1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
6 #include "allowedips.h"
9 static void swap_endian(u8 *dst, const u8 *src, u8 bits)
12 *(u32 *)dst = be32_to_cpu(*(const __be32 *)src);
13 } else if (bits == 128) {
14 ((u64 *)dst)[0] = be64_to_cpu(((const __be64 *)src)[0]);
15 ((u64 *)dst)[1] = be64_to_cpu(((const __be64 *)src)[1]);
19 static void copy_and_assign_cidr(struct allowedips_node *node, const u8 *src,
23 node->bit_at_a = cidr / 8U;
24 #ifdef __LITTLE_ENDIAN
25 node->bit_at_a ^= (bits / 8U - 1U) % 8U;
27 node->bit_at_b = 7U - (cidr % 8U);
29 memcpy(node->bits, src, bits / 8U);
31 #define CHOOSE_NODE(parent, key) \
32 parent->bit[(key[parent->bit_at_a] >> parent->bit_at_b) & 1]
34 static void push_rcu(struct allowedips_node **stack,
35 struct allowedips_node __rcu *p, unsigned int *len)
37 if (rcu_access_pointer(p)) {
38 WARN_ON(IS_ENABLED(DEBUG) && *len >= 128);
39 stack[(*len)++] = rcu_dereference_raw(p);
43 static void root_free_rcu(struct rcu_head *rcu)
45 struct allowedips_node *node, *stack[128] = {
46 container_of(rcu, struct allowedips_node, rcu) };
49 while (len > 0 && (node = stack[--len])) {
50 push_rcu(stack, node->bit[0], &len);
51 push_rcu(stack, node->bit[1], &len);
56 static void root_remove_peer_lists(struct allowedips_node *root)
58 struct allowedips_node *node, *stack[128] = { root };
61 while (len > 0 && (node = stack[--len])) {
62 push_rcu(stack, node->bit[0], &len);
63 push_rcu(stack, node->bit[1], &len);
64 if (rcu_access_pointer(node->peer))
65 list_del(&node->peer_list);
69 static unsigned int fls128(u64 a, u64 b)
71 return a ? fls64(a) + 64U : fls64(b);
74 static u8 common_bits(const struct allowedips_node *node, const u8 *key,
78 return 32U - fls(*(const u32 *)node->bits ^ *(const u32 *)key);
81 *(const u64 *)&node->bits[0] ^ *(const u64 *)&key[0],
82 *(const u64 *)&node->bits[8] ^ *(const u64 *)&key[8]);
86 static bool prefix_matches(const struct allowedips_node *node, const u8 *key,
89 /* This could be much faster if it actually just compared the common
90 * bits properly, by precomputing a mask bswap(~0 << (32 - cidr)), and
91 * the rest, but it turns out that common_bits is already super fast on
92 * modern processors, even taking into account the unfortunate bswap.
93 * So, we just inline it like this instead.
95 return common_bits(node, key, bits) >= node->cidr;
98 static struct allowedips_node *find_node(struct allowedips_node *trie, u8 bits,
101 struct allowedips_node *node = trie, *found = NULL;
103 while (node && prefix_matches(node, key, bits)) {
104 if (rcu_access_pointer(node->peer))
106 if (node->cidr == bits)
108 node = rcu_dereference_bh(CHOOSE_NODE(node, key));
113 /* Returns a strong reference to a peer */
114 static struct wg_peer *lookup(struct allowedips_node __rcu *root, u8 bits,
117 /* Aligned so it can be passed to fls/fls64 */
118 u8 ip[16] __aligned(__alignof(u64));
119 struct allowedips_node *node;
120 struct wg_peer *peer = NULL;
122 swap_endian(ip, be_ip, bits);
126 node = find_node(rcu_dereference_bh(root), bits, ip);
128 peer = wg_peer_get_maybe_zero(rcu_dereference_bh(node->peer));
132 rcu_read_unlock_bh();
136 static bool node_placement(struct allowedips_node __rcu *trie, const u8 *key,
137 u8 cidr, u8 bits, struct allowedips_node **rnode,
140 struct allowedips_node *node = rcu_dereference_protected(trie,
141 lockdep_is_held(lock));
142 struct allowedips_node *parent = NULL;
145 while (node && node->cidr <= cidr && prefix_matches(node, key, bits)) {
147 if (parent->cidr == cidr) {
151 node = rcu_dereference_protected(CHOOSE_NODE(parent, key),
152 lockdep_is_held(lock));
158 static int add(struct allowedips_node __rcu **trie, u8 bits, const u8 *key,
159 u8 cidr, struct wg_peer *peer, struct mutex *lock)
161 struct allowedips_node *node, *parent, *down, *newnode;
163 if (unlikely(cidr > bits || !peer))
166 if (!rcu_access_pointer(*trie)) {
167 node = kzalloc(sizeof(*node), GFP_KERNEL);
170 RCU_INIT_POINTER(node->peer, peer);
171 list_add_tail(&node->peer_list, &peer->allowedips_list);
172 copy_and_assign_cidr(node, key, cidr, bits);
173 rcu_assign_pointer(node->parent_bit, trie);
174 rcu_assign_pointer(*trie, node);
177 if (node_placement(*trie, key, cidr, bits, &node, lock)) {
178 rcu_assign_pointer(node->peer, peer);
179 list_move_tail(&node->peer_list, &peer->allowedips_list);
183 newnode = kzalloc(sizeof(*newnode), GFP_KERNEL);
184 if (unlikely(!newnode))
186 RCU_INIT_POINTER(newnode->peer, peer);
187 list_add_tail(&newnode->peer_list, &peer->allowedips_list);
188 copy_and_assign_cidr(newnode, key, cidr, bits);
191 down = rcu_dereference_protected(*trie, lockdep_is_held(lock));
193 down = rcu_dereference_protected(CHOOSE_NODE(node, key), lockdep_is_held(lock));
195 rcu_assign_pointer(newnode->parent_bit, &CHOOSE_NODE(node, key));
196 rcu_assign_pointer(CHOOSE_NODE(node, key), newnode);
200 cidr = min(cidr, common_bits(down, key, bits));
203 if (newnode->cidr == cidr) {
204 rcu_assign_pointer(down->parent_bit, &CHOOSE_NODE(newnode, down->bits));
205 rcu_assign_pointer(CHOOSE_NODE(newnode, down->bits), down);
207 rcu_assign_pointer(newnode->parent_bit, trie);
208 rcu_assign_pointer(*trie, newnode);
210 rcu_assign_pointer(newnode->parent_bit, &CHOOSE_NODE(parent, newnode->bits));
211 rcu_assign_pointer(CHOOSE_NODE(parent, newnode->bits), newnode);
216 node = kzalloc(sizeof(*node), GFP_KERNEL);
217 if (unlikely(!node)) {
218 list_del(&newnode->peer_list);
222 INIT_LIST_HEAD(&node->peer_list);
223 copy_and_assign_cidr(node, newnode->bits, cidr, bits);
225 rcu_assign_pointer(down->parent_bit, &CHOOSE_NODE(node, down->bits));
226 rcu_assign_pointer(CHOOSE_NODE(node, down->bits), down);
227 rcu_assign_pointer(newnode->parent_bit, &CHOOSE_NODE(node, newnode->bits));
228 rcu_assign_pointer(CHOOSE_NODE(node, newnode->bits), newnode);
230 rcu_assign_pointer(node->parent_bit, trie);
231 rcu_assign_pointer(*trie, node);
233 rcu_assign_pointer(node->parent_bit, &CHOOSE_NODE(parent, node->bits));
234 rcu_assign_pointer(CHOOSE_NODE(parent, node->bits), node);
239 void wg_allowedips_init(struct allowedips *table)
241 table->root4 = table->root6 = NULL;
245 void wg_allowedips_free(struct allowedips *table, struct mutex *lock)
247 struct allowedips_node __rcu *old4 = table->root4, *old6 = table->root6;
250 RCU_INIT_POINTER(table->root4, NULL);
251 RCU_INIT_POINTER(table->root6, NULL);
252 if (rcu_access_pointer(old4)) {
253 struct allowedips_node *node = rcu_dereference_protected(old4,
254 lockdep_is_held(lock));
256 root_remove_peer_lists(node);
257 call_rcu(&node->rcu, root_free_rcu);
259 if (rcu_access_pointer(old6)) {
260 struct allowedips_node *node = rcu_dereference_protected(old6,
261 lockdep_is_held(lock));
263 root_remove_peer_lists(node);
264 call_rcu(&node->rcu, root_free_rcu);
268 int wg_allowedips_insert_v4(struct allowedips *table, const struct in_addr *ip,
269 u8 cidr, struct wg_peer *peer, struct mutex *lock)
271 /* Aligned so it can be passed to fls */
272 u8 key[4] __aligned(__alignof(u32));
275 swap_endian(key, (const u8 *)ip, 32);
276 return add(&table->root4, 32, key, cidr, peer, lock);
279 int wg_allowedips_insert_v6(struct allowedips *table, const struct in6_addr *ip,
280 u8 cidr, struct wg_peer *peer, struct mutex *lock)
282 /* Aligned so it can be passed to fls64 */
283 u8 key[16] __aligned(__alignof(u64));
286 swap_endian(key, (const u8 *)ip, 128);
287 return add(&table->root6, 128, key, cidr, peer, lock);
290 void wg_allowedips_remove_by_peer(struct allowedips *table,
291 struct wg_peer *peer, struct mutex *lock)
293 struct allowedips_node *node, *child, *tmp;
295 if (list_empty(&peer->allowedips_list))
298 list_for_each_entry_safe(node, tmp, &peer->allowedips_list, peer_list) {
299 list_del_init(&node->peer_list);
300 RCU_INIT_POINTER(node->peer, NULL);
301 if (node->bit[0] && node->bit[1])
303 child = rcu_dereference_protected(
304 node->bit[!rcu_access_pointer(node->bit[0])],
305 lockdep_is_held(lock));
307 child->parent_bit = node->parent_bit;
308 *rcu_dereference_protected(node->parent_bit, lockdep_is_held(lock)) = child;
309 kfree_rcu(node, rcu);
311 /* TODO: Note that we currently don't walk up and down in order to
312 * free any potential filler nodes. This means that this function
313 * doesn't free up as much as it could, which could be revisited
319 int wg_allowedips_read_node(struct allowedips_node *node, u8 ip[16], u8 *cidr)
321 const unsigned int cidr_bytes = DIV_ROUND_UP(node->cidr, 8U);
322 swap_endian(ip, node->bits, node->bitlen);
323 memset(ip + cidr_bytes, 0, node->bitlen / 8U - cidr_bytes);
325 ip[cidr_bytes - 1U] &= ~0U << (-node->cidr % 8U);
328 return node->bitlen == 32 ? AF_INET : AF_INET6;
331 /* Returns a strong reference to a peer */
332 struct wg_peer *wg_allowedips_lookup_dst(struct allowedips *table,
335 if (skb->protocol == htons(ETH_P_IP))
336 return lookup(table->root4, 32, &ip_hdr(skb)->daddr);
337 else if (skb->protocol == htons(ETH_P_IPV6))
338 return lookup(table->root6, 128, &ipv6_hdr(skb)->daddr);
342 /* Returns a strong reference to a peer */
343 struct wg_peer *wg_allowedips_lookup_src(struct allowedips *table,
346 if (skb->protocol == htons(ETH_P_IP))
347 return lookup(table->root4, 32, &ip_hdr(skb)->saddr);
348 else if (skb->protocol == htons(ETH_P_IPV6))
349 return lookup(table->root6, 128, &ipv6_hdr(skb)->saddr);
353 #include "selftest/allowedips.c"