projects/linux·blob master · refs

1409 lines · 35.5 KiB

1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Resizable, Scalable, Concurrent Hash Table
4 *
5 * Copyright (c) 2015 Herbert Xu <[email protected]>
6 * Copyright (c) 2014-2015 Thomas Graf <[email protected]>
7 * Copyright (c) 2008-2014 Patrick McHardy <[email protected]>
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/irq_work.h>
18#include <linux/log2.h>
19#include <linux/sched.h>
20#include <linux/rculist.h>
21#include <linux/slab.h>
22#include <linux/vmalloc.h>
23#include <linux/mm.h>
24#include <linux/jhash.h>
25#include <linux/random.h>
26#include <linux/rhashtable.h>
27#include <linux/err.h>
28#include <linux/export.h>
29#include <linux/workqueue.h>
30
31#define HASH_DEFAULT_SIZE 64UL
32#define HASH_MIN_SIZE 4U
33
34union nested_table {
35 union nested_table __rcu *table;
36 struct rhash_lock_head __rcu *bucket;
37};
38
39static u32 head_hashfn(struct rhashtable *ht,
40 const struct bucket_table *tbl,
41 const struct rhash_head *he)
42{
43 return rht_head_hashfn(ht, tbl, he, ht->p);
44}
45
46#ifdef CONFIG_PROVE_LOCKING
47#define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
48
49int lockdep_rht_mutex_is_held(struct rhashtable *ht)
50{
51 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
52}
53EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
54
55int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
56{
57 if (!debug_locks)
58 return 1;
59 if (unlikely(tbl->nest))
60 return 1;
61 return bit_spin_is_locked(0, (unsigned long *)&tbl->buckets[hash]);
62}
63EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
64#else
65#define ASSERT_RHT_MUTEX(HT)
66#endif
67
68static inline union nested_table *nested_table_top(
69 const struct bucket_table *tbl)
70{
71 /* The top-level bucket entry does not need RCU protection
72 * because it's set at the same time as tbl->nest.
73 */
74 return (void *)rcu_dereference_protected(tbl->buckets[0], 1);
75}
76
77static void nested_table_free(union nested_table *ntbl, unsigned int size)
78{
79 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
80 const unsigned int len = 1 << shift;
81 unsigned int i;
82
83 ntbl = rcu_dereference_protected(ntbl->table, 1);
84 if (!ntbl)
85 return;
86
87 if (size > len) {
88 size >>= shift;
89 for (i = 0; i < len; i++)
90 nested_table_free(ntbl + i, size);
91 }
92
93 kfree(ntbl);
94}
95
96static void nested_bucket_table_free(const struct bucket_table *tbl)
97{
98 unsigned int size = tbl->size >> tbl->nest;
99 unsigned int len = 1 << tbl->nest;
100 union nested_table *ntbl;
101 unsigned int i;
102
103 ntbl = nested_table_top(tbl);
104
105 for (i = 0; i < len; i++)
106 nested_table_free(ntbl + i, size);
107
108 kfree(ntbl);
109}
110
111static void bucket_table_free(const struct bucket_table *tbl)
112{
113 if (tbl->nest)
114 nested_bucket_table_free(tbl);
115
116 kvfree(tbl);
117}
118
119static void bucket_table_free_atomic(const struct bucket_table *tbl)
120{
121 if (tbl->nest)
122 nested_bucket_table_free(tbl);
123
124 kvfree_atomic(tbl);
125}
126
127static void bucket_table_free_rcu(struct rcu_head *head)
128{
129 bucket_table_free(container_of(head, struct bucket_table, rcu));
130}
131
132static union nested_table *nested_table_alloc(struct rhashtable *ht,
133 union nested_table __rcu **prev,
134 bool leaf)
135{
136 union nested_table *ntbl;
137 int i;
138
139 ntbl = rcu_dereference(*prev);
140 if (ntbl)
141 return ntbl;
142
143 ntbl = alloc_hooks_tag(ht->alloc_tag,
144 kmalloc_noprof(PAGE_SIZE, GFP_ATOMIC|__GFP_ZERO));
145
146 if (ntbl && leaf) {
147 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
148 INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
149 }
150
151 if (cmpxchg((union nested_table **)prev, NULL, ntbl) == NULL)
152 return ntbl;
153 /* Raced with another thread. */
154 kfree(ntbl);
155 return rcu_dereference(*prev);
156}
157
158static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
159 size_t nbuckets,
160 gfp_t gfp)
161{
162 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
163 struct bucket_table *tbl;
164 size_t size;
165
166 if (nbuckets < (1 << (shift + 1)))
167 return NULL;
168
169 size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
170
171 tbl = alloc_hooks_tag(ht->alloc_tag,
172 kmalloc_noprof(size, gfp|__GFP_ZERO));
173 if (!tbl)
174 return NULL;
175
176 if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
177 false)) {
178 kfree(tbl);
179 return NULL;
180 }
181
182 tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
183
184 return tbl;
185}
186
187static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
188 size_t nbuckets,
189 gfp_t gfp)
190{
191 struct bucket_table *tbl = NULL;
192 size_t size;
193 int i;
194 static struct lock_class_key __key;
195
196 tbl = alloc_hooks_tag(ht->alloc_tag,
197 kvmalloc_node_align_noprof(struct_size(tbl, buckets, nbuckets),
198 1, gfp|__GFP_ZERO, NUMA_NO_NODE));
199
200 size = nbuckets;
201
202 if (tbl == NULL && !gfpflags_allow_blocking(gfp)) {
203 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
204 nbuckets = 0;
205 }
206
207 if (tbl == NULL)
208 return NULL;
209
210 lockdep_init_map(&tbl->dep_map, "rhashtable_bucket", &__key, 0);
211
212 tbl->size = size;
213
214 rcu_head_init(&tbl->rcu);
215 INIT_LIST_HEAD(&tbl->walkers);
216
217 tbl->hash_rnd = get_random_u32();
218
219 for (i = 0; i < nbuckets; i++)
220 INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
221
222 return tbl;
223}
224
225static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
226 struct bucket_table *tbl)
227{
228 struct bucket_table *new_tbl;
229
230 do {
231 new_tbl = tbl;
232 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
233 } while (tbl);
234
235 return new_tbl;
236}
237
238static int rhashtable_rehash_one(struct rhashtable *ht,
239 struct rhash_lock_head __rcu **bkt,
240 unsigned int old_hash)
241{
242 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
243 struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
244 int err = -EAGAIN;
245 struct rhash_head *head, *next, *entry;
246 struct rhash_head __rcu **pprev = NULL;
247 unsigned int new_hash;
248 unsigned long flags;
249
250 if (new_tbl->nest)
251 goto out;
252
253 err = -ENOENT;
254
255 rht_for_each_from(entry, rht_ptr(bkt, old_tbl, old_hash),
256 old_tbl, old_hash) {
257 err = 0;
258 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
259
260 if (rht_is_a_nulls(next))
261 break;
262
263 pprev = &entry->next;
264 }
265
266 if (err)
267 goto out;
268
269 new_hash = head_hashfn(ht, new_tbl, entry);
270
271 flags = rht_lock_nested(new_tbl, &new_tbl->buckets[new_hash],
272 SINGLE_DEPTH_NESTING);
273
274 head = rht_ptr(new_tbl->buckets + new_hash, new_tbl, new_hash);
275
276 RCU_INIT_POINTER(entry->next, head);
277
278 rht_assign_unlock(new_tbl, &new_tbl->buckets[new_hash], entry, flags);
279
280 if (pprev)
281 rcu_assign_pointer(*pprev, next);
282 else
283 /* Need to preserved the bit lock. */
284 rht_assign_locked(bkt, next);
285
286out:
287 return err;
288}
289
290static int rhashtable_rehash_chain(struct rhashtable *ht,
291 unsigned int old_hash)
292{
293 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
294 struct rhash_lock_head __rcu **bkt = rht_bucket_var(old_tbl, old_hash);
295 unsigned long flags;
296 int err;
297
298 if (!bkt)
299 return 0;
300 flags = rht_lock(old_tbl, bkt);
301
302 while (!(err = rhashtable_rehash_one(ht, bkt, old_hash)))
303 ;
304
305 if (err == -ENOENT)
306 err = 0;
307 rht_unlock(old_tbl, bkt, flags);
308
309 return err;
310}
311
312static int rhashtable_rehash_attach(struct rhashtable *ht,
313 struct bucket_table *old_tbl,
314 struct bucket_table *new_tbl)
315{
316 /* Make insertions go into the new, empty table right away. Deletions
317 * and lookups will be attempted in both tables until we synchronize.
318 * As cmpxchg() provides strong barriers, we do not need
319 * rcu_assign_pointer().
320 */
321
322 if (cmpxchg((struct bucket_table **)&old_tbl->future_tbl, NULL,
323 new_tbl) != NULL)
324 return -EEXIST;
325
326 return 0;
327}
328
329static int rhashtable_rehash_table(struct rhashtable *ht)
330{
331 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
332 struct bucket_table *new_tbl;
333 struct rhashtable_walker *walker;
334 unsigned int old_hash;
335 int err;
336
337 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
338 if (!new_tbl)
339 return 0;
340
341 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
342 err = rhashtable_rehash_chain(ht, old_hash);
343 if (err)
344 return err;
345 cond_resched();
346 }
347
348 /* Publish the new table pointer. */
349 rcu_assign_pointer(ht->tbl, new_tbl);
350
351 spin_lock(&ht->lock);
352 list_for_each_entry(walker, &old_tbl->walkers, list)
353 walker->tbl = NULL;
354
355 /* Wait for readers. All new readers will see the new
356 * table, and thus no references to the old table will
357 * remain.
358 * We do this inside the locked region so that
359 * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
360 * to check if it should not re-link the table.
361 */
362 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
363 spin_unlock(&ht->lock);
364
365 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
366}
367
368static int rhashtable_rehash_alloc(struct rhashtable *ht,
369 struct bucket_table *old_tbl,
370 unsigned int size)
371 __must_hold(&ht->mutex)
372{
373 struct bucket_table *new_tbl;
374 int err;
375
376 ASSERT_RHT_MUTEX(ht);
377
378 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
379 if (new_tbl == NULL)
380 return -ENOMEM;
381
382 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
383 if (err)
384 bucket_table_free(new_tbl);
385
386 return err;
387}
388
389/**
390 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
391 * @ht: the hash table to shrink
392 *
393 * This function shrinks the hash table to fit, i.e., the smallest
394 * size would not cause it to expand right away automatically.
395 *
396 * The caller must ensure that no concurrent resizing occurs by holding
397 * ht->mutex.
398 *
399 * The caller must ensure that no concurrent table mutations take place.
400 * It is however valid to have concurrent lookups if they are RCU protected.
401 *
402 * It is valid to have concurrent insertions and deletions protected by per
403 * bucket locks or concurrent RCU protected lookups and traversals.
404 */
405static int rhashtable_shrink(struct rhashtable *ht)
406 __must_hold(&ht->mutex)
407{
408 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
409 unsigned int nelems = atomic_read(&ht->nelems);
410 unsigned int size = 0;
411
412 if (nelems)
413 size = roundup_pow_of_two(nelems * 3 / 2);
414 if (size < ht->p.min_size)
415 size = ht->p.min_size;
416
417 if (old_tbl->size <= size)
418 return 0;
419
420 if (rht_dereference(old_tbl->future_tbl, ht))
421 return -EEXIST;
422
423 return rhashtable_rehash_alloc(ht, old_tbl, size);
424}
425
426static void rht_deferred_worker(struct work_struct *work)
427{
428 struct rhashtable *ht;
429 struct bucket_table *tbl;
430 int err = 0;
431
432 ht = container_of(work, struct rhashtable, run_work);
433 mutex_lock(&ht->mutex);
434
435 tbl = rht_dereference(ht->tbl, ht);
436 tbl = rhashtable_last_table(ht, tbl);
437
438 if (rht_grow_above_75(ht, tbl))
439 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
440 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
441 err = rhashtable_shrink(ht);
442 else if (tbl->nest)
443 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
444
445 if (!err || err == -EEXIST) {
446 int nerr;
447
448 nerr = rhashtable_rehash_table(ht);
449 err = err ?: nerr;
450 }
451
452 mutex_unlock(&ht->mutex);
453
454 /*
455 * Re-arm via @run_work, not @run_irq_work.
456 * rhashtable_free_and_destroy() drains async work as irq_work_sync()
457 * followed by cancel_work_sync(). If this site queued irq_work while
458 * cancel_work_sync() was waiting for us, irq_work_sync() would already
459 * have returned and the stale irq_work could fire post-teardown.
460 * cancel_work_sync() natively handles self-requeue on @run_work.
461 */
462 if (err)
463 schedule_work(&ht->run_work);
464}
465
466/*
467 * Insert-path callers can run under a raw spinlock (e.g. an insecure_elasticity
468 * user). Calling schedule_work() under that lock records caller_lock ->
469 * pool->lock -> pi_lock -> rq->__lock, closing a locking cycle if any of
470 * these is acquired in the reverse direction elsewhere. Bounce through
471 * irq_work so the schedule_work() runs with the caller's lock no longer held.
472 */
473static void rht_deferred_irq_work(struct irq_work *irq_work)
474{
475 struct rhashtable *ht = container_of(irq_work, struct rhashtable,
476 run_irq_work);
477
478 schedule_work(&ht->run_work);
479}
480
481static int rhashtable_insert_rehash(struct rhashtable *ht,
482 struct bucket_table *tbl)
483{
484 struct bucket_table *old_tbl;
485 struct bucket_table *new_tbl;
486 unsigned int size;
487 int err;
488
489 old_tbl = rht_dereference_rcu(ht->tbl, ht);
490
491 size = tbl->size;
492
493 err = -EBUSY;
494
495 if (rht_grow_above_75(ht, tbl))
496 size *= 2;
497 /* Do not schedule more than one rehash */
498 else if (old_tbl != tbl)
499 goto fail;
500
501 err = -ENOMEM;
502
503 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
504 if (new_tbl == NULL)
505 goto fail;
506
507 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
508 if (err) {
509 bucket_table_free_atomic(new_tbl);
510 if (err == -EEXIST)
511 err = 0;
512 } else
513 irq_work_queue(&ht->run_irq_work);
514
515 return err;
516
517fail:
518 /* Do not fail the insert if someone else did a rehash. */
519 if (likely(rcu_access_pointer(tbl->future_tbl)))
520 return 0;
521
522 /* Schedule async rehash to retry allocation in process context. */
523 if (err == -ENOMEM)
524 irq_work_queue(&ht->run_irq_work);
525
526 return err;
527}
528
529static void *rhashtable_lookup_one(struct rhashtable *ht,
530 struct rhash_lock_head __rcu **bkt,
531 struct bucket_table *tbl, unsigned int hash,
532 const void *key, struct rhash_head *obj)
533{
534 struct rhashtable_compare_arg arg = {
535 .ht = ht,
536 .key = key,
537 };
538 struct rhash_head __rcu **pprev = NULL;
539 struct rhash_head *head;
540 int elasticity;
541
542 elasticity = RHT_ELASTICITY;
543 rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
544 struct rhlist_head *list;
545 struct rhlist_head *plist;
546
547 elasticity--;
548 if (!key ||
549 (ht->p.obj_cmpfn ?
550 ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
551 rhashtable_compare(&arg, rht_obj(ht, head)))) {
552 pprev = &head->next;
553 continue;
554 }
555
556 if (!ht->rhlist)
557 return rht_obj(ht, head);
558
559 list = container_of(obj, struct rhlist_head, rhead);
560 plist = container_of(head, struct rhlist_head, rhead);
561
562 RCU_INIT_POINTER(list->next, plist);
563 head = rht_dereference_bucket(head->next, tbl, hash);
564 RCU_INIT_POINTER(list->rhead.next, head);
565 if (pprev)
566 rcu_assign_pointer(*pprev, obj);
567 else
568 /* Need to preserve the bit lock */
569 rht_assign_locked(bkt, obj);
570
571 return NULL;
572 }
573
574 if (elasticity <= 0 && !ht->p.insecure_elasticity)
575 return ERR_PTR(-EAGAIN);
576
577 return ERR_PTR(-ENOENT);
578}
579
580static struct bucket_table *rhashtable_insert_one(
581 struct rhashtable *ht, struct rhash_lock_head __rcu **bkt,
582 struct bucket_table *tbl, unsigned int hash, struct rhash_head *obj,
583 void *data)
584{
585 struct bucket_table *new_tbl;
586 struct rhash_head *head;
587
588 if (!IS_ERR_OR_NULL(data))
589 return ERR_PTR(-EEXIST);
590
591 if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
592 return ERR_CAST(data);
593
594 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
595 if (new_tbl)
596 return new_tbl;
597
598 if (PTR_ERR(data) != -ENOENT)
599 return ERR_CAST(data);
600
601 if (unlikely(rht_grow_above_max(ht, tbl)))
602 return ERR_PTR(-E2BIG);
603
604 if (unlikely(rht_grow_above_100(ht, tbl)) &&
605 !ht->p.insecure_elasticity)
606 return ERR_PTR(-EAGAIN);
607
608 head = rht_ptr(bkt, tbl, hash);
609
610 RCU_INIT_POINTER(obj->next, head);
611 if (ht->rhlist) {
612 struct rhlist_head *list;
613
614 list = container_of(obj, struct rhlist_head, rhead);
615 RCU_INIT_POINTER(list->next, NULL);
616 }
617
618 /* bkt is always the head of the list, so it holds
619 * the lock, which we need to preserve
620 */
621 rht_assign_locked(bkt, obj);
622
623 return NULL;
624}
625
626static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
627 struct rhash_head *obj)
628{
629 struct bucket_table *new_tbl;
630 struct bucket_table *tbl;
631 struct rhash_lock_head __rcu **bkt;
632 unsigned long flags;
633 unsigned int hash;
634 void *data;
635
636 new_tbl = rcu_dereference(ht->tbl);
637
638 do {
639 tbl = new_tbl;
640 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
641 if (rcu_access_pointer(tbl->future_tbl))
642 /* Failure is OK */
643 bkt = rht_bucket_var(tbl, hash);
644 else
645 bkt = rht_bucket_insert(ht, tbl, hash);
646 if (bkt == NULL) {
647 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
648 data = ERR_PTR(-EAGAIN);
649 } else {
650 bool inserted;
651
652 flags = rht_lock(tbl, bkt);
653 data = rhashtable_lookup_one(ht, bkt, tbl,
654 hash, key, obj);
655 new_tbl = rhashtable_insert_one(ht, bkt, tbl,
656 hash, obj, data);
657 inserted = data && !new_tbl;
658 if (inserted)
659 atomic_inc(&ht->nelems);
660 if (PTR_ERR(new_tbl) != -EEXIST)
661 data = ERR_CAST(new_tbl);
662
663 rht_unlock(tbl, bkt, flags);
664
665 if (inserted && rht_grow_above_75(ht, tbl))
666 irq_work_queue(&ht->run_irq_work);
667 }
668 } while (!IS_ERR_OR_NULL(new_tbl));
669
670 if (PTR_ERR(data) == -EAGAIN)
671 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
672 -EAGAIN);
673
674 return data;
675}
676
677void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
678 struct rhash_head *obj)
679{
680 void *data;
681
682 do {
683 rcu_read_lock();
684 data = rhashtable_try_insert(ht, key, obj);
685 rcu_read_unlock();
686 } while (PTR_ERR(data) == -EAGAIN);
687
688 return data;
689}
690EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
691
692/* Scan one element forward from prev_key's position in @tbl.
693 * Returns first rhash_head whose bucket > prev_key's bucket, or the
694 * element immediately after prev_key inside prev_key's bucket.
695 * Returns the first element if prev_key is NULL, NULL when @tbl is
696 * exhausted, or ERR_PTR(-ENOENT) if prev_key is not found in @tbl.
697 */
698static struct rhash_head *__rhashtable_next_in_table(
699 struct rhashtable *ht, struct bucket_table *tbl,
700 const void *prev_key)
701{
702 struct rhashtable_compare_arg arg = { .ht = ht, .key = prev_key };
703 const struct rhashtable_params params = ht->p;
704 struct rhash_head *he;
705 unsigned int b = 0;
706 bool found = false;
707
708 if (prev_key) {
709 b = rht_key_hashfn(ht, tbl, prev_key, params);
710 rht_for_each_rcu(he, tbl, b) {
711 bool match = params.obj_cmpfn
712 ? !params.obj_cmpfn(&arg, rht_obj(ht, he))
713 : !rhashtable_compare(&arg, rht_obj(ht, he));
714 if (found) {
715 if (match)
716 continue;
717 return he;
718 }
719 if (match)
720 found = true;
721 }
722 if (!found)
723 return ERR_PTR(-ENOENT);
724 b++;
725 }
726
727 for (; b < tbl->size; b++)
728 rht_for_each_rcu(he, tbl, b)
729 return he;
730 return NULL;
731}
732
733/**
734 * rhashtable_next_key - return next element after a given key
735 * @ht: hash table
736 * @prev_key: pointer to previous key, or NULL for the first element
737 *
738 * WARNING: this walk is highly unstable. Unlike rhashtable_walk_*(),
739 * it cannot detect a concurrent resize or rehash, so a full iteration
740 * is NOT guaranteed to terminate under adversarial or sustained
741 * rehashing. Callers MUST tolerate skipped and duplicated elements and
742 * SHOULD bound their loop externally.
743 *
744 * Returns the next element in best-effort iteration order, walking the
745 * @tbl chain (including any future_tbl in flight). Caller must hold RCU.
746 *
747 * Pass @prev_key == NULL to obtain the first element. To iterate, set
748 * @prev_key to the key of the previously returned element on each call,
749 * and stop when NULL is returned.
750 *
751 * Best-effort semantics:
752 * - Across the tbl->future_tbl chain, an element being migrated may
753 * transiently appear in both tables and be observed twice.
754 * - Concurrent inserts may or may not be observed.
755 * - Termination of a full iteration loop is NOT guaranteed under
756 * adversarial continuous rehash; callers MUST tolerate skips and
757 * repeats and SHOULD bound their loop externally.
758 * - Behavior on tables that contain duplicate keys is undefined:
759 * duplicates may be skipped, repeated, or trap the walk in a
760 * cycle. Callers requiring duplicate-key iteration must use
761 * rhashtable_walk_*() instead.
762 * - rhltable instances are not supported and return
763 * ERR_PTR(-EOPNOTSUPP).
764 * - If prev_key was concurrently deleted and is not present in any
765 * in-flight table, returns ERR_PTR(-ENOENT).
766 *
767 * Returns entry of the next element, or NULL when iteration is exhausted,
768 * or ERR_PTR(-ENOENT) if prev_key is not found, or
769 * ERR_PTR(-EOPNOTSUPP) if @ht is an rhltable.
770 */
771void *rhashtable_next_key(struct rhashtable *ht, const void *prev_key)
772{
773 struct bucket_table *tbl;
774 struct rhash_head *he;
775
776 if (unlikely(ht->rhlist))
777 return ERR_PTR(-EOPNOTSUPP);
778
779 tbl = rht_dereference_rcu(ht->tbl, ht);
780 do {
781 he = __rhashtable_next_in_table(ht, tbl, prev_key);
782 if (!IS_ERR_OR_NULL(he))
783 return rht_obj(ht, he);
784 if (!he)
785 prev_key = NULL;
786 /* See any new future_tbl attached during a rehash. */
787 smp_rmb();
788 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
789 } while (tbl);
790 return he; /* NULL or -ENOENT */
791}
792EXPORT_SYMBOL_GPL(rhashtable_next_key);
793
794/**
795 * rhashtable_walk_enter - Initialise an iterator
796 * @ht: Table to walk over
797 * @iter: Hash table Iterator
798 *
799 * This function prepares a hash table walk.
800 *
801 * Note that if you restart a walk after rhashtable_walk_stop you
802 * may see the same object twice. Also, you may miss objects if
803 * there are removals in between rhashtable_walk_stop and the next
804 * call to rhashtable_walk_start.
805 *
806 * For a completely stable walk you should construct your own data
807 * structure outside the hash table.
808 *
809 * This function may be called from any process context, including
810 * non-preemptible context, but cannot be called from softirq or
811 * hardirq context.
812 *
813 * You must call rhashtable_walk_exit after this function returns.
814 */
815void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
816{
817 iter->ht = ht;
818 iter->p = NULL;
819 iter->slot = 0;
820 iter->skip = 0;
821 iter->end_of_table = 0;
822
823 spin_lock(&ht->lock);
824 iter->walker.tbl =
825 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
826 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
827 spin_unlock(&ht->lock);
828}
829EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
830
831/**
832 * rhashtable_walk_exit - Free an iterator
833 * @iter: Hash table Iterator
834 *
835 * This function frees resources allocated by rhashtable_walk_enter.
836 */
837void rhashtable_walk_exit(struct rhashtable_iter *iter)
838{
839 spin_lock(&iter->ht->lock);
840 if (iter->walker.tbl)
841 list_del(&iter->walker.list);
842 spin_unlock(&iter->ht->lock);
843}
844EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
845
846/**
847 * rhashtable_walk_start_check - Start a hash table walk
848 * @iter: Hash table iterator
849 *
850 * Start a hash table walk at the current iterator position. Note that we take
851 * the RCU lock in all cases including when we return an error. So you must
852 * always call rhashtable_walk_stop to clean up.
853 *
854 * Returns zero if successful.
855 *
856 * Returns -EAGAIN if resize event occurred. Note that the iterator
857 * will rewind back to the beginning and you may use it immediately
858 * by calling rhashtable_walk_next.
859 *
860 * rhashtable_walk_start is defined as an inline variant that returns
861 * void. This is preferred in cases where the caller would ignore
862 * resize events and always continue.
863 */
864int rhashtable_walk_start_check(struct rhashtable_iter *iter)
865 __acquires_shared(RCU)
866{
867 struct rhashtable *ht = iter->ht;
868 bool rhlist = ht->rhlist;
869
870 rcu_read_lock();
871
872 spin_lock(&ht->lock);
873 if (iter->walker.tbl)
874 list_del(&iter->walker.list);
875 spin_unlock(&ht->lock);
876
877 if (iter->end_of_table)
878 return 0;
879 if (!iter->walker.tbl) {
880 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
881 iter->slot = 0;
882 iter->skip = 0;
883 iter->p = NULL;
884 return -EAGAIN;
885 }
886
887 if (iter->p && !rhlist) {
888 /*
889 * We need to validate that 'p' is still in the table, and
890 * if so, update 'skip'
891 */
892 struct rhash_head *p;
893 int skip = 0;
894 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
895 skip++;
896 if (p == iter->p) {
897 iter->skip = skip;
898 goto found;
899 }
900 }
901 iter->p = NULL;
902 } else if (iter->p && rhlist) {
903 /* Need to validate that 'list' is still in the table, and
904 * if so, update 'skip' and 'p'.
905 */
906 struct rhash_head *p;
907 struct rhlist_head *list;
908 int skip = 0;
909 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
910 for (list = container_of(p, struct rhlist_head, rhead);
911 list;
912 list = rcu_dereference(list->next)) {
913 skip++;
914 if (list == iter->list) {
915 iter->p = p;
916 iter->skip = skip;
917 goto found;
918 }
919 }
920 }
921 iter->p = NULL;
922 }
923found:
924 return 0;
925}
926EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
927
928/**
929 * __rhashtable_walk_find_next - Find the next element in a table (or the first
930 * one in case of a new walk).
931 *
932 * @iter: Hash table iterator
933 *
934 * Returns the found object or NULL when the end of the table is reached.
935 *
936 * Returns -EAGAIN if resize event occurred.
937 */
938static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
939{
940 struct bucket_table *tbl = iter->walker.tbl;
941 struct rhlist_head *list = iter->list;
942 struct rhashtable *ht = iter->ht;
943 struct rhash_head *p = iter->p;
944 bool rhlist = ht->rhlist;
945
946 if (!tbl)
947 return NULL;
948
949 for (; iter->slot < tbl->size; iter->slot++) {
950 int skip = iter->skip;
951
952 rht_for_each_rcu(p, tbl, iter->slot) {
953 if (rhlist) {
954 list = container_of(p, struct rhlist_head,
955 rhead);
956 do {
957 if (!skip)
958 goto next;
959 skip--;
960 list = rcu_dereference(list->next);
961 } while (list);
962
963 continue;
964 }
965 if (!skip)
966 break;
967 skip--;
968 }
969
970next:
971 if (!rht_is_a_nulls(p)) {
972 iter->skip++;
973 iter->p = p;
974 iter->list = list;
975 return rht_obj(ht, rhlist ? &list->rhead : p);
976 }
977
978 iter->skip = 0;
979 }
980
981 iter->p = NULL;
982
983 /* Ensure we see any new tables. */
984 smp_rmb();
985
986 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
987 if (iter->walker.tbl) {
988 iter->slot = 0;
989 iter->skip = 0;
990 return ERR_PTR(-EAGAIN);
991 } else {
992 iter->end_of_table = true;
993 }
994
995 return NULL;
996}
997
998/**
999 * rhashtable_walk_next - Return the next object and advance the iterator
1000 * @iter: Hash table iterator
1001 *
1002 * Note that you must call rhashtable_walk_stop when you are finished
1003 * with the walk.
1004 *
1005 * Returns the next object or NULL when the end of the table is reached.
1006 *
1007 * Returns -EAGAIN if resize event occurred. Note that the iterator
1008 * will rewind back to the beginning and you may continue to use it.
1009 */
1010void *rhashtable_walk_next(struct rhashtable_iter *iter)
1011{
1012 struct rhlist_head *list = iter->list;
1013 struct rhashtable *ht = iter->ht;
1014 struct rhash_head *p = iter->p;
1015 bool rhlist = ht->rhlist;
1016
1017 if (p) {
1018 if (!rhlist || !(list = rcu_dereference(list->next))) {
1019 p = rcu_dereference(p->next);
1020 list = container_of(p, struct rhlist_head, rhead);
1021 }
1022 if (!rht_is_a_nulls(p)) {
1023 iter->skip++;
1024 iter->p = p;
1025 iter->list = list;
1026 return rht_obj(ht, rhlist ? &list->rhead : p);
1027 }
1028
1029 /* At the end of this slot, switch to next one and then find
1030 * next entry from that point.
1031 */
1032 iter->skip = 0;
1033 iter->slot++;
1034 }
1035
1036 return __rhashtable_walk_find_next(iter);
1037}
1038EXPORT_SYMBOL_GPL(rhashtable_walk_next);
1039
1040/**
1041 * rhashtable_walk_peek - Return the next object but don't advance the iterator
1042 * @iter: Hash table iterator
1043 *
1044 * Returns the next object or NULL when the end of the table is reached.
1045 *
1046 * Returns -EAGAIN if resize event occurred. Note that the iterator
1047 * will rewind back to the beginning and you may continue to use it.
1048 */
1049void *rhashtable_walk_peek(struct rhashtable_iter *iter)
1050{
1051 struct rhlist_head *list = iter->list;
1052 struct rhashtable *ht = iter->ht;
1053 struct rhash_head *p = iter->p;
1054
1055 if (p)
1056 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
1057
1058 /* No object found in current iter, find next one in the table. */
1059
1060 if (iter->skip) {
1061 /* A nonzero skip value points to the next entry in the table
1062 * beyond that last one that was found. Decrement skip so
1063 * we find the current value. __rhashtable_walk_find_next
1064 * will restore the original value of skip assuming that
1065 * the table hasn't changed.
1066 */
1067 iter->skip--;
1068 }
1069
1070 return __rhashtable_walk_find_next(iter);
1071}
1072EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
1073
1074/**
1075 * rhashtable_walk_stop - Finish a hash table walk
1076 * @iter: Hash table iterator
1077 *
1078 * Finish a hash table walk. Does not reset the iterator to the start of the
1079 * hash table.
1080 */
1081void rhashtable_walk_stop(struct rhashtable_iter *iter)
1082{
1083 struct rhashtable *ht;
1084 struct bucket_table *tbl = iter->walker.tbl;
1085
1086 if (!tbl)
1087 goto out;
1088
1089 ht = iter->ht;
1090
1091 spin_lock(&ht->lock);
1092 if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
1093 /* This bucket table is being freed, don't re-link it. */
1094 iter->walker.tbl = NULL;
1095 else
1096 list_add(&iter->walker.list, &tbl->walkers);
1097 spin_unlock(&ht->lock);
1098
1099out:
1100 rcu_read_unlock();
1101}
1102EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
1103
1104static size_t rounded_hashtable_size(const struct rhashtable_params *params)
1105{
1106 size_t retsize;
1107
1108 if (params->nelem_hint)
1109 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
1110 (unsigned long)params->min_size);
1111 else
1112 retsize = max(HASH_DEFAULT_SIZE,
1113 (unsigned long)params->min_size);
1114
1115 return retsize;
1116}
1117
1118static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
1119{
1120 return jhash2(key, length, seed);
1121}
1122
1123/**
1124 * rhashtable_init - initialize a new hash table
1125 * @ht: hash table to be initialized
1126 * @params: configuration parameters
1127 *
1128 * Initializes a new hash table based on the provided configuration
1129 * parameters. A table can be configured either with a variable or
1130 * fixed length key:
1131 *
1132 * Configuration Example 1: Fixed length keys
1133 * struct test_obj {
1134 * int key;
1135 * void * my_member;
1136 * struct rhash_head node;
1137 * };
1138 *
1139 * struct rhashtable_params params = {
1140 * .head_offset = offsetof(struct test_obj, node),
1141 * .key_offset = offsetof(struct test_obj, key),
1142 * .key_len = sizeof(int),
1143 * .hashfn = jhash,
1144 * };
1145 *
1146 * Configuration Example 2: Variable length keys
1147 * struct test_obj {
1148 * [...]
1149 * struct rhash_head node;
1150 * };
1151 *
1152 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1153 * {
1154 * struct test_obj *obj = data;
1155 *
1156 * return [... hash ...];
1157 * }
1158 *
1159 * struct rhashtable_params params = {
1160 * .head_offset = offsetof(struct test_obj, node),
1161 * .hashfn = jhash,
1162 * .obj_hashfn = my_hash_fn,
1163 * };
1164 */
1165int __rhashtable_init_noprof(struct rhashtable *ht,
1166 const struct rhashtable_params *params,
1167 struct lock_class_key *key)
1168{
1169 struct bucket_table *tbl;
1170 size_t size;
1171
1172 if ((!params->key_len && !params->obj_hashfn) ||
1173 (params->obj_hashfn && !params->obj_cmpfn))
1174 return -EINVAL;
1175
1176 memset(ht, 0, sizeof(*ht));
1177 mutex_init_with_key(&ht->mutex, key);
1178 spin_lock_init(&ht->lock);
1179 memcpy(&ht->p, params, sizeof(*params));
1180
1181 alloc_tag_record(ht->alloc_tag);
1182
1183 if (params->min_size)
1184 ht->p.min_size = roundup_pow_of_two(params->min_size);
1185
1186 /* Cap total entries at 2^31 to avoid nelems overflow. */
1187 ht->max_elems = 1u << 31;
1188
1189 if (params->max_size) {
1190 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1191 if (ht->p.max_size < ht->max_elems / 2)
1192 ht->max_elems = ht->p.max_size * 2;
1193 }
1194
1195 ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1196
1197 size = rounded_hashtable_size(&ht->p);
1198
1199 ht->key_len = ht->p.key_len;
1200 if (!params->hashfn) {
1201 ht->p.hashfn = jhash;
1202
1203 if (!(ht->key_len & (sizeof(u32) - 1))) {
1204 ht->key_len /= sizeof(u32);
1205 ht->p.hashfn = rhashtable_jhash2;
1206 }
1207 }
1208
1209 /*
1210 * This is api initialization and thus we need to guarantee the
1211 * initial rhashtable allocation. Upon failure, retry with the
1212 * smallest possible size with __GFP_NOFAIL semantics.
1213 */
1214 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1215 if (unlikely(tbl == NULL)) {
1216 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1217 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1218 }
1219
1220 atomic_set(&ht->nelems, 0);
1221
1222 RCU_INIT_POINTER(ht->tbl, tbl);
1223
1224 INIT_WORK(&ht->run_work, rht_deferred_worker);
1225 init_irq_work(&ht->run_irq_work, rht_deferred_irq_work);
1226
1227 return 0;
1228}
1229EXPORT_SYMBOL_GPL(__rhashtable_init_noprof);
1230
1231/**
1232 * rhltable_init - initialize a new hash list table
1233 * @hlt: hash list table to be initialized
1234 * @params: configuration parameters
1235 *
1236 * Initializes a new hash list table.
1237 *
1238 * See documentation for rhashtable_init.
1239 */
1240int __rhltable_init_noprof(struct rhltable *hlt,
1241 const struct rhashtable_params *params,
1242 struct lock_class_key *key)
1243{
1244 int err;
1245
1246 err = __rhashtable_init_noprof(&hlt->ht, params, key);
1247 hlt->ht.rhlist = true;
1248 return err;
1249}
1250EXPORT_SYMBOL_GPL(__rhltable_init_noprof);
1251
1252static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1253 void (*free_fn)(void *ptr, void *arg),
1254 void *arg)
1255{
1256 struct rhlist_head *list;
1257
1258 if (!ht->rhlist) {
1259 free_fn(rht_obj(ht, obj), arg);
1260 return;
1261 }
1262
1263 list = container_of(obj, struct rhlist_head, rhead);
1264 do {
1265 obj = &list->rhead;
1266 list = rcu_dereference_raw(list->next);
1267 free_fn(rht_obj(ht, obj), arg);
1268 } while (list);
1269}
1270
1271/**
1272 * rhashtable_free_and_destroy - free elements and destroy hash table
1273 * @ht: the hash table to destroy
1274 * @free_fn: callback to release resources of element
1275 * @arg: pointer passed to free_fn
1276 *
1277 * Stops an eventual async resize. If defined, invokes free_fn for each
1278 * element to releasal resources. Please note that RCU protected
1279 * readers may still be accessing the elements. Releasing of resources
1280 * must occur in a compatible manner. Then frees the bucket array.
1281 *
1282 * This function will eventually sleep to wait for an async resize
1283 * to complete. The caller is responsible that no further write operations
1284 * occurs in parallel.
1285 *
1286 * After cancel_work_sync() has returned, the deferred rehash worker is
1287 * quiesced and, per the contract above, no other concurrent access to the
1288 * rhashtable is possible. The tables are therefore owned exclusively by
1289 * this function and can be walked without ht->mutex held.
1290 */
1291void rhashtable_free_and_destroy(struct rhashtable *ht,
1292 void (*free_fn)(void *ptr, void *arg),
1293 void *arg)
1294{
1295 struct bucket_table *tbl, *next_tbl;
1296 unsigned int i;
1297
1298 irq_work_sync(&ht->run_irq_work);
1299 cancel_work_sync(&ht->run_work);
1300
1301 /*
1302 * Do NOT take ht->mutex here. The rehash worker establishes
1303 * ht->mutex -> fs_reclaim via GFP_KERNEL bucket allocation under
1304 * the mutex; callers on the reclaim path (e.g. simple_xattr_ht_free()
1305 * from evict() under the dcache shrinker for shmem/kernfs/pidfs
1306 * inodes) would otherwise close a circular dependency
1307 * fs_reclaim -> ht->mutex.
1308 */
1309 tbl = rcu_dereference_raw(ht->tbl);
1310restart:
1311 if (free_fn) {
1312 for (i = 0; i < tbl->size; i++) {
1313 struct rhash_head *pos, *next;
1314
1315 cond_resched();
1316 for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1317 next = !rht_is_a_nulls(pos) ?
1318 rcu_dereference_raw(pos->next) : NULL;
1319 !rht_is_a_nulls(pos);
1320 pos = next,
1321 next = !rht_is_a_nulls(pos) ?
1322 rcu_dereference_raw(pos->next) : NULL)
1323 rhashtable_free_one(ht, pos, free_fn, arg);
1324 }
1325 }
1326
1327 next_tbl = rcu_dereference_raw(tbl->future_tbl);
1328 bucket_table_free(tbl);
1329 if (next_tbl) {
1330 tbl = next_tbl;
1331 goto restart;
1332 }
1333}
1334EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1335
1336void rhashtable_destroy(struct rhashtable *ht)
1337{
1338 return rhashtable_free_and_destroy(ht, NULL, NULL);
1339}
1340EXPORT_SYMBOL_GPL(rhashtable_destroy);
1341
1342struct rhash_lock_head __rcu **__rht_bucket_nested(
1343 const struct bucket_table *tbl, unsigned int hash)
1344{
1345 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1346 unsigned int index = hash & ((1 << tbl->nest) - 1);
1347 unsigned int size = tbl->size >> tbl->nest;
1348 unsigned int subhash = hash;
1349 union nested_table *ntbl;
1350
1351 ntbl = nested_table_top(tbl);
1352 ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1353 subhash >>= tbl->nest;
1354
1355 while (ntbl && size > (1 << shift)) {
1356 index = subhash & ((1 << shift) - 1);
1357 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1358 tbl, hash);
1359 size >>= shift;
1360 subhash >>= shift;
1361 }
1362
1363 if (!ntbl)
1364 return NULL;
1365
1366 return &ntbl[subhash].bucket;
1367
1368}
1369EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1370
1371struct rhash_lock_head __rcu **rht_bucket_nested(
1372 const struct bucket_table *tbl, unsigned int hash)
1373{
1374 static struct rhash_lock_head __rcu *rhnull;
1375
1376 if (!rhnull)
1377 INIT_RHT_NULLS_HEAD(rhnull);
1378 return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1379}
1380EXPORT_SYMBOL_GPL(rht_bucket_nested);
1381
1382struct rhash_lock_head __rcu **rht_bucket_nested_insert(
1383 struct rhashtable *ht, struct bucket_table *tbl, unsigned int hash)
1384{
1385 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1386 unsigned int index = hash & ((1 << tbl->nest) - 1);
1387 unsigned int size = tbl->size >> tbl->nest;
1388 union nested_table *ntbl;
1389
1390 ntbl = nested_table_top(tbl);
1391 hash >>= tbl->nest;
1392 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1393 size <= (1 << shift));
1394
1395 while (ntbl && size > (1 << shift)) {
1396 index = hash & ((1 << shift) - 1);
1397 size >>= shift;
1398 hash >>= shift;
1399 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1400 size <= (1 << shift));
1401 }
1402
1403 if (!ntbl)
1404 return NULL;
1405
1406 return &ntbl[hash].bucket;
1407
1408}
1409EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);