1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2008-2009 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/rbtree.h>
13 #include <linux/netlink.h>
14 #include <linux/netfilter.h>
15 #include <linux/netfilter/nf_tables.h>
16 #include <net/netfilter/nf_tables_core.h>
21 seqcount_rwlock_t count;
22 struct delayed_work gc_work;
25 struct nft_rbtree_elem {
27 struct nft_set_ext ext;
30 static bool nft_rbtree_interval_end(const struct nft_rbtree_elem *rbe)
32 return nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) &&
33 (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END);
36 static bool nft_rbtree_interval_start(const struct nft_rbtree_elem *rbe)
38 return !nft_rbtree_interval_end(rbe);
41 static int nft_rbtree_cmp(const struct nft_set *set,
42 const struct nft_rbtree_elem *e1,
43 const struct nft_rbtree_elem *e2)
45 return memcmp(nft_set_ext_key(&e1->ext), nft_set_ext_key(&e2->ext),
49 static bool nft_rbtree_elem_expired(const struct nft_rbtree_elem *rbe)
51 return nft_set_elem_expired(&rbe->ext) ||
52 nft_set_elem_is_dead(&rbe->ext);
55 static bool __nft_rbtree_lookup(const struct net *net, const struct nft_set *set,
56 const u32 *key, const struct nft_set_ext **ext,
59 struct nft_rbtree *priv = nft_set_priv(set);
60 const struct nft_rbtree_elem *rbe, *interval = NULL;
61 u8 genmask = nft_genmask_cur(net);
62 const struct rb_node *parent;
65 parent = rcu_dereference_raw(priv->root.rb_node);
66 while (parent != NULL) {
67 if (read_seqcount_retry(&priv->count, seq))
70 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
72 d = memcmp(nft_set_ext_key(&rbe->ext), key, set->klen);
74 parent = rcu_dereference_raw(parent->rb_left);
76 !nft_rbtree_cmp(set, rbe, interval) &&
77 nft_rbtree_interval_end(rbe) &&
78 nft_rbtree_interval_start(interval))
82 parent = rcu_dereference_raw(parent->rb_right);
84 if (!nft_set_elem_active(&rbe->ext, genmask)) {
85 parent = rcu_dereference_raw(parent->rb_left);
89 if (nft_rbtree_elem_expired(rbe))
92 if (nft_rbtree_interval_end(rbe)) {
93 if (nft_set_is_anonymous(set))
95 parent = rcu_dereference_raw(parent->rb_left);
105 if (set->flags & NFT_SET_INTERVAL && interval != NULL &&
106 nft_set_elem_active(&interval->ext, genmask) &&
107 !nft_rbtree_elem_expired(interval) &&
108 nft_rbtree_interval_start(interval)) {
109 *ext = &interval->ext;
116 INDIRECT_CALLABLE_SCOPE
117 bool nft_rbtree_lookup(const struct net *net, const struct nft_set *set,
118 const u32 *key, const struct nft_set_ext **ext)
120 struct nft_rbtree *priv = nft_set_priv(set);
121 unsigned int seq = read_seqcount_begin(&priv->count);
124 ret = __nft_rbtree_lookup(net, set, key, ext, seq);
125 if (ret || !read_seqcount_retry(&priv->count, seq))
128 read_lock_bh(&priv->lock);
129 seq = read_seqcount_begin(&priv->count);
130 ret = __nft_rbtree_lookup(net, set, key, ext, seq);
131 read_unlock_bh(&priv->lock);
136 static bool __nft_rbtree_get(const struct net *net, const struct nft_set *set,
137 const u32 *key, struct nft_rbtree_elem **elem,
138 unsigned int seq, unsigned int flags, u8 genmask)
140 struct nft_rbtree_elem *rbe, *interval = NULL;
141 struct nft_rbtree *priv = nft_set_priv(set);
142 const struct rb_node *parent;
146 parent = rcu_dereference_raw(priv->root.rb_node);
147 while (parent != NULL) {
148 if (read_seqcount_retry(&priv->count, seq))
151 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
153 this = nft_set_ext_key(&rbe->ext);
154 d = memcmp(this, key, set->klen);
156 parent = rcu_dereference_raw(parent->rb_left);
157 if (!(flags & NFT_SET_ELEM_INTERVAL_END))
160 parent = rcu_dereference_raw(parent->rb_right);
161 if (flags & NFT_SET_ELEM_INTERVAL_END)
164 if (!nft_set_elem_active(&rbe->ext, genmask)) {
165 parent = rcu_dereference_raw(parent->rb_left);
169 if (nft_set_elem_expired(&rbe->ext))
172 if (!nft_set_ext_exists(&rbe->ext, NFT_SET_EXT_FLAGS) ||
173 (*nft_set_ext_flags(&rbe->ext) & NFT_SET_ELEM_INTERVAL_END) ==
174 (flags & NFT_SET_ELEM_INTERVAL_END)) {
179 if (nft_rbtree_interval_end(rbe))
182 parent = rcu_dereference_raw(parent->rb_left);
186 if (set->flags & NFT_SET_INTERVAL && interval != NULL &&
187 nft_set_elem_active(&interval->ext, genmask) &&
188 !nft_set_elem_expired(&interval->ext) &&
189 ((!nft_rbtree_interval_end(interval) &&
190 !(flags & NFT_SET_ELEM_INTERVAL_END)) ||
191 (nft_rbtree_interval_end(interval) &&
192 (flags & NFT_SET_ELEM_INTERVAL_END)))) {
200 static void *nft_rbtree_get(const struct net *net, const struct nft_set *set,
201 const struct nft_set_elem *elem, unsigned int flags)
203 struct nft_rbtree *priv = nft_set_priv(set);
204 unsigned int seq = read_seqcount_begin(&priv->count);
205 struct nft_rbtree_elem *rbe = ERR_PTR(-ENOENT);
206 const u32 *key = (const u32 *)&elem->key.val;
207 u8 genmask = nft_genmask_cur(net);
210 ret = __nft_rbtree_get(net, set, key, &rbe, seq, flags, genmask);
211 if (ret || !read_seqcount_retry(&priv->count, seq))
214 read_lock_bh(&priv->lock);
215 seq = read_seqcount_begin(&priv->count);
216 ret = __nft_rbtree_get(net, set, key, &rbe, seq, flags, genmask);
218 rbe = ERR_PTR(-ENOENT);
219 read_unlock_bh(&priv->lock);
224 static void nft_rbtree_gc_remove(struct net *net, struct nft_set *set,
225 struct nft_rbtree *priv,
226 struct nft_rbtree_elem *rbe)
228 struct nft_set_elem elem = {
232 nft_setelem_data_deactivate(net, set, &elem);
233 rb_erase(&rbe->node, &priv->root);
236 static int nft_rbtree_gc_elem(const struct nft_set *__set,
237 struct nft_rbtree *priv,
238 struct nft_rbtree_elem *rbe,
241 struct nft_set *set = (struct nft_set *)__set;
242 struct rb_node *prev = rb_prev(&rbe->node);
243 struct net *net = read_pnet(&set->net);
244 struct nft_rbtree_elem *rbe_prev;
245 struct nft_trans_gc *gc;
247 gc = nft_trans_gc_alloc(set, 0, GFP_ATOMIC);
251 /* search for end interval coming before this element.
252 * end intervals don't carry a timeout extension, they
253 * are coupled with the interval start element.
256 rbe_prev = rb_entry(prev, struct nft_rbtree_elem, node);
257 if (nft_rbtree_interval_end(rbe_prev) &&
258 nft_set_elem_active(&rbe_prev->ext, genmask))
261 prev = rb_prev(prev);
265 rbe_prev = rb_entry(prev, struct nft_rbtree_elem, node);
266 nft_rbtree_gc_remove(net, set, priv, rbe_prev);
268 /* There is always room in this trans gc for this element,
269 * memory allocation never actually happens, hence, the warning
270 * splat in such case. No need to set NFT_SET_ELEM_DEAD_BIT,
271 * this is synchronous gc which never fails.
273 gc = nft_trans_gc_queue_sync(gc, GFP_ATOMIC);
274 if (WARN_ON_ONCE(!gc))
277 nft_trans_gc_elem_add(gc, rbe_prev);
280 nft_rbtree_gc_remove(net, set, priv, rbe);
281 gc = nft_trans_gc_queue_sync(gc, GFP_ATOMIC);
282 if (WARN_ON_ONCE(!gc))
285 nft_trans_gc_elem_add(gc, rbe);
287 nft_trans_gc_queue_sync_done(gc);
292 static bool nft_rbtree_update_first(const struct nft_set *set,
293 struct nft_rbtree_elem *rbe,
294 struct rb_node *first)
296 struct nft_rbtree_elem *first_elem;
298 first_elem = rb_entry(first, struct nft_rbtree_elem, node);
299 /* this element is closest to where the new element is to be inserted:
300 * update the first element for the node list path.
302 if (nft_rbtree_cmp(set, rbe, first_elem) < 0)
308 static int __nft_rbtree_insert(const struct net *net, const struct nft_set *set,
309 struct nft_rbtree_elem *new,
310 struct nft_set_ext **ext)
312 struct nft_rbtree_elem *rbe, *rbe_le = NULL, *rbe_ge = NULL;
313 struct rb_node *node, *next, *parent, **p, *first = NULL;
314 struct nft_rbtree *priv = nft_set_priv(set);
315 u8 cur_genmask = nft_genmask_cur(net);
316 u8 genmask = nft_genmask_next(net);
319 /* Descend the tree to search for an existing element greater than the
320 * key value to insert that is greater than the new element. This is the
321 * first element to walk the ordered elements to find possible overlap.
324 p = &priv->root.rb_node;
327 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
328 d = nft_rbtree_cmp(set, rbe, new);
331 p = &parent->rb_left;
334 nft_rbtree_update_first(set, rbe, first))
337 p = &parent->rb_right;
339 if (nft_rbtree_interval_end(rbe))
340 p = &parent->rb_left;
342 p = &parent->rb_right;
347 first = rb_first(&priv->root);
349 /* Detect overlap by going through the list of valid tree nodes.
350 * Values stored in the tree are in reversed order, starting from
351 * highest to lowest value.
353 for (node = first; node != NULL; node = next) {
354 next = rb_next(node);
356 rbe = rb_entry(node, struct nft_rbtree_elem, node);
358 if (!nft_set_elem_active(&rbe->ext, genmask))
361 /* perform garbage collection to avoid bogus overlap reports
362 * but skip new elements in this transaction.
364 if (nft_set_elem_expired(&rbe->ext) &&
365 nft_set_elem_active(&rbe->ext, cur_genmask)) {
366 err = nft_rbtree_gc_elem(set, priv, rbe, genmask);
373 d = nft_rbtree_cmp(set, rbe, new);
375 /* Matching end element: no need to look for an
376 * overlapping greater or equal element.
378 if (nft_rbtree_interval_end(rbe)) {
383 /* first element that is greater or equal to key value. */
389 /* this is a closer more or equal element, update it. */
390 if (nft_rbtree_cmp(set, rbe_ge, new) != 0) {
395 /* element is equal to key value, make sure flags are
396 * the same, an existing more or equal start element
397 * must not be replaced by more or equal end element.
399 if ((nft_rbtree_interval_start(new) &&
400 nft_rbtree_interval_start(rbe_ge)) ||
401 (nft_rbtree_interval_end(new) &&
402 nft_rbtree_interval_end(rbe_ge))) {
407 /* annotate element greater than the new element. */
411 /* annotate element less than the new element. */
417 /* - new start element matching existing start element: full overlap
418 * reported as -EEXIST, cleared by caller if NLM_F_EXCL is not given.
420 if (rbe_ge && !nft_rbtree_cmp(set, new, rbe_ge) &&
421 nft_rbtree_interval_start(rbe_ge) == nft_rbtree_interval_start(new)) {
426 /* - new end element matching existing end element: full overlap
427 * reported as -EEXIST, cleared by caller if NLM_F_EXCL is not given.
429 if (rbe_le && !nft_rbtree_cmp(set, new, rbe_le) &&
430 nft_rbtree_interval_end(rbe_le) == nft_rbtree_interval_end(new)) {
435 /* - new start element with existing closest, less or equal key value
436 * being a start element: partial overlap, reported as -ENOTEMPTY.
437 * Anonymous sets allow for two consecutive start element since they
438 * are constant, skip them to avoid bogus overlap reports.
440 if (!nft_set_is_anonymous(set) && rbe_le &&
441 nft_rbtree_interval_start(rbe_le) && nft_rbtree_interval_start(new))
444 /* - new end element with existing closest, less or equal key value
445 * being a end element: partial overlap, reported as -ENOTEMPTY.
448 nft_rbtree_interval_end(rbe_le) && nft_rbtree_interval_end(new))
451 /* - new end element with existing closest, greater or equal key value
452 * being an end element: partial overlap, reported as -ENOTEMPTY
455 nft_rbtree_interval_end(rbe_ge) && nft_rbtree_interval_end(new))
458 /* Accepted element: pick insertion point depending on key value */
460 p = &priv->root.rb_node;
463 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
464 d = nft_rbtree_cmp(set, rbe, new);
467 p = &parent->rb_left;
469 p = &parent->rb_right;
470 else if (nft_rbtree_interval_end(rbe))
471 p = &parent->rb_left;
473 p = &parent->rb_right;
476 rb_link_node_rcu(&new->node, parent, p);
477 rb_insert_color(&new->node, &priv->root);
481 static int nft_rbtree_insert(const struct net *net, const struct nft_set *set,
482 const struct nft_set_elem *elem,
483 struct nft_set_ext **ext)
485 struct nft_rbtree *priv = nft_set_priv(set);
486 struct nft_rbtree_elem *rbe = elem->priv;
489 write_lock_bh(&priv->lock);
490 write_seqcount_begin(&priv->count);
491 err = __nft_rbtree_insert(net, set, rbe, ext);
492 write_seqcount_end(&priv->count);
493 write_unlock_bh(&priv->lock);
498 static void nft_rbtree_remove(const struct net *net,
499 const struct nft_set *set,
500 const struct nft_set_elem *elem)
502 struct nft_rbtree *priv = nft_set_priv(set);
503 struct nft_rbtree_elem *rbe = elem->priv;
505 write_lock_bh(&priv->lock);
506 write_seqcount_begin(&priv->count);
507 rb_erase(&rbe->node, &priv->root);
508 write_seqcount_end(&priv->count);
509 write_unlock_bh(&priv->lock);
512 static void nft_rbtree_activate(const struct net *net,
513 const struct nft_set *set,
514 const struct nft_set_elem *elem)
516 struct nft_rbtree_elem *rbe = elem->priv;
518 nft_set_elem_change_active(net, set, &rbe->ext);
521 static bool nft_rbtree_flush(const struct net *net,
522 const struct nft_set *set, void *priv)
524 struct nft_rbtree_elem *rbe = priv;
526 nft_set_elem_change_active(net, set, &rbe->ext);
531 static void *nft_rbtree_deactivate(const struct net *net,
532 const struct nft_set *set,
533 const struct nft_set_elem *elem)
535 const struct nft_rbtree *priv = nft_set_priv(set);
536 const struct rb_node *parent = priv->root.rb_node;
537 struct nft_rbtree_elem *rbe, *this = elem->priv;
538 u8 genmask = nft_genmask_next(net);
541 while (parent != NULL) {
542 rbe = rb_entry(parent, struct nft_rbtree_elem, node);
544 d = memcmp(nft_set_ext_key(&rbe->ext), &elem->key.val,
547 parent = parent->rb_left;
549 parent = parent->rb_right;
551 if (nft_rbtree_interval_end(rbe) &&
552 nft_rbtree_interval_start(this)) {
553 parent = parent->rb_left;
555 } else if (nft_rbtree_interval_start(rbe) &&
556 nft_rbtree_interval_end(this)) {
557 parent = parent->rb_right;
559 } else if (!nft_set_elem_active(&rbe->ext, genmask)) {
560 parent = parent->rb_left;
563 nft_rbtree_flush(net, set, rbe);
570 static void nft_rbtree_walk(const struct nft_ctx *ctx,
572 struct nft_set_iter *iter)
574 struct nft_rbtree *priv = nft_set_priv(set);
575 struct nft_rbtree_elem *rbe;
576 struct nft_set_elem elem;
577 struct rb_node *node;
579 read_lock_bh(&priv->lock);
580 for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
581 rbe = rb_entry(node, struct nft_rbtree_elem, node);
583 if (iter->count < iter->skip)
585 if (!nft_set_elem_active(&rbe->ext, iter->genmask))
590 iter->err = iter->fn(ctx, set, iter, &elem);
592 read_unlock_bh(&priv->lock);
598 read_unlock_bh(&priv->lock);
601 static void nft_rbtree_gc(struct work_struct *work)
603 struct nft_rbtree_elem *rbe, *rbe_end = NULL;
604 struct nftables_pernet *nft_net;
605 struct nft_rbtree *priv;
606 struct nft_trans_gc *gc;
607 struct rb_node *node;
612 priv = container_of(work, struct nft_rbtree, gc_work.work);
613 set = nft_set_container_of(priv);
614 net = read_pnet(&set->net);
615 nft_net = nft_pernet(net);
616 gc_seq = READ_ONCE(nft_net->gc_seq);
618 if (nft_set_gc_is_pending(set))
621 gc = nft_trans_gc_alloc(set, gc_seq, GFP_KERNEL);
625 write_lock_bh(&priv->lock);
626 write_seqcount_begin(&priv->count);
627 for (node = rb_first(&priv->root); node != NULL; node = rb_next(node)) {
629 /* Ruleset has been updated, try later. */
630 if (READ_ONCE(nft_net->gc_seq) != gc_seq) {
631 nft_trans_gc_destroy(gc);
636 rbe = rb_entry(node, struct nft_rbtree_elem, node);
638 if (nft_set_elem_is_dead(&rbe->ext))
641 /* elements are reversed in the rbtree for historical reasons,
642 * from highest to lowest value, that is why end element is
643 * always visited before the start element.
645 if (nft_rbtree_interval_end(rbe)) {
649 if (!nft_set_elem_expired(&rbe->ext))
652 nft_set_elem_dead(&rbe->ext);
657 nft_set_elem_dead(&rbe_end->ext);
659 gc = nft_trans_gc_queue_async(gc, gc_seq, GFP_ATOMIC);
663 nft_trans_gc_elem_add(gc, rbe_end);
666 gc = nft_trans_gc_queue_async(gc, gc_seq, GFP_ATOMIC);
670 nft_trans_gc_elem_add(gc, rbe);
673 gc = nft_trans_gc_catchall(gc, gc_seq);
676 write_seqcount_end(&priv->count);
677 write_unlock_bh(&priv->lock);
680 nft_trans_gc_queue_async_done(gc);
682 queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
683 nft_set_gc_interval(set));
686 static u64 nft_rbtree_privsize(const struct nlattr * const nla[],
687 const struct nft_set_desc *desc)
689 return sizeof(struct nft_rbtree);
692 static int nft_rbtree_init(const struct nft_set *set,
693 const struct nft_set_desc *desc,
694 const struct nlattr * const nla[])
696 struct nft_rbtree *priv = nft_set_priv(set);
698 rwlock_init(&priv->lock);
699 seqcount_rwlock_init(&priv->count, &priv->lock);
700 priv->root = RB_ROOT;
702 INIT_DEFERRABLE_WORK(&priv->gc_work, nft_rbtree_gc);
703 if (set->flags & NFT_SET_TIMEOUT)
704 queue_delayed_work(system_power_efficient_wq, &priv->gc_work,
705 nft_set_gc_interval(set));
710 static void nft_rbtree_destroy(const struct nft_ctx *ctx,
711 const struct nft_set *set)
713 struct nft_rbtree *priv = nft_set_priv(set);
714 struct nft_rbtree_elem *rbe;
715 struct rb_node *node;
717 cancel_delayed_work_sync(&priv->gc_work);
719 while ((node = priv->root.rb_node) != NULL) {
720 rb_erase(node, &priv->root);
721 rbe = rb_entry(node, struct nft_rbtree_elem, node);
722 nf_tables_set_elem_destroy(ctx, set, rbe);
726 static bool nft_rbtree_estimate(const struct nft_set_desc *desc, u32 features,
727 struct nft_set_estimate *est)
729 if (desc->field_count > 1)
733 est->size = sizeof(struct nft_rbtree) +
734 desc->size * sizeof(struct nft_rbtree_elem);
738 est->lookup = NFT_SET_CLASS_O_LOG_N;
739 est->space = NFT_SET_CLASS_O_N;
744 const struct nft_set_type nft_set_rbtree_type = {
745 .features = NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_OBJECT | NFT_SET_TIMEOUT,
747 .privsize = nft_rbtree_privsize,
748 .elemsize = offsetof(struct nft_rbtree_elem, ext),
749 .estimate = nft_rbtree_estimate,
750 .init = nft_rbtree_init,
751 .destroy = nft_rbtree_destroy,
752 .insert = nft_rbtree_insert,
753 .remove = nft_rbtree_remove,
754 .deactivate = nft_rbtree_deactivate,
755 .flush = nft_rbtree_flush,
756 .activate = nft_rbtree_activate,
757 .lookup = nft_rbtree_lookup,
758 .walk = nft_rbtree_walk,
759 .get = nft_rbtree_get,