Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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_align_noprof(struct_size(tbl, buckets, nbuckets),
188 1, 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 __must_hold(&ht->mutex)
362{
363 struct bucket_table *new_tbl;
364 int err;
365
366 ASSERT_RHT_MUTEX(ht);
367
368 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
369 if (new_tbl == NULL)
370 return -ENOMEM;
371
372 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
373 if (err)
374 bucket_table_free(new_tbl);
375
376 return err;
377}
378
379/**
380 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
381 * @ht: the hash table to shrink
382 *
383 * This function shrinks the hash table to fit, i.e., the smallest
384 * size would not cause it to expand right away automatically.
385 *
386 * The caller must ensure that no concurrent resizing occurs by holding
387 * ht->mutex.
388 *
389 * The caller must ensure that no concurrent table mutations take place.
390 * It is however valid to have concurrent lookups if they are RCU protected.
391 *
392 * It is valid to have concurrent insertions and deletions protected by per
393 * bucket locks or concurrent RCU protected lookups and traversals.
394 */
395static int rhashtable_shrink(struct rhashtable *ht)
396 __must_hold(&ht->mutex)
397{
398 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
399 unsigned int nelems = atomic_read(&ht->nelems);
400 unsigned int size = 0;
401
402 if (nelems)
403 size = roundup_pow_of_two(nelems * 3 / 2);
404 if (size < ht->p.min_size)
405 size = ht->p.min_size;
406
407 if (old_tbl->size <= size)
408 return 0;
409
410 if (rht_dereference(old_tbl->future_tbl, ht))
411 return -EEXIST;
412
413 return rhashtable_rehash_alloc(ht, old_tbl, size);
414}
415
416static void rht_deferred_worker(struct work_struct *work)
417{
418 struct rhashtable *ht;
419 struct bucket_table *tbl;
420 int err = 0;
421
422 ht = container_of(work, struct rhashtable, run_work);
423 mutex_lock(&ht->mutex);
424
425 tbl = rht_dereference(ht->tbl, ht);
426 tbl = rhashtable_last_table(ht, tbl);
427
428 if (rht_grow_above_75(ht, tbl))
429 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
430 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
431 err = rhashtable_shrink(ht);
432 else if (tbl->nest)
433 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
434
435 if (!err || err == -EEXIST) {
436 int nerr;
437
438 nerr = rhashtable_rehash_table(ht);
439 err = err ?: nerr;
440 }
441
442 mutex_unlock(&ht->mutex);
443
444 if (err)
445 schedule_work(&ht->run_work);
446}
447
448static int rhashtable_insert_rehash(struct rhashtable *ht,
449 struct bucket_table *tbl)
450{
451 struct bucket_table *old_tbl;
452 struct bucket_table *new_tbl;
453 unsigned int size;
454 int err;
455
456 old_tbl = rht_dereference_rcu(ht->tbl, ht);
457
458 size = tbl->size;
459
460 err = -EBUSY;
461
462 if (rht_grow_above_75(ht, tbl))
463 size *= 2;
464 /* Do not schedule more than one rehash */
465 else if (old_tbl != tbl)
466 goto fail;
467
468 err = -ENOMEM;
469
470 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
471 if (new_tbl == NULL)
472 goto fail;
473
474 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
475 if (err) {
476 bucket_table_free(new_tbl);
477 if (err == -EEXIST)
478 err = 0;
479 } else
480 schedule_work(&ht->run_work);
481
482 return err;
483
484fail:
485 /* Do not fail the insert if someone else did a rehash. */
486 if (likely(rcu_access_pointer(tbl->future_tbl)))
487 return 0;
488
489 /* Schedule async rehash to retry allocation in process context. */
490 if (err == -ENOMEM)
491 schedule_work(&ht->run_work);
492
493 return err;
494}
495
496static void *rhashtable_lookup_one(struct rhashtable *ht,
497 struct rhash_lock_head __rcu **bkt,
498 struct bucket_table *tbl, unsigned int hash,
499 const void *key, struct rhash_head *obj)
500{
501 struct rhashtable_compare_arg arg = {
502 .ht = ht,
503 .key = key,
504 };
505 struct rhash_head __rcu **pprev = NULL;
506 struct rhash_head *head;
507 int elasticity;
508
509 elasticity = RHT_ELASTICITY;
510 rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
511 struct rhlist_head *list;
512 struct rhlist_head *plist;
513
514 elasticity--;
515 if (!key ||
516 (ht->p.obj_cmpfn ?
517 ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
518 rhashtable_compare(&arg, rht_obj(ht, head)))) {
519 pprev = &head->next;
520 continue;
521 }
522
523 if (!ht->rhlist)
524 return rht_obj(ht, head);
525
526 list = container_of(obj, struct rhlist_head, rhead);
527 plist = container_of(head, struct rhlist_head, rhead);
528
529 RCU_INIT_POINTER(list->next, plist);
530 head = rht_dereference_bucket(head->next, tbl, hash);
531 RCU_INIT_POINTER(list->rhead.next, head);
532 if (pprev)
533 rcu_assign_pointer(*pprev, obj);
534 else
535 /* Need to preserve the bit lock */
536 rht_assign_locked(bkt, obj);
537
538 return NULL;
539 }
540
541 if (elasticity <= 0)
542 return ERR_PTR(-EAGAIN);
543
544 return ERR_PTR(-ENOENT);
545}
546
547static struct bucket_table *rhashtable_insert_one(
548 struct rhashtable *ht, struct rhash_lock_head __rcu **bkt,
549 struct bucket_table *tbl, unsigned int hash, struct rhash_head *obj,
550 void *data)
551{
552 struct bucket_table *new_tbl;
553 struct rhash_head *head;
554
555 if (!IS_ERR_OR_NULL(data))
556 return ERR_PTR(-EEXIST);
557
558 if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
559 return ERR_CAST(data);
560
561 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
562 if (new_tbl)
563 return new_tbl;
564
565 if (PTR_ERR(data) != -ENOENT)
566 return ERR_CAST(data);
567
568 if (unlikely(rht_grow_above_max(ht, tbl)))
569 return ERR_PTR(-E2BIG);
570
571 if (unlikely(rht_grow_above_100(ht, tbl)))
572 return ERR_PTR(-EAGAIN);
573
574 head = rht_ptr(bkt, tbl, hash);
575
576 RCU_INIT_POINTER(obj->next, head);
577 if (ht->rhlist) {
578 struct rhlist_head *list;
579
580 list = container_of(obj, struct rhlist_head, rhead);
581 RCU_INIT_POINTER(list->next, NULL);
582 }
583
584 /* bkt is always the head of the list, so it holds
585 * the lock, which we need to preserve
586 */
587 rht_assign_locked(bkt, obj);
588
589 return NULL;
590}
591
592static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
593 struct rhash_head *obj)
594{
595 struct bucket_table *new_tbl;
596 struct bucket_table *tbl;
597 struct rhash_lock_head __rcu **bkt;
598 unsigned long flags;
599 unsigned int hash;
600 void *data;
601
602 new_tbl = rcu_dereference(ht->tbl);
603
604 do {
605 tbl = new_tbl;
606 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
607 if (rcu_access_pointer(tbl->future_tbl))
608 /* Failure is OK */
609 bkt = rht_bucket_var(tbl, hash);
610 else
611 bkt = rht_bucket_insert(ht, tbl, hash);
612 if (bkt == NULL) {
613 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
614 data = ERR_PTR(-EAGAIN);
615 } else {
616 bool inserted;
617
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 inserted = data && !new_tbl;
624 if (inserted)
625 atomic_inc(&ht->nelems);
626 if (PTR_ERR(new_tbl) != -EEXIST)
627 data = ERR_CAST(new_tbl);
628
629 rht_unlock(tbl, bkt, flags);
630
631 if (inserted && rht_grow_above_75(ht, tbl))
632 schedule_work(&ht->run_work);
633 }
634 } while (!IS_ERR_OR_NULL(new_tbl));
635
636 if (PTR_ERR(data) == -EAGAIN)
637 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
638 -EAGAIN);
639
640 return data;
641}
642
643void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
644 struct rhash_head *obj)
645{
646 void *data;
647
648 do {
649 rcu_read_lock();
650 data = rhashtable_try_insert(ht, key, obj);
651 rcu_read_unlock();
652 } while (PTR_ERR(data) == -EAGAIN);
653
654 return data;
655}
656EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
657
658/**
659 * rhashtable_walk_enter - Initialise an iterator
660 * @ht: Table to walk over
661 * @iter: Hash table Iterator
662 *
663 * This function prepares a hash table walk.
664 *
665 * Note that if you restart a walk after rhashtable_walk_stop you
666 * may see the same object twice. Also, you may miss objects if
667 * there are removals in between rhashtable_walk_stop and the next
668 * call to rhashtable_walk_start.
669 *
670 * For a completely stable walk you should construct your own data
671 * structure outside the hash table.
672 *
673 * This function may be called from any process context, including
674 * non-preemptible context, but cannot be called from softirq or
675 * hardirq context.
676 *
677 * You must call rhashtable_walk_exit after this function returns.
678 */
679void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
680{
681 iter->ht = ht;
682 iter->p = NULL;
683 iter->slot = 0;
684 iter->skip = 0;
685 iter->end_of_table = 0;
686
687 spin_lock(&ht->lock);
688 iter->walker.tbl =
689 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
690 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
691 spin_unlock(&ht->lock);
692}
693EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
694
695/**
696 * rhashtable_walk_exit - Free an iterator
697 * @iter: Hash table Iterator
698 *
699 * This function frees resources allocated by rhashtable_walk_enter.
700 */
701void rhashtable_walk_exit(struct rhashtable_iter *iter)
702{
703 spin_lock(&iter->ht->lock);
704 if (iter->walker.tbl)
705 list_del(&iter->walker.list);
706 spin_unlock(&iter->ht->lock);
707}
708EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
709
710/**
711 * rhashtable_walk_start_check - Start a hash table walk
712 * @iter: Hash table iterator
713 *
714 * Start a hash table walk at the current iterator position. Note that we take
715 * the RCU lock in all cases including when we return an error. So you must
716 * always call rhashtable_walk_stop to clean up.
717 *
718 * Returns zero if successful.
719 *
720 * Returns -EAGAIN if resize event occurred. Note that the iterator
721 * will rewind back to the beginning and you may use it immediately
722 * by calling rhashtable_walk_next.
723 *
724 * rhashtable_walk_start is defined as an inline variant that returns
725 * void. This is preferred in cases where the caller would ignore
726 * resize events and always continue.
727 */
728int rhashtable_walk_start_check(struct rhashtable_iter *iter)
729 __acquires_shared(RCU)
730{
731 struct rhashtable *ht = iter->ht;
732 bool rhlist = ht->rhlist;
733
734 rcu_read_lock();
735
736 spin_lock(&ht->lock);
737 if (iter->walker.tbl)
738 list_del(&iter->walker.list);
739 spin_unlock(&ht->lock);
740
741 if (iter->end_of_table)
742 return 0;
743 if (!iter->walker.tbl) {
744 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
745 iter->slot = 0;
746 iter->skip = 0;
747 return -EAGAIN;
748 }
749
750 if (iter->p && !rhlist) {
751 /*
752 * We need to validate that 'p' is still in the table, and
753 * if so, update 'skip'
754 */
755 struct rhash_head *p;
756 int skip = 0;
757 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
758 skip++;
759 if (p == iter->p) {
760 iter->skip = skip;
761 goto found;
762 }
763 }
764 iter->p = NULL;
765 } else if (iter->p && rhlist) {
766 /* Need to validate that 'list' is still in the table, and
767 * if so, update 'skip' and 'p'.
768 */
769 struct rhash_head *p;
770 struct rhlist_head *list;
771 int skip = 0;
772 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
773 for (list = container_of(p, struct rhlist_head, rhead);
774 list;
775 list = rcu_dereference(list->next)) {
776 skip++;
777 if (list == iter->list) {
778 iter->p = p;
779 iter->skip = skip;
780 goto found;
781 }
782 }
783 }
784 iter->p = NULL;
785 }
786found:
787 return 0;
788}
789EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
790
791/**
792 * __rhashtable_walk_find_next - Find the next element in a table (or the first
793 * one in case of a new walk).
794 *
795 * @iter: Hash table iterator
796 *
797 * Returns the found object or NULL when the end of the table is reached.
798 *
799 * Returns -EAGAIN if resize event occurred.
800 */
801static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
802{
803 struct bucket_table *tbl = iter->walker.tbl;
804 struct rhlist_head *list = iter->list;
805 struct rhashtable *ht = iter->ht;
806 struct rhash_head *p = iter->p;
807 bool rhlist = ht->rhlist;
808
809 if (!tbl)
810 return NULL;
811
812 for (; iter->slot < tbl->size; iter->slot++) {
813 int skip = iter->skip;
814
815 rht_for_each_rcu(p, tbl, iter->slot) {
816 if (rhlist) {
817 list = container_of(p, struct rhlist_head,
818 rhead);
819 do {
820 if (!skip)
821 goto next;
822 skip--;
823 list = rcu_dereference(list->next);
824 } while (list);
825
826 continue;
827 }
828 if (!skip)
829 break;
830 skip--;
831 }
832
833next:
834 if (!rht_is_a_nulls(p)) {
835 iter->skip++;
836 iter->p = p;
837 iter->list = list;
838 return rht_obj(ht, rhlist ? &list->rhead : p);
839 }
840
841 iter->skip = 0;
842 }
843
844 iter->p = NULL;
845
846 /* Ensure we see any new tables. */
847 smp_rmb();
848
849 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
850 if (iter->walker.tbl) {
851 iter->slot = 0;
852 iter->skip = 0;
853 return ERR_PTR(-EAGAIN);
854 } else {
855 iter->end_of_table = true;
856 }
857
858 return NULL;
859}
860
861/**
862 * rhashtable_walk_next - Return the next object and advance the iterator
863 * @iter: Hash table iterator
864 *
865 * Note that you must call rhashtable_walk_stop when you are finished
866 * with the walk.
867 *
868 * Returns the next object or NULL when the end of the table is reached.
869 *
870 * Returns -EAGAIN if resize event occurred. Note that the iterator
871 * will rewind back to the beginning and you may continue to use it.
872 */
873void *rhashtable_walk_next(struct rhashtable_iter *iter)
874{
875 struct rhlist_head *list = iter->list;
876 struct rhashtable *ht = iter->ht;
877 struct rhash_head *p = iter->p;
878 bool rhlist = ht->rhlist;
879
880 if (p) {
881 if (!rhlist || !(list = rcu_dereference(list->next))) {
882 p = rcu_dereference(p->next);
883 list = container_of(p, struct rhlist_head, rhead);
884 }
885 if (!rht_is_a_nulls(p)) {
886 iter->skip++;
887 iter->p = p;
888 iter->list = list;
889 return rht_obj(ht, rhlist ? &list->rhead : p);
890 }
891
892 /* At the end of this slot, switch to next one and then find
893 * next entry from that point.
894 */
895 iter->skip = 0;
896 iter->slot++;
897 }
898
899 return __rhashtable_walk_find_next(iter);
900}
901EXPORT_SYMBOL_GPL(rhashtable_walk_next);
902
903/**
904 * rhashtable_walk_peek - Return the next object but don't advance the iterator
905 * @iter: Hash table iterator
906 *
907 * Returns the next object or NULL when the end of the table is reached.
908 *
909 * Returns -EAGAIN if resize event occurred. Note that the iterator
910 * will rewind back to the beginning and you may continue to use it.
911 */
912void *rhashtable_walk_peek(struct rhashtable_iter *iter)
913{
914 struct rhlist_head *list = iter->list;
915 struct rhashtable *ht = iter->ht;
916 struct rhash_head *p = iter->p;
917
918 if (p)
919 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
920
921 /* No object found in current iter, find next one in the table. */
922
923 if (iter->skip) {
924 /* A nonzero skip value points to the next entry in the table
925 * beyond that last one that was found. Decrement skip so
926 * we find the current value. __rhashtable_walk_find_next
927 * will restore the original value of skip assuming that
928 * the table hasn't changed.
929 */
930 iter->skip--;
931 }
932
933 return __rhashtable_walk_find_next(iter);
934}
935EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
936
937/**
938 * rhashtable_walk_stop - Finish a hash table walk
939 * @iter: Hash table iterator
940 *
941 * Finish a hash table walk. Does not reset the iterator to the start of the
942 * hash table.
943 */
944void rhashtable_walk_stop(struct rhashtable_iter *iter)
945{
946 struct rhashtable *ht;
947 struct bucket_table *tbl = iter->walker.tbl;
948
949 if (!tbl)
950 goto out;
951
952 ht = iter->ht;
953
954 spin_lock(&ht->lock);
955 if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
956 /* This bucket table is being freed, don't re-link it. */
957 iter->walker.tbl = NULL;
958 else
959 list_add(&iter->walker.list, &tbl->walkers);
960 spin_unlock(&ht->lock);
961
962out:
963 rcu_read_unlock();
964}
965EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
966
967static size_t rounded_hashtable_size(const struct rhashtable_params *params)
968{
969 size_t retsize;
970
971 if (params->nelem_hint)
972 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
973 (unsigned long)params->min_size);
974 else
975 retsize = max(HASH_DEFAULT_SIZE,
976 (unsigned long)params->min_size);
977
978 return retsize;
979}
980
981static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
982{
983 return jhash2(key, length, seed);
984}
985
986/**
987 * rhashtable_init - initialize a new hash table
988 * @ht: hash table to be initialized
989 * @params: configuration parameters
990 *
991 * Initializes a new hash table based on the provided configuration
992 * parameters. A table can be configured either with a variable or
993 * fixed length key:
994 *
995 * Configuration Example 1: Fixed length keys
996 * struct test_obj {
997 * int key;
998 * void * my_member;
999 * struct rhash_head node;
1000 * };
1001 *
1002 * struct rhashtable_params params = {
1003 * .head_offset = offsetof(struct test_obj, node),
1004 * .key_offset = offsetof(struct test_obj, key),
1005 * .key_len = sizeof(int),
1006 * .hashfn = jhash,
1007 * };
1008 *
1009 * Configuration Example 2: Variable length keys
1010 * struct test_obj {
1011 * [...]
1012 * struct rhash_head node;
1013 * };
1014 *
1015 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1016 * {
1017 * struct test_obj *obj = data;
1018 *
1019 * return [... hash ...];
1020 * }
1021 *
1022 * struct rhashtable_params params = {
1023 * .head_offset = offsetof(struct test_obj, node),
1024 * .hashfn = jhash,
1025 * .obj_hashfn = my_hash_fn,
1026 * };
1027 */
1028int rhashtable_init_noprof(struct rhashtable *ht,
1029 const struct rhashtable_params *params)
1030{
1031 struct bucket_table *tbl;
1032 size_t size;
1033
1034 if ((!params->key_len && !params->obj_hashfn) ||
1035 (params->obj_hashfn && !params->obj_cmpfn))
1036 return -EINVAL;
1037
1038 memset(ht, 0, sizeof(*ht));
1039 mutex_init(&ht->mutex);
1040 spin_lock_init(&ht->lock);
1041 memcpy(&ht->p, params, sizeof(*params));
1042
1043 alloc_tag_record(ht->alloc_tag);
1044
1045 if (params->min_size)
1046 ht->p.min_size = roundup_pow_of_two(params->min_size);
1047
1048 /* Cap total entries at 2^31 to avoid nelems overflow. */
1049 ht->max_elems = 1u << 31;
1050
1051 if (params->max_size) {
1052 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1053 if (ht->p.max_size < ht->max_elems / 2)
1054 ht->max_elems = ht->p.max_size * 2;
1055 }
1056
1057 ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1058
1059 size = rounded_hashtable_size(&ht->p);
1060
1061 ht->key_len = ht->p.key_len;
1062 if (!params->hashfn) {
1063 ht->p.hashfn = jhash;
1064
1065 if (!(ht->key_len & (sizeof(u32) - 1))) {
1066 ht->key_len /= sizeof(u32);
1067 ht->p.hashfn = rhashtable_jhash2;
1068 }
1069 }
1070
1071 /*
1072 * This is api initialization and thus we need to guarantee the
1073 * initial rhashtable allocation. Upon failure, retry with the
1074 * smallest possible size with __GFP_NOFAIL semantics.
1075 */
1076 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1077 if (unlikely(tbl == NULL)) {
1078 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1079 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1080 }
1081
1082 atomic_set(&ht->nelems, 0);
1083
1084 RCU_INIT_POINTER(ht->tbl, tbl);
1085
1086 INIT_WORK(&ht->run_work, rht_deferred_worker);
1087
1088 return 0;
1089}
1090EXPORT_SYMBOL_GPL(rhashtable_init_noprof);
1091
1092/**
1093 * rhltable_init - initialize a new hash list table
1094 * @hlt: hash list table to be initialized
1095 * @params: configuration parameters
1096 *
1097 * Initializes a new hash list table.
1098 *
1099 * See documentation for rhashtable_init.
1100 */
1101int rhltable_init_noprof(struct rhltable *hlt, const struct rhashtable_params *params)
1102{
1103 int err;
1104
1105 err = rhashtable_init_noprof(&hlt->ht, params);
1106 hlt->ht.rhlist = true;
1107 return err;
1108}
1109EXPORT_SYMBOL_GPL(rhltable_init_noprof);
1110
1111static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1112 void (*free_fn)(void *ptr, void *arg),
1113 void *arg)
1114{
1115 struct rhlist_head *list;
1116
1117 if (!ht->rhlist) {
1118 free_fn(rht_obj(ht, obj), arg);
1119 return;
1120 }
1121
1122 list = container_of(obj, struct rhlist_head, rhead);
1123 do {
1124 obj = &list->rhead;
1125 list = rht_dereference(list->next, ht);
1126 free_fn(rht_obj(ht, obj), arg);
1127 } while (list);
1128}
1129
1130/**
1131 * rhashtable_free_and_destroy - free elements and destroy hash table
1132 * @ht: the hash table to destroy
1133 * @free_fn: callback to release resources of element
1134 * @arg: pointer passed to free_fn
1135 *
1136 * Stops an eventual async resize. If defined, invokes free_fn for each
1137 * element to releasal resources. Please note that RCU protected
1138 * readers may still be accessing the elements. Releasing of resources
1139 * must occur in a compatible manner. Then frees the bucket array.
1140 *
1141 * This function will eventually sleep to wait for an async resize
1142 * to complete. The caller is responsible that no further write operations
1143 * occurs in parallel.
1144 */
1145void rhashtable_free_and_destroy(struct rhashtable *ht,
1146 void (*free_fn)(void *ptr, void *arg),
1147 void *arg)
1148{
1149 struct bucket_table *tbl, *next_tbl;
1150 unsigned int i;
1151
1152 cancel_work_sync(&ht->run_work);
1153
1154 mutex_lock(&ht->mutex);
1155 tbl = rht_dereference(ht->tbl, ht);
1156restart:
1157 if (free_fn) {
1158 for (i = 0; i < tbl->size; i++) {
1159 struct rhash_head *pos, *next;
1160
1161 cond_resched();
1162 for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1163 next = !rht_is_a_nulls(pos) ?
1164 rht_dereference(pos->next, ht) : NULL;
1165 !rht_is_a_nulls(pos);
1166 pos = next,
1167 next = !rht_is_a_nulls(pos) ?
1168 rht_dereference(pos->next, ht) : NULL)
1169 rhashtable_free_one(ht, pos, free_fn, arg);
1170 }
1171 }
1172
1173 next_tbl = rht_dereference(tbl->future_tbl, ht);
1174 bucket_table_free(tbl);
1175 if (next_tbl) {
1176 tbl = next_tbl;
1177 goto restart;
1178 }
1179 mutex_unlock(&ht->mutex);
1180}
1181EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1182
1183void rhashtable_destroy(struct rhashtable *ht)
1184{
1185 return rhashtable_free_and_destroy(ht, NULL, NULL);
1186}
1187EXPORT_SYMBOL_GPL(rhashtable_destroy);
1188
1189struct rhash_lock_head __rcu **__rht_bucket_nested(
1190 const struct bucket_table *tbl, unsigned int hash)
1191{
1192 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1193 unsigned int index = hash & ((1 << tbl->nest) - 1);
1194 unsigned int size = tbl->size >> tbl->nest;
1195 unsigned int subhash = hash;
1196 union nested_table *ntbl;
1197
1198 ntbl = nested_table_top(tbl);
1199 ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1200 subhash >>= tbl->nest;
1201
1202 while (ntbl && size > (1 << shift)) {
1203 index = subhash & ((1 << shift) - 1);
1204 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1205 tbl, hash);
1206 size >>= shift;
1207 subhash >>= shift;
1208 }
1209
1210 if (!ntbl)
1211 return NULL;
1212
1213 return &ntbl[subhash].bucket;
1214
1215}
1216EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1217
1218struct rhash_lock_head __rcu **rht_bucket_nested(
1219 const struct bucket_table *tbl, unsigned int hash)
1220{
1221 static struct rhash_lock_head __rcu *rhnull;
1222
1223 if (!rhnull)
1224 INIT_RHT_NULLS_HEAD(rhnull);
1225 return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1226}
1227EXPORT_SYMBOL_GPL(rht_bucket_nested);
1228
1229struct rhash_lock_head __rcu **rht_bucket_nested_insert(
1230 struct rhashtable *ht, struct bucket_table *tbl, unsigned int hash)
1231{
1232 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1233 unsigned int index = hash & ((1 << tbl->nest) - 1);
1234 unsigned int size = tbl->size >> tbl->nest;
1235 union nested_table *ntbl;
1236
1237 ntbl = nested_table_top(tbl);
1238 hash >>= tbl->nest;
1239 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1240 size <= (1 << shift));
1241
1242 while (ntbl && size > (1 << shift)) {
1243 index = hash & ((1 << shift) - 1);
1244 size >>= shift;
1245 hash >>= shift;
1246 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1247 size <= (1 << shift));
1248 }
1249
1250 if (!ntbl)
1251 return NULL;
1252
1253 return &ntbl[hash].bucket;
1254
1255}
1256EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);