1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
5 * Development of this code funded by Astaro AG (http://www.astaro.com/)
8 #include <linux/kernel.h>
9 #include <linux/init.h>
10 #include <linux/module.h>
11 #include <linux/list.h>
12 #include <linux/log2.h>
13 #include <linux/jhash.h>
14 #include <linux/netlink.h>
15 #include <linux/workqueue.h>
16 #include <linux/rhashtable.h>
17 #include <linux/netfilter.h>
18 #include <linux/netfilter/nf_tables.h>
19 #include <net/netfilter/nf_tables_core.h>
21 /* We target a hash table size of 4, element hint is 75% of final size */
22 #define NFT_RHASH_ELEMENT_HINT 3
26 struct delayed_work gc_work;
29 struct nft_rhash_elem {
30 struct rhash_head node;
31 struct nft_set_ext ext;
34 struct nft_rhash_cmp_arg {
35 const struct nft_set *set;
40 static inline u32 nft_rhash_key(const void *data, u32 len, u32 seed)
42 const struct nft_rhash_cmp_arg *arg = data;
44 return jhash(arg->key, len, seed);
47 static inline u32 nft_rhash_obj(const void *data, u32 len, u32 seed)
49 const struct nft_rhash_elem *he = data;
51 return jhash(nft_set_ext_key(&he->ext), len, seed);
54 static inline int nft_rhash_cmp(struct rhashtable_compare_arg *arg,
57 const struct nft_rhash_cmp_arg *x = arg->key;
58 const struct nft_rhash_elem *he = ptr;
60 if (memcmp(nft_set_ext_key(&he->ext), x->key, x->set->klen))
62 if (nft_set_elem_expired(&he->ext))
64 if (!nft_set_elem_active(&he->ext, x->genmask))
69 static const struct rhashtable_params nft_rhash_params = {
70 .head_offset = offsetof(struct nft_rhash_elem, node),
71 .hashfn = nft_rhash_key,
72 .obj_hashfn = nft_rhash_obj,
73 .obj_cmpfn = nft_rhash_cmp,
74 .automatic_shrinking = true,
77 static bool nft_rhash_lookup(const struct net *net, const struct nft_set *set,
78 const u32 *key, const struct nft_set_ext **ext)
80 struct nft_rhash *priv = nft_set_priv(set);
81 const struct nft_rhash_elem *he;
82 struct nft_rhash_cmp_arg arg = {
83 .genmask = nft_genmask_cur(net),
88 he = rhashtable_lookup(&priv->ht, &arg, nft_rhash_params);
95 static void *nft_rhash_get(const struct net *net, const struct nft_set *set,
96 const struct nft_set_elem *elem, unsigned int flags)
98 struct nft_rhash *priv = nft_set_priv(set);
99 struct nft_rhash_elem *he;
100 struct nft_rhash_cmp_arg arg = {
101 .genmask = nft_genmask_cur(net),
103 .key = elem->key.val.data,
106 he = rhashtable_lookup(&priv->ht, &arg, nft_rhash_params);
110 return ERR_PTR(-ENOENT);
113 static bool nft_rhash_update(struct nft_set *set, const u32 *key,
114 void *(*new)(struct nft_set *,
115 const struct nft_expr *,
116 struct nft_regs *regs),
117 const struct nft_expr *expr,
118 struct nft_regs *regs,
119 const struct nft_set_ext **ext)
121 struct nft_rhash *priv = nft_set_priv(set);
122 struct nft_rhash_elem *he, *prev;
123 struct nft_rhash_cmp_arg arg = {
124 .genmask = NFT_GENMASK_ANY,
129 he = rhashtable_lookup(&priv->ht, &arg, nft_rhash_params);
133 he = new(set, expr, regs);
137 prev = rhashtable_lookup_get_insert_key(&priv->ht, &arg, &he->node,
142 /* Another cpu may race to insert the element with the same key */
144 nft_set_elem_destroy(set, he, true);
153 nft_set_elem_destroy(set, he, true);
158 static int nft_rhash_insert(const struct net *net, const struct nft_set *set,
159 const struct nft_set_elem *elem,
160 struct nft_set_ext **ext)
162 struct nft_rhash *priv = nft_set_priv(set);
163 struct nft_rhash_elem *he = elem->priv;
164 struct nft_rhash_cmp_arg arg = {
165 .genmask = nft_genmask_next(net),
167 .key = elem->key.val.data,
169 struct nft_rhash_elem *prev;
171 prev = rhashtable_lookup_get_insert_key(&priv->ht, &arg, &he->node,
174 return PTR_ERR(prev);
182 static void nft_rhash_activate(const struct net *net, const struct nft_set *set,
183 const struct nft_set_elem *elem)
185 struct nft_rhash_elem *he = elem->priv;
187 nft_set_elem_change_active(net, set, &he->ext);
188 nft_set_elem_clear_busy(&he->ext);
191 static bool nft_rhash_flush(const struct net *net,
192 const struct nft_set *set, void *priv)
194 struct nft_rhash_elem *he = priv;
196 if (!nft_set_elem_mark_busy(&he->ext) ||
197 !nft_is_active(net, &he->ext)) {
198 nft_set_elem_change_active(net, set, &he->ext);
204 static void *nft_rhash_deactivate(const struct net *net,
205 const struct nft_set *set,
206 const struct nft_set_elem *elem)
208 struct nft_rhash *priv = nft_set_priv(set);
209 struct nft_rhash_elem *he;
210 struct nft_rhash_cmp_arg arg = {
211 .genmask = nft_genmask_next(net),
213 .key = elem->key.val.data,
217 he = rhashtable_lookup(&priv->ht, &arg, nft_rhash_params);
219 !nft_rhash_flush(net, set, he))
227 static void nft_rhash_remove(const struct net *net,
228 const struct nft_set *set,
229 const struct nft_set_elem *elem)
231 struct nft_rhash *priv = nft_set_priv(set);
232 struct nft_rhash_elem *he = elem->priv;
234 rhashtable_remove_fast(&priv->ht, &he->node, nft_rhash_params);
237 static bool nft_rhash_delete(const struct nft_set *set,
240 struct nft_rhash *priv = nft_set_priv(set);
241 struct nft_rhash_cmp_arg arg = {
242 .genmask = NFT_GENMASK_ANY,
246 struct nft_rhash_elem *he;
248 he = rhashtable_lookup(&priv->ht, &arg, nft_rhash_params);
252 return rhashtable_remove_fast(&priv->ht, &he->node, nft_rhash_params) == 0;
255 static void nft_rhash_walk(const struct nft_ctx *ctx, struct nft_set *set,
256 struct nft_set_iter *iter)
258 struct nft_rhash *priv = nft_set_priv(set);
259 struct nft_rhash_elem *he;
260 struct rhashtable_iter hti;
261 struct nft_set_elem elem;
263 rhashtable_walk_enter(&priv->ht, &hti);
264 rhashtable_walk_start(&hti);
266 while ((he = rhashtable_walk_next(&hti))) {
268 if (PTR_ERR(he) != -EAGAIN) {
269 iter->err = PTR_ERR(he);
276 if (iter->count < iter->skip)
278 if (nft_set_elem_expired(&he->ext))
280 if (!nft_set_elem_active(&he->ext, iter->genmask))
285 iter->err = iter->fn(ctx, set, iter, &elem);
292 rhashtable_walk_stop(&hti);
293 rhashtable_walk_exit(&hti);
296 static bool nft_rhash_expr_needs_gc_run(const struct nft_set *set,
297 struct nft_set_ext *ext)
299 struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
300 struct nft_expr *expr;
303 nft_setelem_expr_foreach(expr, elem_expr, size) {
305 expr->ops->gc(read_pnet(&set->net), expr))
312 static void nft_rhash_gc(struct work_struct *work)
315 struct nft_rhash_elem *he;
316 struct nft_rhash *priv;
317 struct nft_set_gc_batch *gcb = NULL;
318 struct rhashtable_iter hti;
320 priv = container_of(work, struct nft_rhash, gc_work.work);
321 set = nft_set_container_of(priv);
323 rhashtable_walk_enter(&priv->ht, &hti);
324 rhashtable_walk_start(&hti);
326 while ((he = rhashtable_walk_next(&hti))) {
328 if (PTR_ERR(he) != -EAGAIN)
333 if (nft_set_ext_exists(&he->ext, NFT_SET_EXT_EXPRESSIONS) &&
334 nft_rhash_expr_needs_gc_run(set, &he->ext))
337 if (!nft_set_elem_expired(&he->ext))
340 if (nft_set_elem_mark_busy(&he->ext))
343 gcb = nft_set_gc_batch_check(set, gcb, GFP_ATOMIC);
346 rhashtable_remove_fast(&priv->ht, &he->node, nft_rhash_params);
347 atomic_dec(&set->nelems);
348 nft_set_gc_batch_add(gcb, he);
350 rhashtable_walk_stop(&hti);
351 rhashtable_walk_exit(&hti);
353 he = nft_set_catchall_gc(set);
355 gcb = nft_set_gc_batch_check(set, gcb, GFP_ATOMIC);
357 nft_set_gc_batch_add(gcb, he);
359 nft_set_gc_batch_complete(gcb);
360 queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
361 nft_set_gc_interval(set));
364 static u64 nft_rhash_privsize(const struct nlattr * const nla[],
365 const struct nft_set_desc *desc)
367 return sizeof(struct nft_rhash);
370 static void nft_rhash_gc_init(const struct nft_set *set)
372 struct nft_rhash *priv = nft_set_priv(set);
374 queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
375 nft_set_gc_interval(set));
378 static int nft_rhash_init(const struct nft_set *set,
379 const struct nft_set_desc *desc,
380 const struct nlattr * const tb[])
382 struct nft_rhash *priv = nft_set_priv(set);
383 struct rhashtable_params params = nft_rhash_params;
386 params.nelem_hint = desc->size ?: NFT_RHASH_ELEMENT_HINT;
387 params.key_len = set->klen;
389 err = rhashtable_init(&priv->ht, ¶ms);
393 INIT_DEFERRABLE_WORK(&priv->gc_work, nft_rhash_gc);
394 if (set->flags & NFT_SET_TIMEOUT)
395 nft_rhash_gc_init(set);
400 static void nft_rhash_elem_destroy(void *ptr, void *arg)
402 nft_set_elem_destroy(arg, ptr, true);
405 static void nft_rhash_destroy(const struct nft_set *set)
407 struct nft_rhash *priv = nft_set_priv(set);
409 cancel_delayed_work_sync(&priv->gc_work);
411 rhashtable_free_and_destroy(&priv->ht, nft_rhash_elem_destroy,
415 /* Number of buckets is stored in u32, so cap our result to 1U<<31 */
416 #define NFT_MAX_BUCKETS (1U << 31)
418 static u32 nft_hash_buckets(u32 size)
420 u64 val = div_u64((u64)size * 4, 3);
422 if (val >= NFT_MAX_BUCKETS)
423 return NFT_MAX_BUCKETS;
425 return roundup_pow_of_two(val);
428 static bool nft_rhash_estimate(const struct nft_set_desc *desc, u32 features,
429 struct nft_set_estimate *est)
432 est->lookup = NFT_SET_CLASS_O_1;
433 est->space = NFT_SET_CLASS_O_N;
441 struct hlist_head table[];
444 struct nft_hash_elem {
445 struct hlist_node node;
446 struct nft_set_ext ext;
449 static bool nft_hash_lookup(const struct net *net, const struct nft_set *set,
450 const u32 *key, const struct nft_set_ext **ext)
452 struct nft_hash *priv = nft_set_priv(set);
453 u8 genmask = nft_genmask_cur(net);
454 const struct nft_hash_elem *he;
457 hash = jhash(key, set->klen, priv->seed);
458 hash = reciprocal_scale(hash, priv->buckets);
459 hlist_for_each_entry_rcu(he, &priv->table[hash], node) {
460 if (!memcmp(nft_set_ext_key(&he->ext), key, set->klen) &&
461 nft_set_elem_active(&he->ext, genmask)) {
469 static void *nft_hash_get(const struct net *net, const struct nft_set *set,
470 const struct nft_set_elem *elem, unsigned int flags)
472 struct nft_hash *priv = nft_set_priv(set);
473 u8 genmask = nft_genmask_cur(net);
474 struct nft_hash_elem *he;
477 hash = jhash(elem->key.val.data, set->klen, priv->seed);
478 hash = reciprocal_scale(hash, priv->buckets);
479 hlist_for_each_entry_rcu(he, &priv->table[hash], node) {
480 if (!memcmp(nft_set_ext_key(&he->ext), elem->key.val.data, set->klen) &&
481 nft_set_elem_active(&he->ext, genmask))
484 return ERR_PTR(-ENOENT);
487 static bool nft_hash_lookup_fast(const struct net *net,
488 const struct nft_set *set,
489 const u32 *key, const struct nft_set_ext **ext)
491 struct nft_hash *priv = nft_set_priv(set);
492 u8 genmask = nft_genmask_cur(net);
493 const struct nft_hash_elem *he;
497 hash = jhash_1word(k1, priv->seed);
498 hash = reciprocal_scale(hash, priv->buckets);
499 hlist_for_each_entry_rcu(he, &priv->table[hash], node) {
500 k2 = *(u32 *)nft_set_ext_key(&he->ext)->data;
502 nft_set_elem_active(&he->ext, genmask)) {
510 static u32 nft_jhash(const struct nft_set *set, const struct nft_hash *priv,
511 const struct nft_set_ext *ext)
513 const struct nft_data *key = nft_set_ext_key(ext);
516 if (set->klen == 4) {
518 hash = jhash_1word(k1, priv->seed);
520 hash = jhash(key, set->klen, priv->seed);
522 hash = reciprocal_scale(hash, priv->buckets);
527 static int nft_hash_insert(const struct net *net, const struct nft_set *set,
528 const struct nft_set_elem *elem,
529 struct nft_set_ext **ext)
531 struct nft_hash_elem *this = elem->priv, *he;
532 struct nft_hash *priv = nft_set_priv(set);
533 u8 genmask = nft_genmask_next(net);
536 hash = nft_jhash(set, priv, &this->ext);
537 hlist_for_each_entry(he, &priv->table[hash], node) {
538 if (!memcmp(nft_set_ext_key(&this->ext),
539 nft_set_ext_key(&he->ext), set->klen) &&
540 nft_set_elem_active(&he->ext, genmask)) {
545 hlist_add_head_rcu(&this->node, &priv->table[hash]);
549 static void nft_hash_activate(const struct net *net, const struct nft_set *set,
550 const struct nft_set_elem *elem)
552 struct nft_hash_elem *he = elem->priv;
554 nft_set_elem_change_active(net, set, &he->ext);
557 static bool nft_hash_flush(const struct net *net,
558 const struct nft_set *set, void *priv)
560 struct nft_hash_elem *he = priv;
562 nft_set_elem_change_active(net, set, &he->ext);
566 static void *nft_hash_deactivate(const struct net *net,
567 const struct nft_set *set,
568 const struct nft_set_elem *elem)
570 struct nft_hash *priv = nft_set_priv(set);
571 struct nft_hash_elem *this = elem->priv, *he;
572 u8 genmask = nft_genmask_next(net);
575 hash = nft_jhash(set, priv, &this->ext);
576 hlist_for_each_entry(he, &priv->table[hash], node) {
577 if (!memcmp(nft_set_ext_key(&he->ext), &elem->key.val,
579 nft_set_elem_active(&he->ext, genmask)) {
580 nft_set_elem_change_active(net, set, &he->ext);
587 static void nft_hash_remove(const struct net *net,
588 const struct nft_set *set,
589 const struct nft_set_elem *elem)
591 struct nft_hash_elem *he = elem->priv;
593 hlist_del_rcu(&he->node);
596 static void nft_hash_walk(const struct nft_ctx *ctx, struct nft_set *set,
597 struct nft_set_iter *iter)
599 struct nft_hash *priv = nft_set_priv(set);
600 struct nft_hash_elem *he;
601 struct nft_set_elem elem;
604 for (i = 0; i < priv->buckets; i++) {
605 hlist_for_each_entry_rcu(he, &priv->table[i], node) {
606 if (iter->count < iter->skip)
608 if (!nft_set_elem_active(&he->ext, iter->genmask))
613 iter->err = iter->fn(ctx, set, iter, &elem);
622 static u64 nft_hash_privsize(const struct nlattr * const nla[],
623 const struct nft_set_desc *desc)
625 return sizeof(struct nft_hash) +
626 (u64)nft_hash_buckets(desc->size) * sizeof(struct hlist_head);
629 static int nft_hash_init(const struct nft_set *set,
630 const struct nft_set_desc *desc,
631 const struct nlattr * const tb[])
633 struct nft_hash *priv = nft_set_priv(set);
635 priv->buckets = nft_hash_buckets(desc->size);
636 get_random_bytes(&priv->seed, sizeof(priv->seed));
641 static void nft_hash_destroy(const struct nft_set *set)
643 struct nft_hash *priv = nft_set_priv(set);
644 struct nft_hash_elem *he;
645 struct hlist_node *next;
648 for (i = 0; i < priv->buckets; i++) {
649 hlist_for_each_entry_safe(he, next, &priv->table[i], node) {
650 hlist_del_rcu(&he->node);
651 nft_set_elem_destroy(set, he, true);
656 static bool nft_hash_estimate(const struct nft_set_desc *desc, u32 features,
657 struct nft_set_estimate *est)
665 est->size = sizeof(struct nft_hash) +
666 (u64)nft_hash_buckets(desc->size) * sizeof(struct hlist_head) +
667 (u64)desc->size * sizeof(struct nft_hash_elem);
668 est->lookup = NFT_SET_CLASS_O_1;
669 est->space = NFT_SET_CLASS_O_N;
674 static bool nft_hash_fast_estimate(const struct nft_set_desc *desc, u32 features,
675 struct nft_set_estimate *est)
683 est->size = sizeof(struct nft_hash) +
684 (u64)nft_hash_buckets(desc->size) * sizeof(struct hlist_head) +
685 (u64)desc->size * sizeof(struct nft_hash_elem);
686 est->lookup = NFT_SET_CLASS_O_1;
687 est->space = NFT_SET_CLASS_O_N;
692 const struct nft_set_type nft_set_rhash_type = {
693 .features = NFT_SET_MAP | NFT_SET_OBJECT |
694 NFT_SET_TIMEOUT | NFT_SET_EVAL,
696 .privsize = nft_rhash_privsize,
697 .elemsize = offsetof(struct nft_rhash_elem, ext),
698 .estimate = nft_rhash_estimate,
699 .init = nft_rhash_init,
700 .gc_init = nft_rhash_gc_init,
701 .destroy = nft_rhash_destroy,
702 .insert = nft_rhash_insert,
703 .activate = nft_rhash_activate,
704 .deactivate = nft_rhash_deactivate,
705 .flush = nft_rhash_flush,
706 .remove = nft_rhash_remove,
707 .lookup = nft_rhash_lookup,
708 .update = nft_rhash_update,
709 .delete = nft_rhash_delete,
710 .walk = nft_rhash_walk,
711 .get = nft_rhash_get,
715 const struct nft_set_type nft_set_hash_type = {
716 .features = NFT_SET_MAP | NFT_SET_OBJECT,
718 .privsize = nft_hash_privsize,
719 .elemsize = offsetof(struct nft_hash_elem, ext),
720 .estimate = nft_hash_estimate,
721 .init = nft_hash_init,
722 .destroy = nft_hash_destroy,
723 .insert = nft_hash_insert,
724 .activate = nft_hash_activate,
725 .deactivate = nft_hash_deactivate,
726 .flush = nft_hash_flush,
727 .remove = nft_hash_remove,
728 .lookup = nft_hash_lookup,
729 .walk = nft_hash_walk,
734 const struct nft_set_type nft_set_hash_fast_type = {
735 .features = NFT_SET_MAP | NFT_SET_OBJECT,
737 .privsize = nft_hash_privsize,
738 .elemsize = offsetof(struct nft_hash_elem, ext),
739 .estimate = nft_hash_fast_estimate,
740 .init = nft_hash_init,
741 .destroy = nft_hash_destroy,
742 .insert = nft_hash_insert,
743 .activate = nft_hash_activate,
744 .deactivate = nft_hash_deactivate,
745 .flush = nft_hash_flush,
746 .remove = nft_hash_remove,
747 .lookup = nft_hash_lookup_fast,
748 .walk = nft_hash_walk,