Loading...
Note: File does not exist in v3.15.
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
32union nested_table {
33 union nested_table __rcu *table;
34 struct rhash_lock_head __rcu *bucket;
35};
36
37static 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
47int lockdep_rht_mutex_is_held(struct rhashtable *ht)
48{
49 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
50}
51EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
52
53int 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}
61EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
62#else
63#define ASSERT_RHT_MUTEX(HT)
64#endif
65
66static 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
75static 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
94static 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
109static 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
117static void bucket_table_free_rcu(struct rcu_head *head)
118{
119 bucket_table_free(container_of(head, struct bucket_table, rcu));
120}
121
122static 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
148static 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
177static 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
215static 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
228static 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
276out:
277 return err;
278}
279
280static 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
302static 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
319static 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
358static 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 */
394static 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
414static 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
446static 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
482fail:
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
494static 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
545static 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 return NULL;
588}
589
590static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
591 struct rhash_head *obj)
592{
593 struct bucket_table *new_tbl;
594 struct bucket_table *tbl;
595 struct rhash_lock_head __rcu **bkt;
596 unsigned long flags;
597 unsigned int hash;
598 void *data;
599
600 new_tbl = rcu_dereference(ht->tbl);
601
602 do {
603 tbl = new_tbl;
604 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
605 if (rcu_access_pointer(tbl->future_tbl))
606 /* Failure is OK */
607 bkt = rht_bucket_var(tbl, hash);
608 else
609 bkt = rht_bucket_insert(ht, tbl, hash);
610 if (bkt == NULL) {
611 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
612 data = ERR_PTR(-EAGAIN);
613 } else {
614 bool inserted;
615
616 flags = rht_lock(tbl, bkt);
617 data = rhashtable_lookup_one(ht, bkt, tbl,
618 hash, key, obj);
619 new_tbl = rhashtable_insert_one(ht, bkt, tbl,
620 hash, obj, data);
621 inserted = data && !new_tbl;
622 if (inserted)
623 atomic_inc(&ht->nelems);
624 if (PTR_ERR(new_tbl) != -EEXIST)
625 data = ERR_CAST(new_tbl);
626
627 rht_unlock(tbl, bkt, flags);
628
629 if (inserted && rht_grow_above_75(ht, tbl))
630 schedule_work(&ht->run_work);
631 }
632 } while (!IS_ERR_OR_NULL(new_tbl));
633
634 if (PTR_ERR(data) == -EAGAIN)
635 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
636 -EAGAIN);
637
638 return data;
639}
640
641void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
642 struct rhash_head *obj)
643{
644 void *data;
645
646 do {
647 rcu_read_lock();
648 data = rhashtable_try_insert(ht, key, obj);
649 rcu_read_unlock();
650 } while (PTR_ERR(data) == -EAGAIN);
651
652 return data;
653}
654EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
655
656/**
657 * rhashtable_walk_enter - Initialise an iterator
658 * @ht: Table to walk over
659 * @iter: Hash table Iterator
660 *
661 * This function prepares a hash table walk.
662 *
663 * Note that if you restart a walk after rhashtable_walk_stop you
664 * may see the same object twice. Also, you may miss objects if
665 * there are removals in between rhashtable_walk_stop and the next
666 * call to rhashtable_walk_start.
667 *
668 * For a completely stable walk you should construct your own data
669 * structure outside the hash table.
670 *
671 * This function may be called from any process context, including
672 * non-preemptable context, but cannot be called from softirq or
673 * hardirq context.
674 *
675 * You must call rhashtable_walk_exit after this function returns.
676 */
677void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
678{
679 iter->ht = ht;
680 iter->p = NULL;
681 iter->slot = 0;
682 iter->skip = 0;
683 iter->end_of_table = 0;
684
685 spin_lock(&ht->lock);
686 iter->walker.tbl =
687 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
688 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
689 spin_unlock(&ht->lock);
690}
691EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
692
693/**
694 * rhashtable_walk_exit - Free an iterator
695 * @iter: Hash table Iterator
696 *
697 * This function frees resources allocated by rhashtable_walk_enter.
698 */
699void rhashtable_walk_exit(struct rhashtable_iter *iter)
700{
701 spin_lock(&iter->ht->lock);
702 if (iter->walker.tbl)
703 list_del(&iter->walker.list);
704 spin_unlock(&iter->ht->lock);
705}
706EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
707
708/**
709 * rhashtable_walk_start_check - Start a hash table walk
710 * @iter: Hash table iterator
711 *
712 * Start a hash table walk at the current iterator position. Note that we take
713 * the RCU lock in all cases including when we return an error. So you must
714 * always call rhashtable_walk_stop to clean up.
715 *
716 * Returns zero if successful.
717 *
718 * Returns -EAGAIN if resize event occurred. Note that the iterator
719 * will rewind back to the beginning and you may use it immediately
720 * by calling rhashtable_walk_next.
721 *
722 * rhashtable_walk_start is defined as an inline variant that returns
723 * void. This is preferred in cases where the caller would ignore
724 * resize events and always continue.
725 */
726int rhashtable_walk_start_check(struct rhashtable_iter *iter)
727 __acquires(RCU)
728{
729 struct rhashtable *ht = iter->ht;
730 bool rhlist = ht->rhlist;
731
732 rcu_read_lock();
733
734 spin_lock(&ht->lock);
735 if (iter->walker.tbl)
736 list_del(&iter->walker.list);
737 spin_unlock(&ht->lock);
738
739 if (iter->end_of_table)
740 return 0;
741 if (!iter->walker.tbl) {
742 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
743 iter->slot = 0;
744 iter->skip = 0;
745 return -EAGAIN;
746 }
747
748 if (iter->p && !rhlist) {
749 /*
750 * We need to validate that 'p' is still in the table, and
751 * if so, update 'skip'
752 */
753 struct rhash_head *p;
754 int skip = 0;
755 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
756 skip++;
757 if (p == iter->p) {
758 iter->skip = skip;
759 goto found;
760 }
761 }
762 iter->p = NULL;
763 } else if (iter->p && rhlist) {
764 /* Need to validate that 'list' is still in the table, and
765 * if so, update 'skip' and 'p'.
766 */
767 struct rhash_head *p;
768 struct rhlist_head *list;
769 int skip = 0;
770 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
771 for (list = container_of(p, struct rhlist_head, rhead);
772 list;
773 list = rcu_dereference(list->next)) {
774 skip++;
775 if (list == iter->list) {
776 iter->p = p;
777 iter->skip = skip;
778 goto found;
779 }
780 }
781 }
782 iter->p = NULL;
783 }
784found:
785 return 0;
786}
787EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
788
789/**
790 * __rhashtable_walk_find_next - Find the next element in a table (or the first
791 * one in case of a new walk).
792 *
793 * @iter: Hash table iterator
794 *
795 * Returns the found object or NULL when the end of the table is reached.
796 *
797 * Returns -EAGAIN if resize event occurred.
798 */
799static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
800{
801 struct bucket_table *tbl = iter->walker.tbl;
802 struct rhlist_head *list = iter->list;
803 struct rhashtable *ht = iter->ht;
804 struct rhash_head *p = iter->p;
805 bool rhlist = ht->rhlist;
806
807 if (!tbl)
808 return NULL;
809
810 for (; iter->slot < tbl->size; iter->slot++) {
811 int skip = iter->skip;
812
813 rht_for_each_rcu(p, tbl, iter->slot) {
814 if (rhlist) {
815 list = container_of(p, struct rhlist_head,
816 rhead);
817 do {
818 if (!skip)
819 goto next;
820 skip--;
821 list = rcu_dereference(list->next);
822 } while (list);
823
824 continue;
825 }
826 if (!skip)
827 break;
828 skip--;
829 }
830
831next:
832 if (!rht_is_a_nulls(p)) {
833 iter->skip++;
834 iter->p = p;
835 iter->list = list;
836 return rht_obj(ht, rhlist ? &list->rhead : p);
837 }
838
839 iter->skip = 0;
840 }
841
842 iter->p = NULL;
843
844 /* Ensure we see any new tables. */
845 smp_rmb();
846
847 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
848 if (iter->walker.tbl) {
849 iter->slot = 0;
850 iter->skip = 0;
851 return ERR_PTR(-EAGAIN);
852 } else {
853 iter->end_of_table = true;
854 }
855
856 return NULL;
857}
858
859/**
860 * rhashtable_walk_next - Return the next object and advance the iterator
861 * @iter: Hash table iterator
862 *
863 * Note that you must call rhashtable_walk_stop when you are finished
864 * with the walk.
865 *
866 * Returns the next object or NULL when the end of the table is reached.
867 *
868 * Returns -EAGAIN if resize event occurred. Note that the iterator
869 * will rewind back to the beginning and you may continue to use it.
870 */
871void *rhashtable_walk_next(struct rhashtable_iter *iter)
872{
873 struct rhlist_head *list = iter->list;
874 struct rhashtable *ht = iter->ht;
875 struct rhash_head *p = iter->p;
876 bool rhlist = ht->rhlist;
877
878 if (p) {
879 if (!rhlist || !(list = rcu_dereference(list->next))) {
880 p = rcu_dereference(p->next);
881 list = container_of(p, struct rhlist_head, rhead);
882 }
883 if (!rht_is_a_nulls(p)) {
884 iter->skip++;
885 iter->p = p;
886 iter->list = list;
887 return rht_obj(ht, rhlist ? &list->rhead : p);
888 }
889
890 /* At the end of this slot, switch to next one and then find
891 * next entry from that point.
892 */
893 iter->skip = 0;
894 iter->slot++;
895 }
896
897 return __rhashtable_walk_find_next(iter);
898}
899EXPORT_SYMBOL_GPL(rhashtable_walk_next);
900
901/**
902 * rhashtable_walk_peek - Return the next object but don't advance the iterator
903 * @iter: Hash table iterator
904 *
905 * Returns the next object or NULL when the end of the table is reached.
906 *
907 * Returns -EAGAIN if resize event occurred. Note that the iterator
908 * will rewind back to the beginning and you may continue to use it.
909 */
910void *rhashtable_walk_peek(struct rhashtable_iter *iter)
911{
912 struct rhlist_head *list = iter->list;
913 struct rhashtable *ht = iter->ht;
914 struct rhash_head *p = iter->p;
915
916 if (p)
917 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
918
919 /* No object found in current iter, find next one in the table. */
920
921 if (iter->skip) {
922 /* A nonzero skip value points to the next entry in the table
923 * beyond that last one that was found. Decrement skip so
924 * we find the current value. __rhashtable_walk_find_next
925 * will restore the original value of skip assuming that
926 * the table hasn't changed.
927 */
928 iter->skip--;
929 }
930
931 return __rhashtable_walk_find_next(iter);
932}
933EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
934
935/**
936 * rhashtable_walk_stop - Finish a hash table walk
937 * @iter: Hash table iterator
938 *
939 * Finish a hash table walk. Does not reset the iterator to the start of the
940 * hash table.
941 */
942void rhashtable_walk_stop(struct rhashtable_iter *iter)
943 __releases(RCU)
944{
945 struct rhashtable *ht;
946 struct bucket_table *tbl = iter->walker.tbl;
947
948 if (!tbl)
949 goto out;
950
951 ht = iter->ht;
952
953 spin_lock(&ht->lock);
954 if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
955 /* This bucket table is being freed, don't re-link it. */
956 iter->walker.tbl = NULL;
957 else
958 list_add(&iter->walker.list, &tbl->walkers);
959 spin_unlock(&ht->lock);
960
961out:
962 rcu_read_unlock();
963}
964EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
965
966static size_t rounded_hashtable_size(const struct rhashtable_params *params)
967{
968 size_t retsize;
969
970 if (params->nelem_hint)
971 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
972 (unsigned long)params->min_size);
973 else
974 retsize = max(HASH_DEFAULT_SIZE,
975 (unsigned long)params->min_size);
976
977 return retsize;
978}
979
980static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
981{
982 return jhash2(key, length, seed);
983}
984
985/**
986 * rhashtable_init - initialize a new hash table
987 * @ht: hash table to be initialized
988 * @params: configuration parameters
989 *
990 * Initializes a new hash table based on the provided configuration
991 * parameters. A table can be configured either with a variable or
992 * fixed length key:
993 *
994 * Configuration Example 1: Fixed length keys
995 * struct test_obj {
996 * int key;
997 * void * my_member;
998 * struct rhash_head node;
999 * };
1000 *
1001 * struct rhashtable_params params = {
1002 * .head_offset = offsetof(struct test_obj, node),
1003 * .key_offset = offsetof(struct test_obj, key),
1004 * .key_len = sizeof(int),
1005 * .hashfn = jhash,
1006 * };
1007 *
1008 * Configuration Example 2: Variable length keys
1009 * struct test_obj {
1010 * [...]
1011 * struct rhash_head node;
1012 * };
1013 *
1014 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1015 * {
1016 * struct test_obj *obj = data;
1017 *
1018 * return [... hash ...];
1019 * }
1020 *
1021 * struct rhashtable_params params = {
1022 * .head_offset = offsetof(struct test_obj, node),
1023 * .hashfn = jhash,
1024 * .obj_hashfn = my_hash_fn,
1025 * };
1026 */
1027int rhashtable_init_noprof(struct rhashtable *ht,
1028 const struct rhashtable_params *params)
1029{
1030 struct bucket_table *tbl;
1031 size_t size;
1032
1033 if ((!params->key_len && !params->obj_hashfn) ||
1034 (params->obj_hashfn && !params->obj_cmpfn))
1035 return -EINVAL;
1036
1037 memset(ht, 0, sizeof(*ht));
1038 mutex_init(&ht->mutex);
1039 spin_lock_init(&ht->lock);
1040 memcpy(&ht->p, params, sizeof(*params));
1041
1042 alloc_tag_record(ht->alloc_tag);
1043
1044 if (params->min_size)
1045 ht->p.min_size = roundup_pow_of_two(params->min_size);
1046
1047 /* Cap total entries at 2^31 to avoid nelems overflow. */
1048 ht->max_elems = 1u << 31;
1049
1050 if (params->max_size) {
1051 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1052 if (ht->p.max_size < ht->max_elems / 2)
1053 ht->max_elems = ht->p.max_size * 2;
1054 }
1055
1056 ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1057
1058 size = rounded_hashtable_size(&ht->p);
1059
1060 ht->key_len = ht->p.key_len;
1061 if (!params->hashfn) {
1062 ht->p.hashfn = jhash;
1063
1064 if (!(ht->key_len & (sizeof(u32) - 1))) {
1065 ht->key_len /= sizeof(u32);
1066 ht->p.hashfn = rhashtable_jhash2;
1067 }
1068 }
1069
1070 /*
1071 * This is api initialization and thus we need to guarantee the
1072 * initial rhashtable allocation. Upon failure, retry with the
1073 * smallest possible size with __GFP_NOFAIL semantics.
1074 */
1075 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1076 if (unlikely(tbl == NULL)) {
1077 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1078 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1079 }
1080
1081 atomic_set(&ht->nelems, 0);
1082
1083 RCU_INIT_POINTER(ht->tbl, tbl);
1084
1085 INIT_WORK(&ht->run_work, rht_deferred_worker);
1086
1087 return 0;
1088}
1089EXPORT_SYMBOL_GPL(rhashtable_init_noprof);
1090
1091/**
1092 * rhltable_init - initialize a new hash list table
1093 * @hlt: hash list table to be initialized
1094 * @params: configuration parameters
1095 *
1096 * Initializes a new hash list table.
1097 *
1098 * See documentation for rhashtable_init.
1099 */
1100int rhltable_init_noprof(struct rhltable *hlt, const struct rhashtable_params *params)
1101{
1102 int err;
1103
1104 err = rhashtable_init_noprof(&hlt->ht, params);
1105 hlt->ht.rhlist = true;
1106 return err;
1107}
1108EXPORT_SYMBOL_GPL(rhltable_init_noprof);
1109
1110static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1111 void (*free_fn)(void *ptr, void *arg),
1112 void *arg)
1113{
1114 struct rhlist_head *list;
1115
1116 if (!ht->rhlist) {
1117 free_fn(rht_obj(ht, obj), arg);
1118 return;
1119 }
1120
1121 list = container_of(obj, struct rhlist_head, rhead);
1122 do {
1123 obj = &list->rhead;
1124 list = rht_dereference(list->next, ht);
1125 free_fn(rht_obj(ht, obj), arg);
1126 } while (list);
1127}
1128
1129/**
1130 * rhashtable_free_and_destroy - free elements and destroy hash table
1131 * @ht: the hash table to destroy
1132 * @free_fn: callback to release resources of element
1133 * @arg: pointer passed to free_fn
1134 *
1135 * Stops an eventual async resize. If defined, invokes free_fn for each
1136 * element to releasal resources. Please note that RCU protected
1137 * readers may still be accessing the elements. Releasing of resources
1138 * must occur in a compatible manner. Then frees the bucket array.
1139 *
1140 * This function will eventually sleep to wait for an async resize
1141 * to complete. The caller is responsible that no further write operations
1142 * occurs in parallel.
1143 */
1144void rhashtable_free_and_destroy(struct rhashtable *ht,
1145 void (*free_fn)(void *ptr, void *arg),
1146 void *arg)
1147{
1148 struct bucket_table *tbl, *next_tbl;
1149 unsigned int i;
1150
1151 cancel_work_sync(&ht->run_work);
1152
1153 mutex_lock(&ht->mutex);
1154 tbl = rht_dereference(ht->tbl, ht);
1155restart:
1156 if (free_fn) {
1157 for (i = 0; i < tbl->size; i++) {
1158 struct rhash_head *pos, *next;
1159
1160 cond_resched();
1161 for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1162 next = !rht_is_a_nulls(pos) ?
1163 rht_dereference(pos->next, ht) : NULL;
1164 !rht_is_a_nulls(pos);
1165 pos = next,
1166 next = !rht_is_a_nulls(pos) ?
1167 rht_dereference(pos->next, ht) : NULL)
1168 rhashtable_free_one(ht, pos, free_fn, arg);
1169 }
1170 }
1171
1172 next_tbl = rht_dereference(tbl->future_tbl, ht);
1173 bucket_table_free(tbl);
1174 if (next_tbl) {
1175 tbl = next_tbl;
1176 goto restart;
1177 }
1178 mutex_unlock(&ht->mutex);
1179}
1180EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1181
1182void rhashtable_destroy(struct rhashtable *ht)
1183{
1184 return rhashtable_free_and_destroy(ht, NULL, NULL);
1185}
1186EXPORT_SYMBOL_GPL(rhashtable_destroy);
1187
1188struct rhash_lock_head __rcu **__rht_bucket_nested(
1189 const struct bucket_table *tbl, unsigned int hash)
1190{
1191 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1192 unsigned int index = hash & ((1 << tbl->nest) - 1);
1193 unsigned int size = tbl->size >> tbl->nest;
1194 unsigned int subhash = hash;
1195 union nested_table *ntbl;
1196
1197 ntbl = nested_table_top(tbl);
1198 ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1199 subhash >>= tbl->nest;
1200
1201 while (ntbl && size > (1 << shift)) {
1202 index = subhash & ((1 << shift) - 1);
1203 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1204 tbl, hash);
1205 size >>= shift;
1206 subhash >>= shift;
1207 }
1208
1209 if (!ntbl)
1210 return NULL;
1211
1212 return &ntbl[subhash].bucket;
1213
1214}
1215EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1216
1217struct rhash_lock_head __rcu **rht_bucket_nested(
1218 const struct bucket_table *tbl, unsigned int hash)
1219{
1220 static struct rhash_lock_head __rcu *rhnull;
1221
1222 if (!rhnull)
1223 INIT_RHT_NULLS_HEAD(rhnull);
1224 return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1225}
1226EXPORT_SYMBOL_GPL(rht_bucket_nested);
1227
1228struct rhash_lock_head __rcu **rht_bucket_nested_insert(
1229 struct rhashtable *ht, struct bucket_table *tbl, unsigned int hash)
1230{
1231 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1232 unsigned int index = hash & ((1 << tbl->nest) - 1);
1233 unsigned int size = tbl->size >> tbl->nest;
1234 union nested_table *ntbl;
1235
1236 ntbl = nested_table_top(tbl);
1237 hash >>= tbl->nest;
1238 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1239 size <= (1 << shift));
1240
1241 while (ntbl && size > (1 << shift)) {
1242 index = hash & ((1 << shift) - 1);
1243 size >>= shift;
1244 hash >>= shift;
1245 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1246 size <= (1 << shift));
1247 }
1248
1249 if (!ntbl)
1250 return NULL;
1251
1252 return &ntbl[hash].bucket;
1253
1254}
1255EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);