 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
963 {
964 	size_t retsize;
965 
966 	if (params->nelem_hint)
967 		retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
968 			      (unsigned long)params->min_size);
969 	else
970 		retsize = max(HASH_DEFAULT_SIZE,
971 			      (unsigned long)params->min_size);
972 
973 	return retsize;
974 }
975 
976 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
977 {
978 	return jhash2(key, length, seed);
979 }
980 
981 /**
982  * rhashtable_init - initialize a new hash table
983  * @ht:		hash table to be initialized
984  * @params:	configuration parameters
985  *
986  * Initializes a new hash table based on the provided configuration
987  * parameters. A table can be configured either with a variable or
988  * fixed length key:
989  *
990  * Configuration Example 1: Fixed length keys
991  * struct test_obj {
992  *	int			key;
993  *	void *			my_member;
994  *	struct rhash_head	node;
995  * };
996  *
997  * struct rhashtable_params params = {
998  *	.head_offset = offsetof(struct test_obj, node),
999  *	.key_offset = offsetof(struct test_obj, key),
1000  *	.key_len = sizeof(int),
1001  *	.hashfn = jhash,
1002  * };
1003  *
1004  * Configuration Example 2: Variable length keys
1005  * struct test_obj {
1006  *	[...]
1007  *	struct rhash_head	node;
1008  * };
1009  *
1010  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1011  * {
1012  *	struct test_obj *obj = data;
1013  *
1014  *	return [... hash ...];
1015  * }
1016  *
1017  * struct rhashtable_params params = {
1018  *	.head_offset = offsetof(struct test_obj, node),
1019  *	.hashfn = jhash,
1020  *	.obj_hashfn = my_hash_fn,
1021  * };
1022  */
1023 int rhashtable_init_noprof(struct rhashtable *ht,
1024 		    const struct rhashtable_params *params)
1025 {
1026 	struct bucket_table *tbl;
1027 	size_t size;
1028 
1029 	if ((!params->key_len && !params->obj_hashfn) ||
1030 	    (params->obj_hashfn && !params->obj_cmpfn))
1031 		return -EINVAL;
1032 
1033 	memset(ht, 0, sizeof(*ht));
1034 	mutex_init(&ht->mutex);
1035 	spin_lock_init(&ht->lock);
1036 	memcpy(&ht->p, params, sizeof(*params));
1037 
1038 	alloc_tag_record(ht->alloc_tag);
1039 
1040 	if (params->min_size)
1041 		ht->p.min_size = roundup_pow_of_two(params->min_size);
1042 
1043 	/* Cap total entries at 2^31 to avoid nelems overflow. */
1044 	ht->max_elems = 1u << 31;
1045 
1046 	if (params->max_size) {
1047 		ht->p.max_size = rounddown_pow_of_two(params->max_size);
1048 		if (ht->p.max_size < ht->max_elems / 2)
1049 			ht->max_elems = ht->p.max_size * 2;
1050 	}
1051 
1052 	ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1053 
1054 	size = rounded_hashtable_size(&ht->p);
1055 
1056 	ht->key_len = ht->p.key_len;
1057 	if (!params->hashfn) {
1058 		ht->p.hashfn = jhash;
1059 
1060 		if (!(ht->key_len & (sizeof(u32) - 1))) {
1061 			ht->key_len /= sizeof(u32);
1062 			ht->p.hashfn = rhashtable_jhash2;
1063 		}
1064 	}
1065 
1066 	/*
1067 	 * This is api initialization and thus we need to guarantee the
1068 	 * initial rhashtable allocation. Upon failure, retry with the
1069 	 * smallest possible size with __GFP_NOFAIL semantics.
1070 	 */
1071 	tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1072 	if (unlikely(tbl == NULL)) {
1073 		size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1074 		tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1075 	}
1076 
1077 	atomic_set(&ht->nelems, 0);
1078 
1079 	RCU_INIT_POINTER(ht->tbl, tbl);
1080 
1081 	INIT_WORK(&ht->run_work, rht_deferred_worker);
1082 
1083 	return 0;
1084 }
1085 EXPORT_SYMBOL_GPL(rhashtable_init_noprof);
1086 
1087 /**
1088  * rhltable_init - initialize a new hash list table
1089  * @hlt:	hash list table to be initialized
1090  * @params:	configuration parameters
1091  *
1092  * Initializes a new hash list table.
1093  *
1094  * See documentation for rhashtable_init.
1095  */
1096 int rhltable_init_noprof(struct rhltable *hlt, const struct rhashtable_params *params)
1097 {
1098 	int err;
1099 
1100 	err = rhashtable_init_noprof(&hlt->ht, params);
1101 	hlt->ht.rhlist = true;
1102 	return err;
1103 }
1104 EXPORT_SYMBOL_GPL(rhltable_init_noprof);
1105 
1106 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1107 				void (*free_fn)(void *ptr, void *arg),
1108 				void *arg)
1109 {
1110 	struct rhlist_head *list;
1111 
1112 	if (!ht->rhlist) {
1113 		free_fn(rht_obj(ht, obj), arg);
1114 		return;
1115 	}
1116 
1117 	list = container_of(obj, struct rhlist_head, rhead);
1118 	do {
1119 		obj = &list->rhead;
1120 		list = rht_dereference(list->next, ht);
1121 		free_fn(rht_obj(ht, obj), arg);
1122 	} while (list);
1123 }
1124 
1125 /**
1126  * rhashtable_free_and_destroy - free elements and destroy hash table
1127  * @ht:		the hash table to destroy
1128  * @free_fn:	callback to release resources of element
1129  * @arg:	pointer passed to free_fn
1130  *
1131  * Stops an eventual async resize. If defined, invokes free_fn for each
1132  * element to releasal resources. Please note that RCU protected
1133  * readers may still be accessing the elements. Releasing of resources
1134  * must occur in a compatible manner. Then frees the bucket array.
1135  *
1136  * This function will eventually sleep to wait for an async resize
1137  * to complete. The caller is responsible that no further write operations
1138  * occurs in parallel.
1139  */
1140 void rhashtable_free_and_destroy(struct rhashtable *ht,
1141 				 void (*free_fn)(void *ptr, void *arg),
1142 				 void *arg)
1143 {
1144 	struct bucket_table *tbl, *next_tbl;
1145 	unsigned int i;
1146 
1147 	cancel_work_sync(&ht->run_work);
1148 
1149 	mutex_lock(&ht->mutex);
1150 	tbl = rht_dereference(ht->tbl, ht);
1151 restart:
1152 	if (free_fn) {
1153 		for (i = 0; i < tbl->size; i++) {
1154 			struct rhash_head *pos, *next;
1155 
1156 			cond_resched();
1157 			for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1158 			     next = !rht_is_a_nulls(pos) ?
1159 					rht_dereference(pos->next, ht) : NULL;
1160 			     !rht_is_a_nulls(pos);
1161 			     pos = next,
1162 			     next = !rht_is_a_nulls(pos) ?
1163 					rht_dereference(pos->next, ht) : NULL)
1164 				rhashtable_free_one(ht, pos, free_fn, arg);
1165 		}
1166 	}
1167 
1168 	next_tbl = rht_dereference(tbl->future_tbl, ht);
1169 	bucket_table_free(tbl);
1170 	if (next_tbl) {
1171 		tbl = next_tbl;
1172 		goto restart;
1173 	}
1174 	mutex_unlock(&ht->mutex);
1175 }
1176 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1177 
1178 void rhashtable_destroy(struct rhashtable *ht)
1179 {
1180 	return rhashtable_free_and_destroy(ht, NULL, NULL);
1181 }
1182 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1183 
1184 struct rhash_lock_head __rcu **__rht_bucket_nested(
1185 	const struct bucket_table *tbl, unsigned int hash)
1186 {
1187 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1188 	unsigned int index = hash & ((1 << tbl->nest) - 1);
1189 	unsigned int size = tbl->size >> tbl->nest;
1190 	unsigned int subhash = hash;
1191 	union nested_table *ntbl;
1192 
1193 	ntbl = nested_table_top(tbl);
1194 	ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1195 	subhash >>= tbl->nest;
1196 
1197 	while (ntbl && size > (1 << shift)) {
1198 		index = subhash & ((1 << shift) - 1);
1199 		ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1200 						  tbl, hash);
1201 		size >>= shift;
1202 		subhash >>= shift;
1203 	}
1204 
1205 	if (!ntbl)
1206 		return NULL;
1207 
1208 	return &ntbl[subhash].bucket;
1209 
1210 }
1211 EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1212 
1213 struct rhash_lock_head __rcu **rht_bucket_nested(
1214 	const struct bucket_table *tbl, unsigned int hash)
1215 {
1216 	static struct rhash_lock_head __rcu *rhnull;
1217 
1218 	if (!rhnull)
1219 		INIT_RHT_NULLS_HEAD(rhnull);
1220 	return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1221 }
1222 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1223 
1224 struct rhash_lock_head __rcu **rht_bucket_nested_insert(
1225 	struct rhashtable *ht, struct bucket_table *tbl, unsigned int hash)
1226 {
1227 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1228 	unsigned int index = hash & ((1 << tbl->nest) - 1);
1229 	unsigned int size = tbl->size >> tbl->nest;
1230 	union nested_table *ntbl;
1231 
1232 	ntbl = nested_table_top(tbl);
1233 	hash >>= tbl->nest;
1234 	ntbl = nested_table_alloc(ht, &ntbl[index].table,
1235 				  size <= (1 << shift));
1236 
1237 	while (ntbl && size > (1 << shift)) {
1238 		index = hash & ((1 << shift) - 1);
1239 		size >>= shift;
1240 		hash >>= shift;
1241 		ntbl = nested_table_alloc(ht, &ntbl[index].table,
1242 					  size <= (1 << shift));
1243 	}
1244 
1245 	if (!ntbl)
1246 		return NULL;
1247 
1248 	return &ntbl[hash].bucket;
1249 
1250 }
1251 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);
1252