xref: /openbmc/linux/lib/rhashtable.c (revision 9958d30f)
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 = kzalloc(PAGE_SIZE, GFP_ATOMIC);
134 
135 	if (ntbl && leaf) {
136 		for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
137 			INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
138 	}
139 
140 	if (cmpxchg((union nested_table **)prev, NULL, ntbl) == NULL)
141 		return ntbl;
142 	/* Raced with another thread. */
143 	kfree(ntbl);
144 	return rcu_dereference(*prev);
145 }
146 
147 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
148 						      size_t nbuckets,
149 						      gfp_t gfp)
150 {
151 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
152 	struct bucket_table *tbl;
153 	size_t size;
154 
155 	if (nbuckets < (1 << (shift + 1)))
156 		return NULL;
157 
158 	size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
159 
160 	tbl = kzalloc(size, gfp);
161 	if (!tbl)
162 		return NULL;
163 
164 	if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
165 				false)) {
166 		kfree(tbl);
167 		return NULL;
168 	}
169 
170 	tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
171 
172 	return tbl;
173 }
174 
175 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
176 					       size_t nbuckets,
177 					       gfp_t gfp)
178 {
179 	struct bucket_table *tbl = NULL;
180 	size_t size;
181 	int i;
182 	static struct lock_class_key __key;
183 
184 	tbl = kvzalloc(struct_size(tbl, buckets, nbuckets), gfp);
185 
186 	size = nbuckets;
187 
188 	if (tbl == NULL && (gfp & ~__GFP_NOFAIL) != GFP_KERNEL) {
189 		tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
190 		nbuckets = 0;
191 	}
192 
193 	if (tbl == NULL)
194 		return NULL;
195 
196 	lockdep_init_map(&tbl->dep_map, "rhashtable_bucket", &__key, 0);
197 
198 	tbl->size = size;
199 
200 	rcu_head_init(&tbl->rcu);
201 	INIT_LIST_HEAD(&tbl->walkers);
202 
203 	tbl->hash_rnd = get_random_u32();
204 
205 	for (i = 0; i < nbuckets; i++)
206 		INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
207 
208 	return tbl;
209 }
210 
211 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
212 						  struct bucket_table *tbl)
213 {
214 	struct bucket_table *new_tbl;
215 
216 	do {
217 		new_tbl = tbl;
218 		tbl = rht_dereference_rcu(tbl->future_tbl, ht);
219 	} while (tbl);
220 
221 	return new_tbl;
222 }
223 
224 static int rhashtable_rehash_one(struct rhashtable *ht,
225 				 struct rhash_lock_head __rcu **bkt,
226 				 unsigned int old_hash)
227 {
228 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
229 	struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
230 	int err = -EAGAIN;
231 	struct rhash_head *head, *next, *entry;
232 	struct rhash_head __rcu **pprev = NULL;
233 	unsigned int new_hash;
234 
235 	if (new_tbl->nest)
236 		goto out;
237 
238 	err = -ENOENT;
239 
240 	rht_for_each_from(entry, rht_ptr(bkt, old_tbl, old_hash),
241 			  old_tbl, old_hash) {
242 		err = 0;
243 		next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
244 
245 		if (rht_is_a_nulls(next))
246 			break;
247 
248 		pprev = &entry->next;
249 	}
250 
251 	if (err)
252 		goto out;
253 
254 	new_hash = head_hashfn(ht, new_tbl, entry);
255 
256 	rht_lock_nested(new_tbl, &new_tbl->buckets[new_hash], SINGLE_DEPTH_NESTING);
257 
258 	head = rht_ptr(new_tbl->buckets + new_hash, new_tbl, new_hash);
259 
260 	RCU_INIT_POINTER(entry->next, head);
261 
262 	rht_assign_unlock(new_tbl, &new_tbl->buckets[new_hash], entry);
263 
264 	if (pprev)
265 		rcu_assign_pointer(*pprev, next);
266 	else
267 		/* Need to preserved the bit lock. */
268 		rht_assign_locked(bkt, next);
269 
270 out:
271 	return err;
272 }
273 
274 static int rhashtable_rehash_chain(struct rhashtable *ht,
275 				    unsigned int old_hash)
276 {
277 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
278 	struct rhash_lock_head __rcu **bkt = rht_bucket_var(old_tbl, old_hash);
279 	int err;
280 
281 	if (!bkt)
282 		return 0;
283 	rht_lock(old_tbl, bkt);
284 
285 	while (!(err = rhashtable_rehash_one(ht, bkt, old_hash)))
286 		;
287 
288 	if (err == -ENOENT)
289 		err = 0;
290 	rht_unlock(old_tbl, bkt);
291 
292 	return err;
293 }
294 
295 static int rhashtable_rehash_attach(struct rhashtable *ht,
296 				    struct bucket_table *old_tbl,
297 				    struct bucket_table *new_tbl)
298 {
299 	/* Make insertions go into the new, empty table right away. Deletions
300 	 * and lookups will be attempted in both tables until we synchronize.
301 	 * As cmpxchg() provides strong barriers, we do not need
302 	 * rcu_assign_pointer().
303 	 */
304 
305 	if (cmpxchg((struct bucket_table **)&old_tbl->future_tbl, NULL,
306 		    new_tbl) != NULL)
307 		return -EEXIST;
308 
309 	return 0;
310 }
311 
312 static int rhashtable_rehash_table(struct rhashtable *ht)
313 {
314 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
315 	struct bucket_table *new_tbl;
316 	struct rhashtable_walker *walker;
317 	unsigned int old_hash;
318 	int err;
319 
320 	new_tbl = rht_dereference(old_tbl->future_tbl, ht);
321 	if (!new_tbl)
322 		return 0;
323 
324 	for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
325 		err = rhashtable_rehash_chain(ht, old_hash);
326 		if (err)
327 			return err;
328 		cond_resched();
329 	}
330 
331 	/* Publish the new table pointer. */
332 	rcu_assign_pointer(ht->tbl, new_tbl);
333 
334 	spin_lock(&ht->lock);
335 	list_for_each_entry(walker, &old_tbl->walkers, list)
336 		walker->tbl = NULL;
337 
338 	/* Wait for readers. All new readers will see the new
339 	 * table, and thus no references to the old table will
340 	 * remain.
341 	 * We do this inside the locked region so that
342 	 * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
343 	 * to check if it should not re-link the table.
344 	 */
345 	call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
346 	spin_unlock(&ht->lock);
347 
348 	return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
349 }
350 
351 static int rhashtable_rehash_alloc(struct rhashtable *ht,
352 				   struct bucket_table *old_tbl,
353 				   unsigned int size)
354 {
355 	struct bucket_table *new_tbl;
356 	int err;
357 
358 	ASSERT_RHT_MUTEX(ht);
359 
360 	new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
361 	if (new_tbl == NULL)
362 		return -ENOMEM;
363 
364 	err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
365 	if (err)
366 		bucket_table_free(new_tbl);
367 
368 	return err;
369 }
370 
371 /**
372  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
373  * @ht:		the hash table to shrink
374  *
375  * This function shrinks the hash table to fit, i.e., the smallest
376  * size would not cause it to expand right away automatically.
377  *
378  * The caller must ensure that no concurrent resizing occurs by holding
379  * ht->mutex.
380  *
381  * The caller must ensure that no concurrent table mutations take place.
382  * It is however valid to have concurrent lookups if they are RCU protected.
383  *
384  * It is valid to have concurrent insertions and deletions protected by per
385  * bucket locks or concurrent RCU protected lookups and traversals.
386  */
387 static int rhashtable_shrink(struct rhashtable *ht)
388 {
389 	struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
390 	unsigned int nelems = atomic_read(&ht->nelems);
391 	unsigned int size = 0;
392 
393 	if (nelems)
394 		size = roundup_pow_of_two(nelems * 3 / 2);
395 	if (size < ht->p.min_size)
396 		size = ht->p.min_size;
397 
398 	if (old_tbl->size <= size)
399 		return 0;
400 
401 	if (rht_dereference(old_tbl->future_tbl, ht))
402 		return -EEXIST;
403 
404 	return rhashtable_rehash_alloc(ht, old_tbl, size);
405 }
406 
407 static void rht_deferred_worker(struct work_struct *work)
408 {
409 	struct rhashtable *ht;
410 	struct bucket_table *tbl;
411 	int err = 0;
412 
413 	ht = container_of(work, struct rhashtable, run_work);
414 	mutex_lock(&ht->mutex);
415 
416 	tbl = rht_dereference(ht->tbl, ht);
417 	tbl = rhashtable_last_table(ht, tbl);
418 
419 	if (rht_grow_above_75(ht, tbl))
420 		err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
421 	else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
422 		err = rhashtable_shrink(ht);
423 	else if (tbl->nest)
424 		err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
425 
426 	if (!err || err == -EEXIST) {
427 		int nerr;
428 
429 		nerr = rhashtable_rehash_table(ht);
430 		err = err ?: nerr;
431 	}
432 
433 	mutex_unlock(&ht->mutex);
434 
435 	if (err)
436 		schedule_work(&ht->run_work);
437 }
438 
439 static int rhashtable_insert_rehash(struct rhashtable *ht,
440 				    struct bucket_table *tbl)
441 {
442 	struct bucket_table *old_tbl;
443 	struct bucket_table *new_tbl;
444 	unsigned int size;
445 	int err;
446 
447 	old_tbl = rht_dereference_rcu(ht->tbl, ht);
448 
449 	size = tbl->size;
450 
451 	err = -EBUSY;
452 
453 	if (rht_grow_above_75(ht, tbl))
454 		size *= 2;
455 	/* Do not schedule more than one rehash */
456 	else if (old_tbl != tbl)
457 		goto fail;
458 
459 	err = -ENOMEM;
460 
461 	new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
462 	if (new_tbl == NULL)
463 		goto fail;
464 
465 	err = rhashtable_rehash_attach(ht, tbl, new_tbl);
466 	if (err) {
467 		bucket_table_free(new_tbl);
468 		if (err == -EEXIST)
469 			err = 0;
470 	} else
471 		schedule_work(&ht->run_work);
472 
473 	return err;
474 
475 fail:
476 	/* Do not fail the insert if someone else did a rehash. */
477 	if (likely(rcu_access_pointer(tbl->future_tbl)))
478 		return 0;
479 
480 	/* Schedule async rehash to retry allocation in process context. */
481 	if (err == -ENOMEM)
482 		schedule_work(&ht->run_work);
483 
484 	return err;
485 }
486 
487 static void *rhashtable_lookup_one(struct rhashtable *ht,
488 				   struct rhash_lock_head __rcu **bkt,
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 = NULL;
497 	struct rhash_head *head;
498 	int elasticity;
499 
500 	elasticity = RHT_ELASTICITY;
501 	rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
502 		struct rhlist_head *list;
503 		struct rhlist_head *plist;
504 
505 		elasticity--;
506 		if (!key ||
507 		    (ht->p.obj_cmpfn ?
508 		     ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
509 		     rhashtable_compare(&arg, rht_obj(ht, head)))) {
510 			pprev = &head->next;
511 			continue;
512 		}
513 
514 		if (!ht->rhlist)
515 			return rht_obj(ht, head);
516 
517 		list = container_of(obj, struct rhlist_head, rhead);
518 		plist = container_of(head, struct rhlist_head, rhead);
519 
520 		RCU_INIT_POINTER(list->next, plist);
521 		head = rht_dereference_bucket(head->next, tbl, hash);
522 		RCU_INIT_POINTER(list->rhead.next, head);
523 		if (pprev)
524 			rcu_assign_pointer(*pprev, obj);
525 		else
526 			/* Need to preserve the bit lock */
527 			rht_assign_locked(bkt, obj);
528 
529 		return NULL;
530 	}
531 
532 	if (elasticity <= 0)
533 		return ERR_PTR(-EAGAIN);
534 
535 	return ERR_PTR(-ENOENT);
536 }
537 
538 static struct bucket_table *rhashtable_insert_one(
539 	struct rhashtable *ht, struct rhash_lock_head __rcu **bkt,
540 	struct bucket_table *tbl, unsigned int hash, struct rhash_head *obj,
541 	void *data)
542 {
543 	struct bucket_table *new_tbl;
544 	struct rhash_head *head;
545 
546 	if (!IS_ERR_OR_NULL(data))
547 		return ERR_PTR(-EEXIST);
548 
549 	if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
550 		return ERR_CAST(data);
551 
552 	new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
553 	if (new_tbl)
554 		return new_tbl;
555 
556 	if (PTR_ERR(data) != -ENOENT)
557 		return ERR_CAST(data);
558 
559 	if (unlikely(rht_grow_above_max(ht, tbl)))
560 		return ERR_PTR(-E2BIG);
561 
562 	if (unlikely(rht_grow_above_100(ht, tbl)))
563 		return ERR_PTR(-EAGAIN);
564 
565 	head = rht_ptr(bkt, tbl, hash);
566 
567 	RCU_INIT_POINTER(obj->next, head);
568 	if (ht->rhlist) {
569 		struct rhlist_head *list;
570 
571 		list = container_of(obj, struct rhlist_head, rhead);
572 		RCU_INIT_POINTER(list->next, NULL);
573 	}
574 
575 	/* bkt is always the head of the list, so it holds
576 	 * the lock, which we need to preserve
577 	 */
578 	rht_assign_locked(bkt, 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 
587 static 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 	struct rhash_lock_head __rcu **bkt;
593 	unsigned int hash;
594 	void *data;
595 
596 	new_tbl = rcu_dereference(ht->tbl);
597 
598 	do {
599 		tbl = new_tbl;
600 		hash = rht_head_hashfn(ht, tbl, obj, ht->p);
601 		if (rcu_access_pointer(tbl->future_tbl))
602 			/* Failure is OK */
603 			bkt = rht_bucket_var(tbl, hash);
604 		else
605 			bkt = rht_bucket_insert(ht, tbl, hash);
606 		if (bkt == NULL) {
607 			new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
608 			data = ERR_PTR(-EAGAIN);
609 		} else {
610 			rht_lock(tbl, bkt);
611 			data = rhashtable_lookup_one(ht, bkt, tbl,
612 						     hash, key, obj);
613 			new_tbl = rhashtable_insert_one(ht, bkt, tbl,
614 							hash, obj, data);
615 			if (PTR_ERR(new_tbl) != -EEXIST)
616 				data = ERR_CAST(new_tbl);
617 
618 			rht_unlock(tbl, bkt);
619 		}
620 	} while (!IS_ERR_OR_NULL(new_tbl));
621 
622 	if (PTR_ERR(data) == -EAGAIN)
623 		data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
624 			       -EAGAIN);
625 
626 	return data;
627 }
628 
629 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
630 			     struct rhash_head *obj)
631 {
632 	void *data;
633 
634 	do {
635 		rcu_read_lock();
636 		data = rhashtable_try_insert(ht, key, obj);
637 		rcu_read_unlock();
638 	} while (PTR_ERR(data) == -EAGAIN);
639 
640 	return data;
641 }
642 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
643 
644 /**
645  * rhashtable_walk_enter - Initialise an iterator
646  * @ht:		Table to walk over
647  * @iter:	Hash table Iterator
648  *
649  * This function prepares a hash table walk.
650  *
651  * Note that if you restart a walk after rhashtable_walk_stop you
652  * may see the same object twice.  Also, you may miss objects if
653  * there are removals in between rhashtable_walk_stop and the next
654  * call to rhashtable_walk_start.
655  *
656  * For a completely stable walk you should construct your own data
657  * structure outside the hash table.
658  *
659  * This function may be called from any process context, including
660  * non-preemptable context, but cannot be called from softirq or
661  * hardirq context.
662  *
663  * You must call rhashtable_walk_exit after this function returns.
664  */
665 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
666 {
667 	iter->ht = ht;
668 	iter->p = NULL;
669 	iter->slot = 0;
670 	iter->skip = 0;
671 	iter->end_of_table = 0;
672 
673 	spin_lock(&ht->lock);
674 	iter->walker.tbl =
675 		rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
676 	list_add(&iter->walker.list, &iter->walker.tbl->walkers);
677 	spin_unlock(&ht->lock);
678 }
679 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
680 
681 /**
682  * rhashtable_walk_exit - Free an iterator
683  * @iter:	Hash table Iterator
684  *
685  * This function frees resources allocated by rhashtable_walk_enter.
686  */
687 void rhashtable_walk_exit(struct rhashtable_iter *iter)
688 {
689 	spin_lock(&iter->ht->lock);
690 	if (iter->walker.tbl)
691 		list_del(&iter->walker.list);
692 	spin_unlock(&iter->ht->lock);
693 }
694 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
695 
696 /**
697  * rhashtable_walk_start_check - Start a hash table walk
698  * @iter:	Hash table iterator
699  *
700  * Start a hash table walk at the current iterator position.  Note that we take
701  * the RCU lock in all cases including when we return an error.  So you must
702  * always call rhashtable_walk_stop to clean up.
703  *
704  * Returns zero if successful.
705  *
706  * Returns -EAGAIN if resize event occurred.  Note that the iterator
707  * will rewind back to the beginning and you may use it immediately
708  * by calling rhashtable_walk_next.
709  *
710  * rhashtable_walk_start is defined as an inline variant that returns
711  * void. This is preferred in cases where the caller would ignore
712  * resize events and always continue.
713  */
714 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
715 	__acquires(RCU)
716 {
717 	struct rhashtable *ht = iter->ht;
718 	bool rhlist = ht->rhlist;
719 
720 	rcu_read_lock();
721 
722 	spin_lock(&ht->lock);
723 	if (iter->walker.tbl)
724 		list_del(&iter->walker.list);
725 	spin_unlock(&ht->lock);
726 
727 	if (iter->end_of_table)
728 		return 0;
729 	if (!iter->walker.tbl) {
730 		iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
731 		iter->slot = 0;
732 		iter->skip = 0;
733 		return -EAGAIN;
734 	}
735 
736 	if (iter->p && !rhlist) {
737 		/*
738 		 * We need to validate that 'p' is still in the table, and
739 		 * if so, update 'skip'
740 		 */
741 		struct rhash_head *p;
742 		int skip = 0;
743 		rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
744 			skip++;
745 			if (p == iter->p) {
746 				iter->skip = skip;
747 				goto found;
748 			}
749 		}
750 		iter->p = NULL;
751 	} else if (iter->p && rhlist) {
752 		/* Need to validate that 'list' is still in the table, and
753 		 * if so, update 'skip' and 'p'.
754 		 */
755 		struct rhash_head *p;
756 		struct rhlist_head *list;
757 		int skip = 0;
758 		rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
759 			for (list = container_of(p, struct rhlist_head, rhead);
760 			     list;
761 			     list = rcu_dereference(list->next)) {
762 				skip++;
763 				if (list == iter->list) {
764 					iter->p = p;
765 					iter->skip = skip;
766 					goto found;
767 				}
768 			}
769 		}
770 		iter->p = NULL;
771 	}
772 found:
773 	return 0;
774 }
775 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
776 
777 /**
778  * __rhashtable_walk_find_next - Find the next element in a table (or the first
779  * one in case of a new walk).
780  *
781  * @iter:	Hash table iterator
782  *
783  * Returns the found object or NULL when the end of the table is reached.
784  *
785  * Returns -EAGAIN if resize event occurred.
786  */
787 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
788 {
789 	struct bucket_table *tbl = iter->walker.tbl;
790 	struct rhlist_head *list = iter->list;
791 	struct rhashtable *ht = iter->ht;
792 	struct rhash_head *p = iter->p;
793 	bool rhlist = ht->rhlist;
794 
795 	if (!tbl)
796 		return NULL;
797 
798 	for (; iter->slot < tbl->size; iter->slot++) {
799 		int skip = iter->skip;
800 
801 		rht_for_each_rcu(p, tbl, iter->slot) {
802 			if (rhlist) {
803 				list = container_of(p, struct rhlist_head,
804 						    rhead);
805 				do {
806 					if (!skip)
807 						goto next;
808 					skip--;
809 					list = rcu_dereference(list->next);
810 				} while (list);
811 
812 				continue;
813 			}
814 			if (!skip)
815 				break;
816 			skip--;
817 		}
818 
819 next:
820 		if (!rht_is_a_nulls(p)) {
821 			iter->skip++;
822 			iter->p = p;
823 			iter->list = list;
824 			return rht_obj(ht, rhlist ? &list->rhead : p);
825 		}
826 
827 		iter->skip = 0;
828 	}
829 
830 	iter->p = NULL;
831 
832 	/* Ensure we see any new tables. */
833 	smp_rmb();
834 
835 	iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
836 	if (iter->walker.tbl) {
837 		iter->slot = 0;
838 		iter->skip = 0;
839 		return ERR_PTR(-EAGAIN);
840 	} else {
841 		iter->end_of_table = true;
842 	}
843 
844 	return NULL;
845 }
846 
847 /**
848  * rhashtable_walk_next - Return the next object and advance the iterator
849  * @iter:	Hash table iterator
850  *
851  * Note that you must call rhashtable_walk_stop when you are finished
852  * with the walk.
853  *
854  * Returns the next object or NULL when the end of the table is reached.
855  *
856  * Returns -EAGAIN if resize event occurred.  Note that the iterator
857  * will rewind back to the beginning and you may continue to use it.
858  */
859 void *rhashtable_walk_next(struct rhashtable_iter *iter)
860 {
861 	struct rhlist_head *list = iter->list;
862 	struct rhashtable *ht = iter->ht;
863 	struct rhash_head *p = iter->p;
864 	bool rhlist = ht->rhlist;
865 
866 	if (p) {
867 		if (!rhlist || !(list = rcu_dereference(list->next))) {
868 			p = rcu_dereference(p->next);
869 			list = container_of(p, struct rhlist_head, rhead);
870 		}
871 		if (!rht_is_a_nulls(p)) {
872 			iter->skip++;
873 			iter->p = p;
874 			iter->list = list;
875 			return rht_obj(ht, rhlist ? &list->rhead : p);
876 		}
877 
878 		/* At the end of this slot, switch to next one and then find
879 		 * next entry from that point.
880 		 */
881 		iter->skip = 0;
882 		iter->slot++;
883 	}
884 
885 	return __rhashtable_walk_find_next(iter);
886 }
887 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
888 
889 /**
890  * rhashtable_walk_peek - Return the next object but don't advance the iterator
891  * @iter:	Hash table iterator
892  *
893  * Returns the next object or NULL when the end of the table is reached.
894  *
895  * Returns -EAGAIN if resize event occurred.  Note that the iterator
896  * will rewind back to the beginning and you may continue to use it.
897  */
898 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
899 {
900 	struct rhlist_head *list = iter->list;
901 	struct rhashtable *ht = iter->ht;
902 	struct rhash_head *p = iter->p;
903 
904 	if (p)
905 		return rht_obj(ht, ht->rhlist ? &list->rhead : p);
906 
907 	/* No object found in current iter, find next one in the table. */
908 
909 	if (iter->skip) {
910 		/* A nonzero skip value points to the next entry in the table
911 		 * beyond that last one that was found. Decrement skip so
912 		 * we find the current value. __rhashtable_walk_find_next
913 		 * will restore the original value of skip assuming that
914 		 * the table hasn't changed.
915 		 */
916 		iter->skip--;
917 	}
918 
919 	return __rhashtable_walk_find_next(iter);
920 }
921 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
922 
923 /**
924  * rhashtable_walk_stop - Finish a hash table walk
925  * @iter:	Hash table iterator
926  *
927  * Finish a hash table walk.  Does not reset the iterator to the start of the
928  * hash table.
929  */
930 void rhashtable_walk_stop(struct rhashtable_iter *iter)
931 	__releases(RCU)
932 {
933 	struct rhashtable *ht;
934 	struct bucket_table *tbl = iter->walker.tbl;
935 
936 	if (!tbl)
937 		goto out;
938 
939 	ht = iter->ht;
940 
941 	spin_lock(&ht->lock);
942 	if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
943 		/* This bucket table is being freed, don't re-link it. */
944 		iter->walker.tbl = NULL;
945 	else
946 		list_add(&iter->walker.list, &tbl->walkers);
947 	spin_unlock(&ht->lock);
948 
949 out:
950 	rcu_read_unlock();
951 }
952 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
953 
954 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
955 {
956 	size_t retsize;
957 
958 	if (params->nelem_hint)
959 		retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
960 			      (unsigned long)params->min_size);
961 	else
962 		retsize = max(HASH_DEFAULT_SIZE,
963 			      (unsigned long)params->min_size);
964 
965 	return retsize;
966 }
967 
968 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
969 {
970 	return jhash2(key, length, seed);
971 }
972 
973 /**
974  * rhashtable_init - initialize a new hash table
975  * @ht:		hash table to be initialized
976  * @params:	configuration parameters
977  *
978  * Initializes a new hash table based on the provided configuration
979  * parameters. A table can be configured either with a variable or
980  * fixed length key:
981  *
982  * Configuration Example 1: Fixed length keys
983  * struct test_obj {
984  *	int			key;
985  *	void *			my_member;
986  *	struct rhash_head	node;
987  * };
988  *
989  * struct rhashtable_params params = {
990  *	.head_offset = offsetof(struct test_obj, node),
991  *	.key_offset = offsetof(struct test_obj, key),
992  *	.key_len = sizeof(int),
993  *	.hashfn = jhash,
994  * };
995  *
996  * Configuration Example 2: Variable length keys
997  * struct test_obj {
998  *	[...]
999  *	struct rhash_head	node;
1000  * };
1001  *
1002  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1003  * {
1004  *	struct test_obj *obj = data;
1005  *
1006  *	return [... hash ...];
1007  * }
1008  *
1009  * struct rhashtable_params params = {
1010  *	.head_offset = offsetof(struct test_obj, node),
1011  *	.hashfn = jhash,
1012  *	.obj_hashfn = my_hash_fn,
1013  * };
1014  */
1015 int rhashtable_init(struct rhashtable *ht,
1016 		    const struct rhashtable_params *params)
1017 {
1018 	struct bucket_table *tbl;
1019 	size_t size;
1020 
1021 	if ((!params->key_len && !params->obj_hashfn) ||
1022 	    (params->obj_hashfn && !params->obj_cmpfn))
1023 		return -EINVAL;
1024 
1025 	memset(ht, 0, sizeof(*ht));
1026 	mutex_init(&ht->mutex);
1027 	spin_lock_init(&ht->lock);
1028 	memcpy(&ht->p, params, sizeof(*params));
1029 
1030 	if (params->min_size)
1031 		ht->p.min_size = roundup_pow_of_two(params->min_size);
1032 
1033 	/* Cap total entries at 2^31 to avoid nelems overflow. */
1034 	ht->max_elems = 1u << 31;
1035 
1036 	if (params->max_size) {
1037 		ht->p.max_size = rounddown_pow_of_two(params->max_size);
1038 		if (ht->p.max_size < ht->max_elems / 2)
1039 			ht->max_elems = ht->p.max_size * 2;
1040 	}
1041 
1042 	ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1043 
1044 	size = rounded_hashtable_size(&ht->p);
1045 
1046 	ht->key_len = ht->p.key_len;
1047 	if (!params->hashfn) {
1048 		ht->p.hashfn = jhash;
1049 
1050 		if (!(ht->key_len & (sizeof(u32) - 1))) {
1051 			ht->key_len /= sizeof(u32);
1052 			ht->p.hashfn = rhashtable_jhash2;
1053 		}
1054 	}
1055 
1056 	/*
1057 	 * This is api initialization and thus we need to guarantee the
1058 	 * initial rhashtable allocation. Upon failure, retry with the
1059 	 * smallest possible size with __GFP_NOFAIL semantics.
1060 	 */
1061 	tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1062 	if (unlikely(tbl == NULL)) {
1063 		size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1064 		tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1065 	}
1066 
1067 	atomic_set(&ht->nelems, 0);
1068 
1069 	RCU_INIT_POINTER(ht->tbl, tbl);
1070 
1071 	INIT_WORK(&ht->run_work, rht_deferred_worker);
1072 
1073 	return 0;
1074 }
1075 EXPORT_SYMBOL_GPL(rhashtable_init);
1076 
1077 /**
1078  * rhltable_init - initialize a new hash list table
1079  * @hlt:	hash list table to be initialized
1080  * @params:	configuration parameters
1081  *
1082  * Initializes a new hash list table.
1083  *
1084  * See documentation for rhashtable_init.
1085  */
1086 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1087 {
1088 	int err;
1089 
1090 	err = rhashtable_init(&hlt->ht, params);
1091 	hlt->ht.rhlist = true;
1092 	return err;
1093 }
1094 EXPORT_SYMBOL_GPL(rhltable_init);
1095 
1096 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1097 				void (*free_fn)(void *ptr, void *arg),
1098 				void *arg)
1099 {
1100 	struct rhlist_head *list;
1101 
1102 	if (!ht->rhlist) {
1103 		free_fn(rht_obj(ht, obj), arg);
1104 		return;
1105 	}
1106 
1107 	list = container_of(obj, struct rhlist_head, rhead);
1108 	do {
1109 		obj = &list->rhead;
1110 		list = rht_dereference(list->next, ht);
1111 		free_fn(rht_obj(ht, obj), arg);
1112 	} while (list);
1113 }
1114 
1115 /**
1116  * rhashtable_free_and_destroy - free elements and destroy hash table
1117  * @ht:		the hash table to destroy
1118  * @free_fn:	callback to release resources of element
1119  * @arg:	pointer passed to free_fn
1120  *
1121  * Stops an eventual async resize. If defined, invokes free_fn for each
1122  * element to releasal resources. Please note that RCU protected
1123  * readers may still be accessing the elements. Releasing of resources
1124  * must occur in a compatible manner. Then frees the bucket array.
1125  *
1126  * This function will eventually sleep to wait for an async resize
1127  * to complete. The caller is responsible that no further write operations
1128  * occurs in parallel.
1129  */
1130 void rhashtable_free_and_destroy(struct rhashtable *ht,
1131 				 void (*free_fn)(void *ptr, void *arg),
1132 				 void *arg)
1133 {
1134 	struct bucket_table *tbl, *next_tbl;
1135 	unsigned int i;
1136 
1137 	cancel_work_sync(&ht->run_work);
1138 
1139 	mutex_lock(&ht->mutex);
1140 	tbl = rht_dereference(ht->tbl, ht);
1141 restart:
1142 	if (free_fn) {
1143 		for (i = 0; i < tbl->size; i++) {
1144 			struct rhash_head *pos, *next;
1145 
1146 			cond_resched();
1147 			for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1148 			     next = !rht_is_a_nulls(pos) ?
1149 					rht_dereference(pos->next, ht) : NULL;
1150 			     !rht_is_a_nulls(pos);
1151 			     pos = next,
1152 			     next = !rht_is_a_nulls(pos) ?
1153 					rht_dereference(pos->next, ht) : NULL)
1154 				rhashtable_free_one(ht, pos, free_fn, arg);
1155 		}
1156 	}
1157 
1158 	next_tbl = rht_dereference(tbl->future_tbl, ht);
1159 	bucket_table_free(tbl);
1160 	if (next_tbl) {
1161 		tbl = next_tbl;
1162 		goto restart;
1163 	}
1164 	mutex_unlock(&ht->mutex);
1165 }
1166 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1167 
1168 void rhashtable_destroy(struct rhashtable *ht)
1169 {
1170 	return rhashtable_free_and_destroy(ht, NULL, NULL);
1171 }
1172 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1173 
1174 struct rhash_lock_head __rcu **__rht_bucket_nested(
1175 	const struct bucket_table *tbl, unsigned int hash)
1176 {
1177 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1178 	unsigned int index = hash & ((1 << tbl->nest) - 1);
1179 	unsigned int size = tbl->size >> tbl->nest;
1180 	unsigned int subhash = hash;
1181 	union nested_table *ntbl;
1182 
1183 	ntbl = nested_table_top(tbl);
1184 	ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1185 	subhash >>= tbl->nest;
1186 
1187 	while (ntbl && size > (1 << shift)) {
1188 		index = subhash & ((1 << shift) - 1);
1189 		ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1190 						  tbl, hash);
1191 		size >>= shift;
1192 		subhash >>= shift;
1193 	}
1194 
1195 	if (!ntbl)
1196 		return NULL;
1197 
1198 	return &ntbl[subhash].bucket;
1199 
1200 }
1201 EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1202 
1203 struct rhash_lock_head __rcu **rht_bucket_nested(
1204 	const struct bucket_table *tbl, unsigned int hash)
1205 {
1206 	static struct rhash_lock_head __rcu *rhnull;
1207 
1208 	if (!rhnull)
1209 		INIT_RHT_NULLS_HEAD(rhnull);
1210 	return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1211 }
1212 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1213 
1214 struct rhash_lock_head __rcu **rht_bucket_nested_insert(
1215 	struct rhashtable *ht, struct bucket_table *tbl, unsigned int hash)
1216 {
1217 	const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1218 	unsigned int index = hash & ((1 << tbl->nest) - 1);
1219 	unsigned int size = tbl->size >> tbl->nest;
1220 	union nested_table *ntbl;
1221 
1222 	ntbl = nested_table_top(tbl);
1223 	hash >>= tbl->nest;
1224 	ntbl = nested_table_alloc(ht, &ntbl[index].table,
1225 				  size <= (1 << shift));
1226 
1227 	while (ntbl && size > (1 << shift)) {
1228 		index = hash & ((1 << shift) - 1);
1229 		size >>= shift;
1230 		hash >>= shift;
1231 		ntbl = nested_table_alloc(ht, &ntbl[index].table,
1232 					  size <= (1 << shift));
1233 	}
1234 
1235 	if (!ntbl)
1236 		return NULL;
1237 
1238 	return &ntbl[hash].bucket;
1239 
1240 }
1241 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);
1242