]>
Commit | Line | Data |
---|---|---|
d342894c | 1 | /* |
eb5ce439 | 2 | * VXLAN: Virtual eXtensible Local Area Network |
d342894c | 3 | * |
3b8df3c6 | 4 | * Copyright (c) 2012-2013 Vyatta Inc. |
d342894c | 5 | * |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 as | |
8 | * published by the Free Software Foundation. | |
d342894c | 9 | */ |
10 | ||
11 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
12 | ||
13 | #include <linux/kernel.h> | |
d342894c | 14 | #include <linux/module.h> |
15 | #include <linux/errno.h> | |
16 | #include <linux/slab.h> | |
d342894c | 17 | #include <linux/udp.h> |
18 | #include <linux/igmp.h> | |
d342894c | 19 | #include <linux/if_ether.h> |
1b13c97f | 20 | #include <linux/ethtool.h> |
e4f67add DS |
21 | #include <net/arp.h> |
22 | #include <net/ndisc.h> | |
d342894c | 23 | #include <net/ip.h> |
24 | #include <net/icmp.h> | |
d342894c | 25 | #include <net/rtnetlink.h> |
d342894c | 26 | #include <net/inet_ecn.h> |
27 | #include <net/net_namespace.h> | |
28 | #include <net/netns/generic.h> | |
fa20e0e3 | 29 | #include <net/tun_proto.h> |
012a5729 | 30 | #include <net/vxlan.h> |
40d29af0 | 31 | |
e4c7ed41 | 32 | #if IS_ENABLED(CONFIG_IPV6) |
e4c7ed41 | 33 | #include <net/ip6_tunnel.h> |
660d98ca | 34 | #include <net/ip6_checksum.h> |
e4c7ed41 | 35 | #endif |
d342894c | 36 | |
37 | #define VXLAN_VERSION "0.1" | |
38 | ||
553675fb | 39 | #define PORT_HASH_BITS 8 |
40 | #define PORT_HASH_SIZE (1<<PORT_HASH_BITS) | |
d342894c | 41 | #define FDB_AGE_DEFAULT 300 /* 5 min */ |
42 | #define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */ | |
43 | ||
23c578bf | 44 | /* UDP port for VXLAN traffic. |
45 | * The IANA assigned port is 4789, but the Linux default is 8472 | |
234f5b73 | 46 | * for compatibility with early adopters. |
23c578bf | 47 | */ |
9daaa397 SH |
48 | static unsigned short vxlan_port __read_mostly = 8472; |
49 | module_param_named(udp_port, vxlan_port, ushort, 0444); | |
d342894c | 50 | MODULE_PARM_DESC(udp_port, "Destination UDP port"); |
51 | ||
52 | static bool log_ecn_error = true; | |
53 | module_param(log_ecn_error, bool, 0644); | |
54 | MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN"); | |
55 | ||
c7d03a00 | 56 | static unsigned int vxlan_net_id; |
0dfbdf41 | 57 | static struct rtnl_link_ops vxlan_link_ops; |
553675fb | 58 | |
7256eac1 | 59 | static const u8 all_zeros_mac[ETH_ALEN + 2]; |
afbd8bae | 60 | |
205f356d | 61 | static int vxlan_sock_add(struct vxlan_dev *vxlan); |
614732ea | 62 | |
a53cb29b MB |
63 | static void vxlan_vs_del_dev(struct vxlan_dev *vxlan); |
64 | ||
553675fb | 65 | /* per-network namespace private data for this module */ |
66 | struct vxlan_net { | |
67 | struct list_head vxlan_list; | |
68 | struct hlist_head sock_list[PORT_HASH_SIZE]; | |
1c51a915 | 69 | spinlock_t sock_lock; |
553675fb | 70 | }; |
71 | ||
d342894c | 72 | /* Forwarding table entry */ |
73 | struct vxlan_fdb { | |
74 | struct hlist_node hlist; /* linked list of entries */ | |
75 | struct rcu_head rcu; | |
76 | unsigned long updated; /* jiffies */ | |
77 | unsigned long used; | |
3e61aa8f | 78 | struct list_head remotes; |
7177a3b0 | 79 | u8 eth_addr[ETH_ALEN]; |
d342894c | 80 | u16 state; /* see ndm_state */ |
3ad7a4b1 | 81 | __be32 vni; |
ae884082 | 82 | u8 flags; /* see ndm_flags */ |
d342894c | 83 | }; |
84 | ||
d342894c | 85 | /* salt for hash table */ |
86 | static u32 vxlan_salt __read_mostly; | |
87 | ||
ee122c79 TG |
88 | static inline bool vxlan_collect_metadata(struct vxlan_sock *vs) |
89 | { | |
e7030878 TG |
90 | return vs->flags & VXLAN_F_COLLECT_METADATA || |
91 | ip_tunnel_collect_metadata(); | |
ee122c79 TG |
92 | } |
93 | ||
e4c7ed41 CW |
94 | #if IS_ENABLED(CONFIG_IPV6) |
95 | static inline | |
96 | bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b) | |
97 | { | |
f0ef3126 JB |
98 | if (a->sa.sa_family != b->sa.sa_family) |
99 | return false; | |
100 | if (a->sa.sa_family == AF_INET6) | |
101 | return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr); | |
102 | else | |
103 | return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr; | |
e4c7ed41 CW |
104 | } |
105 | ||
106 | static inline bool vxlan_addr_any(const union vxlan_addr *ipa) | |
107 | { | |
f0ef3126 JB |
108 | if (ipa->sa.sa_family == AF_INET6) |
109 | return ipv6_addr_any(&ipa->sin6.sin6_addr); | |
110 | else | |
111 | return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY); | |
e4c7ed41 CW |
112 | } |
113 | ||
114 | static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa) | |
115 | { | |
f0ef3126 JB |
116 | if (ipa->sa.sa_family == AF_INET6) |
117 | return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr); | |
118 | else | |
119 | return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr)); | |
e4c7ed41 CW |
120 | } |
121 | ||
122 | static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla) | |
123 | { | |
f0ef3126 | 124 | if (nla_len(nla) >= sizeof(struct in6_addr)) { |
67b61f6c | 125 | ip->sin6.sin6_addr = nla_get_in6_addr(nla); |
f0ef3126 JB |
126 | ip->sa.sa_family = AF_INET6; |
127 | return 0; | |
128 | } else if (nla_len(nla) >= sizeof(__be32)) { | |
67b61f6c | 129 | ip->sin.sin_addr.s_addr = nla_get_in_addr(nla); |
f0ef3126 JB |
130 | ip->sa.sa_family = AF_INET; |
131 | return 0; | |
132 | } else { | |
133 | return -EAFNOSUPPORT; | |
134 | } | |
e4c7ed41 CW |
135 | } |
136 | ||
137 | static int vxlan_nla_put_addr(struct sk_buff *skb, int attr, | |
f0ef3126 | 138 | const union vxlan_addr *ip) |
e4c7ed41 | 139 | { |
f0ef3126 | 140 | if (ip->sa.sa_family == AF_INET6) |
930345ea | 141 | return nla_put_in6_addr(skb, attr, &ip->sin6.sin6_addr); |
f0ef3126 | 142 | else |
930345ea | 143 | return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr); |
e4c7ed41 CW |
144 | } |
145 | ||
146 | #else /* !CONFIG_IPV6 */ | |
147 | ||
148 | static inline | |
149 | bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b) | |
150 | { | |
f0ef3126 | 151 | return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr; |
e4c7ed41 CW |
152 | } |
153 | ||
154 | static inline bool vxlan_addr_any(const union vxlan_addr *ipa) | |
155 | { | |
f0ef3126 | 156 | return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY); |
e4c7ed41 CW |
157 | } |
158 | ||
159 | static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa) | |
160 | { | |
f0ef3126 | 161 | return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr)); |
e4c7ed41 CW |
162 | } |
163 | ||
164 | static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla) | |
165 | { | |
f0ef3126 JB |
166 | if (nla_len(nla) >= sizeof(struct in6_addr)) { |
167 | return -EAFNOSUPPORT; | |
168 | } else if (nla_len(nla) >= sizeof(__be32)) { | |
67b61f6c | 169 | ip->sin.sin_addr.s_addr = nla_get_in_addr(nla); |
f0ef3126 JB |
170 | ip->sa.sa_family = AF_INET; |
171 | return 0; | |
172 | } else { | |
173 | return -EAFNOSUPPORT; | |
174 | } | |
e4c7ed41 CW |
175 | } |
176 | ||
177 | static int vxlan_nla_put_addr(struct sk_buff *skb, int attr, | |
f0ef3126 | 178 | const union vxlan_addr *ip) |
e4c7ed41 | 179 | { |
930345ea | 180 | return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr); |
e4c7ed41 CW |
181 | } |
182 | #endif | |
183 | ||
553675fb | 184 | /* Virtual Network hash table head */ |
54bfd872 | 185 | static inline struct hlist_head *vni_head(struct vxlan_sock *vs, __be32 vni) |
553675fb | 186 | { |
54bfd872 | 187 | return &vs->vni_list[hash_32((__force u32)vni, VNI_HASH_BITS)]; |
553675fb | 188 | } |
189 | ||
190 | /* Socket hash table head */ | |
191 | static inline struct hlist_head *vs_head(struct net *net, __be16 port) | |
d342894c | 192 | { |
193 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
194 | ||
553675fb | 195 | return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)]; |
196 | } | |
197 | ||
3e61aa8f SH |
198 | /* First remote destination for a forwarding entry. |
199 | * Guaranteed to be non-NULL because remotes are never deleted. | |
200 | */ | |
5ca5461c | 201 | static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb) |
3e61aa8f | 202 | { |
5ca5461c | 203 | return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list); |
204 | } | |
205 | ||
206 | static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb) | |
207 | { | |
208 | return list_first_entry(&fdb->remotes, struct vxlan_rdst, list); | |
3e61aa8f SH |
209 | } |
210 | ||
ac5132d1 TG |
211 | /* Find VXLAN socket based on network namespace, address family and UDP port |
212 | * and enabled unshareable flags. | |
213 | */ | |
214 | static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family, | |
215 | __be16 port, u32 flags) | |
553675fb | 216 | { |
217 | struct vxlan_sock *vs; | |
af33c1ad TH |
218 | |
219 | flags &= VXLAN_F_RCV_FLAGS; | |
553675fb | 220 | |
221 | hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) { | |
19ca9fc1 | 222 | if (inet_sk(vs->sock->sk)->inet_sport == port && |
705cc62f | 223 | vxlan_get_sk_family(vs) == family && |
af33c1ad | 224 | vs->flags == flags) |
553675fb | 225 | return vs; |
226 | } | |
227 | return NULL; | |
d342894c | 228 | } |
229 | ||
49f810f0 MS |
230 | static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, int ifindex, |
231 | __be32 vni) | |
d342894c | 232 | { |
69e76661 | 233 | struct vxlan_dev_node *node; |
d342894c | 234 | |
b9167b2e JB |
235 | /* For flow based devices, map all packets to VNI 0 */ |
236 | if (vs->flags & VXLAN_F_COLLECT_METADATA) | |
237 | vni = 0; | |
238 | ||
69e76661 JB |
239 | hlist_for_each_entry_rcu(node, vni_head(vs, vni), hlist) { |
240 | if (node->vxlan->default_dst.remote_vni != vni) | |
49f810f0 MS |
241 | continue; |
242 | ||
243 | if (IS_ENABLED(CONFIG_IPV6)) { | |
69e76661 | 244 | const struct vxlan_config *cfg = &node->vxlan->cfg; |
49f810f0 MS |
245 | |
246 | if ((cfg->flags & VXLAN_F_IPV6_LINKLOCAL) && | |
247 | cfg->remote_ifindex != ifindex) | |
248 | continue; | |
249 | } | |
250 | ||
69e76661 | 251 | return node->vxlan; |
d342894c | 252 | } |
253 | ||
254 | return NULL; | |
255 | } | |
256 | ||
5cfccc5a | 257 | /* Look up VNI in a per net namespace table */ |
49f810f0 MS |
258 | static struct vxlan_dev *vxlan_find_vni(struct net *net, int ifindex, |
259 | __be32 vni, sa_family_t family, | |
260 | __be16 port, u32 flags) | |
5cfccc5a PS |
261 | { |
262 | struct vxlan_sock *vs; | |
263 | ||
ac5132d1 | 264 | vs = vxlan_find_sock(net, family, port, flags); |
5cfccc5a PS |
265 | if (!vs) |
266 | return NULL; | |
267 | ||
49f810f0 | 268 | return vxlan_vs_find_vni(vs, ifindex, vni); |
5cfccc5a PS |
269 | } |
270 | ||
d342894c | 271 | /* Fill in neighbour message in skbuff. */ |
272 | static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan, | |
234f5b73 SH |
273 | const struct vxlan_fdb *fdb, |
274 | u32 portid, u32 seq, int type, unsigned int flags, | |
275 | const struct vxlan_rdst *rdst) | |
d342894c | 276 | { |
277 | unsigned long now = jiffies; | |
278 | struct nda_cacheinfo ci; | |
279 | struct nlmsghdr *nlh; | |
280 | struct ndmsg *ndm; | |
e4f67add | 281 | bool send_ip, send_eth; |
d342894c | 282 | |
283 | nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags); | |
284 | if (nlh == NULL) | |
285 | return -EMSGSIZE; | |
286 | ||
287 | ndm = nlmsg_data(nlh); | |
288 | memset(ndm, 0, sizeof(*ndm)); | |
e4f67add DS |
289 | |
290 | send_eth = send_ip = true; | |
291 | ||
292 | if (type == RTM_GETNEIGH) { | |
e4c7ed41 | 293 | send_ip = !vxlan_addr_any(&rdst->remote_ip); |
e4f67add | 294 | send_eth = !is_zero_ether_addr(fdb->eth_addr); |
8f48ba71 | 295 | ndm->ndm_family = send_ip ? rdst->remote_ip.sa.sa_family : AF_INET; |
e4f67add DS |
296 | } else |
297 | ndm->ndm_family = AF_BRIDGE; | |
d342894c | 298 | ndm->ndm_state = fdb->state; |
299 | ndm->ndm_ifindex = vxlan->dev->ifindex; | |
ae884082 | 300 | ndm->ndm_flags = fdb->flags; |
545469f7 | 301 | ndm->ndm_type = RTN_UNICAST; |
d342894c | 302 | |
193523bf | 303 | if (!net_eq(dev_net(vxlan->dev), vxlan->net) && |
4967082b | 304 | nla_put_s32(skb, NDA_LINK_NETNSID, |
38f507f1 | 305 | peernet2id(dev_net(vxlan->dev), vxlan->net))) |
193523bf ND |
306 | goto nla_put_failure; |
307 | ||
e4f67add | 308 | if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr)) |
d342894c | 309 | goto nla_put_failure; |
310 | ||
e4c7ed41 | 311 | if (send_ip && vxlan_nla_put_addr(skb, NDA_DST, &rdst->remote_ip)) |
6681712d DS |
312 | goto nla_put_failure; |
313 | ||
0dfbdf41 | 314 | if (rdst->remote_port && rdst->remote_port != vxlan->cfg.dst_port && |
6681712d DS |
315 | nla_put_be16(skb, NDA_PORT, rdst->remote_port)) |
316 | goto nla_put_failure; | |
c7995c43 | 317 | if (rdst->remote_vni != vxlan->default_dst.remote_vni && |
54bfd872 | 318 | nla_put_u32(skb, NDA_VNI, be32_to_cpu(rdst->remote_vni))) |
6681712d | 319 | goto nla_put_failure; |
dc5321d7 | 320 | if ((vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) && fdb->vni && |
3ad7a4b1 RP |
321 | nla_put_u32(skb, NDA_SRC_VNI, |
322 | be32_to_cpu(fdb->vni))) | |
323 | goto nla_put_failure; | |
6681712d DS |
324 | if (rdst->remote_ifindex && |
325 | nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex)) | |
d342894c | 326 | goto nla_put_failure; |
327 | ||
328 | ci.ndm_used = jiffies_to_clock_t(now - fdb->used); | |
329 | ci.ndm_confirmed = 0; | |
330 | ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated); | |
331 | ci.ndm_refcnt = 0; | |
332 | ||
333 | if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci)) | |
334 | goto nla_put_failure; | |
335 | ||
053c095a JB |
336 | nlmsg_end(skb, nlh); |
337 | return 0; | |
d342894c | 338 | |
339 | nla_put_failure: | |
340 | nlmsg_cancel(skb, nlh); | |
341 | return -EMSGSIZE; | |
342 | } | |
343 | ||
344 | static inline size_t vxlan_nlmsg_size(void) | |
345 | { | |
346 | return NLMSG_ALIGN(sizeof(struct ndmsg)) | |
347 | + nla_total_size(ETH_ALEN) /* NDA_LLADDR */ | |
e4c7ed41 | 348 | + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */ |
73cf3317 | 349 | + nla_total_size(sizeof(__be16)) /* NDA_PORT */ |
6681712d DS |
350 | + nla_total_size(sizeof(__be32)) /* NDA_VNI */ |
351 | + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */ | |
4967082b | 352 | + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */ |
d342894c | 353 | + nla_total_size(sizeof(struct nda_cacheinfo)); |
354 | } | |
355 | ||
9e4b93f9 ND |
356 | static void vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb, |
357 | struct vxlan_rdst *rd, int type) | |
d342894c | 358 | { |
359 | struct net *net = dev_net(vxlan->dev); | |
360 | struct sk_buff *skb; | |
361 | int err = -ENOBUFS; | |
362 | ||
363 | skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC); | |
364 | if (skb == NULL) | |
365 | goto errout; | |
366 | ||
9e4b93f9 | 367 | err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd); |
d342894c | 368 | if (err < 0) { |
369 | /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */ | |
370 | WARN_ON(err == -EMSGSIZE); | |
371 | kfree_skb(skb); | |
372 | goto errout; | |
373 | } | |
374 | ||
375 | rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC); | |
376 | return; | |
377 | errout: | |
378 | if (err < 0) | |
379 | rtnl_set_sk_err(net, RTNLGRP_NEIGH, err); | |
380 | } | |
381 | ||
e4c7ed41 | 382 | static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa) |
e4f67add DS |
383 | { |
384 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
bb3fd687 SH |
385 | struct vxlan_fdb f = { |
386 | .state = NUD_STALE, | |
387 | }; | |
388 | struct vxlan_rdst remote = { | |
e4c7ed41 | 389 | .remote_ip = *ipa, /* goes to NDA_DST */ |
54bfd872 | 390 | .remote_vni = cpu_to_be32(VXLAN_N_VID), |
bb3fd687 | 391 | }; |
3e61aa8f | 392 | |
9e4b93f9 | 393 | vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH); |
e4f67add DS |
394 | } |
395 | ||
396 | static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN]) | |
397 | { | |
bb3fd687 SH |
398 | struct vxlan_fdb f = { |
399 | .state = NUD_STALE, | |
400 | }; | |
9e4b93f9 | 401 | struct vxlan_rdst remote = { }; |
e4f67add | 402 | |
e4f67add DS |
403 | memcpy(f.eth_addr, eth_addr, ETH_ALEN); |
404 | ||
9e4b93f9 | 405 | vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH); |
e4f67add DS |
406 | } |
407 | ||
d342894c | 408 | /* Hash Ethernet address */ |
409 | static u32 eth_hash(const unsigned char *addr) | |
410 | { | |
411 | u64 value = get_unaligned((u64 *)addr); | |
412 | ||
413 | /* only want 6 bytes */ | |
414 | #ifdef __BIG_ENDIAN | |
d342894c | 415 | value >>= 16; |
321fb991 | 416 | #else |
417 | value <<= 16; | |
d342894c | 418 | #endif |
419 | return hash_64(value, FDB_HASH_BITS); | |
420 | } | |
421 | ||
3ad7a4b1 RP |
422 | static u32 eth_vni_hash(const unsigned char *addr, __be32 vni) |
423 | { | |
424 | /* use 1 byte of OUI and 3 bytes of NIC */ | |
425 | u32 key = get_unaligned((u32 *)(addr + 2)); | |
426 | ||
427 | return jhash_2words(key, vni, vxlan_salt) & (FDB_HASH_SIZE - 1); | |
428 | } | |
429 | ||
d342894c | 430 | /* Hash chain to use given mac address */ |
431 | static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan, | |
3ad7a4b1 | 432 | const u8 *mac, __be32 vni) |
d342894c | 433 | { |
dc5321d7 | 434 | if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) |
3ad7a4b1 RP |
435 | return &vxlan->fdb_head[eth_vni_hash(mac, vni)]; |
436 | else | |
437 | return &vxlan->fdb_head[eth_hash(mac)]; | |
d342894c | 438 | } |
439 | ||
440 | /* Look up Ethernet address in forwarding table */ | |
014be2c8 | 441 | static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan, |
3ad7a4b1 | 442 | const u8 *mac, __be32 vni) |
d342894c | 443 | { |
3ad7a4b1 | 444 | struct hlist_head *head = vxlan_fdb_head(vxlan, mac, vni); |
d342894c | 445 | struct vxlan_fdb *f; |
d342894c | 446 | |
b67bfe0d | 447 | hlist_for_each_entry_rcu(f, head, hlist) { |
3ad7a4b1 | 448 | if (ether_addr_equal(mac, f->eth_addr)) { |
dc5321d7 | 449 | if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) { |
3ad7a4b1 RP |
450 | if (vni == f->vni) |
451 | return f; | |
452 | } else { | |
453 | return f; | |
454 | } | |
455 | } | |
d342894c | 456 | } |
457 | ||
458 | return NULL; | |
459 | } | |
460 | ||
014be2c8 | 461 | static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan, |
3ad7a4b1 | 462 | const u8 *mac, __be32 vni) |
014be2c8 SS |
463 | { |
464 | struct vxlan_fdb *f; | |
465 | ||
3ad7a4b1 | 466 | f = __vxlan_find_mac(vxlan, mac, vni); |
014be2c8 SS |
467 | if (f) |
468 | f->used = jiffies; | |
469 | ||
470 | return f; | |
471 | } | |
472 | ||
a5e7c10a MR |
473 | /* caller should hold vxlan->hash_lock */ |
474 | static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f, | |
e4c7ed41 | 475 | union vxlan_addr *ip, __be16 port, |
54bfd872 | 476 | __be32 vni, __u32 ifindex) |
6681712d | 477 | { |
3e61aa8f | 478 | struct vxlan_rdst *rd; |
6681712d | 479 | |
3e61aa8f | 480 | list_for_each_entry(rd, &f->remotes, list) { |
e4c7ed41 | 481 | if (vxlan_addr_equal(&rd->remote_ip, ip) && |
6681712d DS |
482 | rd->remote_port == port && |
483 | rd->remote_vni == vni && | |
484 | rd->remote_ifindex == ifindex) | |
a5e7c10a | 485 | return rd; |
6681712d | 486 | } |
3e61aa8f | 487 | |
a5e7c10a MR |
488 | return NULL; |
489 | } | |
490 | ||
906dc186 TR |
491 | /* Replace destination of unicast mac */ |
492 | static int vxlan_fdb_replace(struct vxlan_fdb *f, | |
54bfd872 JB |
493 | union vxlan_addr *ip, __be16 port, __be32 vni, |
494 | __u32 ifindex) | |
906dc186 TR |
495 | { |
496 | struct vxlan_rdst *rd; | |
497 | ||
498 | rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex); | |
499 | if (rd) | |
500 | return 0; | |
501 | ||
502 | rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list); | |
503 | if (!rd) | |
504 | return 0; | |
0c1d70af PA |
505 | |
506 | dst_cache_reset(&rd->dst_cache); | |
e4c7ed41 | 507 | rd->remote_ip = *ip; |
906dc186 TR |
508 | rd->remote_port = port; |
509 | rd->remote_vni = vni; | |
510 | rd->remote_ifindex = ifindex; | |
511 | return 1; | |
512 | } | |
513 | ||
a5e7c10a MR |
514 | /* Add/update destinations for multicast */ |
515 | static int vxlan_fdb_append(struct vxlan_fdb *f, | |
54bfd872 | 516 | union vxlan_addr *ip, __be16 port, __be32 vni, |
9e4b93f9 | 517 | __u32 ifindex, struct vxlan_rdst **rdp) |
a5e7c10a MR |
518 | { |
519 | struct vxlan_rdst *rd; | |
520 | ||
521 | rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex); | |
522 | if (rd) | |
523 | return 0; | |
524 | ||
6681712d DS |
525 | rd = kmalloc(sizeof(*rd), GFP_ATOMIC); |
526 | if (rd == NULL) | |
527 | return -ENOBUFS; | |
0c1d70af PA |
528 | |
529 | if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) { | |
530 | kfree(rd); | |
531 | return -ENOBUFS; | |
532 | } | |
533 | ||
e4c7ed41 | 534 | rd->remote_ip = *ip; |
6681712d DS |
535 | rd->remote_port = port; |
536 | rd->remote_vni = vni; | |
537 | rd->remote_ifindex = ifindex; | |
3e61aa8f SH |
538 | |
539 | list_add_tail_rcu(&rd->list, &f->remotes); | |
540 | ||
9e4b93f9 | 541 | *rdp = rd; |
6681712d DS |
542 | return 1; |
543 | } | |
544 | ||
dfd8645e TH |
545 | static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb, |
546 | unsigned int off, | |
547 | struct vxlanhdr *vh, size_t hdrlen, | |
54bfd872 JB |
548 | __be32 vni_field, |
549 | struct gro_remcsum *grc, | |
0ace2ca8 | 550 | bool nopartial) |
dfd8645e | 551 | { |
b7fe10e5 | 552 | size_t start, offset; |
dfd8645e TH |
553 | |
554 | if (skb->remcsum_offload) | |
b7fe10e5 | 555 | return vh; |
dfd8645e TH |
556 | |
557 | if (!NAPI_GRO_CB(skb)->csum_valid) | |
558 | return NULL; | |
559 | ||
54bfd872 JB |
560 | start = vxlan_rco_start(vni_field); |
561 | offset = start + vxlan_rco_offset(vni_field); | |
dfd8645e | 562 | |
b7fe10e5 TH |
563 | vh = skb_gro_remcsum_process(skb, (void *)vh, off, hdrlen, |
564 | start, offset, grc, nopartial); | |
dfd8645e TH |
565 | |
566 | skb->remcsum_offload = 1; | |
567 | ||
568 | return vh; | |
569 | } | |
570 | ||
5602c48c TH |
571 | static struct sk_buff **vxlan_gro_receive(struct sock *sk, |
572 | struct sk_buff **head, | |
573 | struct sk_buff *skb) | |
dc01e7d3 OG |
574 | { |
575 | struct sk_buff *p, **pp = NULL; | |
576 | struct vxlanhdr *vh, *vh2; | |
9b174d88 | 577 | unsigned int hlen, off_vx; |
dc01e7d3 | 578 | int flush = 1; |
5602c48c | 579 | struct vxlan_sock *vs = rcu_dereference_sk_user_data(sk); |
54bfd872 | 580 | __be32 flags; |
26c4f7da TH |
581 | struct gro_remcsum grc; |
582 | ||
583 | skb_gro_remcsum_init(&grc); | |
dc01e7d3 OG |
584 | |
585 | off_vx = skb_gro_offset(skb); | |
586 | hlen = off_vx + sizeof(*vh); | |
587 | vh = skb_gro_header_fast(skb, off_vx); | |
588 | if (skb_gro_header_hard(skb, hlen)) { | |
589 | vh = skb_gro_header_slow(skb, hlen, off_vx); | |
590 | if (unlikely(!vh)) | |
591 | goto out; | |
592 | } | |
dc01e7d3 | 593 | |
dfd8645e TH |
594 | skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr)); |
595 | ||
54bfd872 | 596 | flags = vh->vx_flags; |
dfd8645e TH |
597 | |
598 | if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) { | |
599 | vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr), | |
54bfd872 | 600 | vh->vx_vni, &grc, |
0ace2ca8 TH |
601 | !!(vs->flags & |
602 | VXLAN_F_REMCSUM_NOPARTIAL)); | |
dfd8645e TH |
603 | |
604 | if (!vh) | |
605 | goto out; | |
606 | } | |
607 | ||
b7fe10e5 TH |
608 | skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */ |
609 | ||
dc01e7d3 OG |
610 | for (p = *head; p; p = p->next) { |
611 | if (!NAPI_GRO_CB(p)->same_flow) | |
612 | continue; | |
613 | ||
614 | vh2 = (struct vxlanhdr *)(p->data + off_vx); | |
3511494c TG |
615 | if (vh->vx_flags != vh2->vx_flags || |
616 | vh->vx_vni != vh2->vx_vni) { | |
dc01e7d3 OG |
617 | NAPI_GRO_CB(p)->same_flow = 0; |
618 | continue; | |
619 | } | |
dc01e7d3 OG |
620 | } |
621 | ||
fcd91dd4 | 622 | pp = call_gro_receive(eth_gro_receive, head, skb); |
c194cf93 | 623 | flush = 0; |
dc01e7d3 | 624 | |
dc01e7d3 | 625 | out: |
26c4f7da | 626 | skb_gro_remcsum_cleanup(skb, &grc); |
be73b304 | 627 | skb->remcsum_offload = 0; |
dc01e7d3 OG |
628 | NAPI_GRO_CB(skb)->flush |= flush; |
629 | ||
630 | return pp; | |
631 | } | |
632 | ||
5602c48c | 633 | static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff) |
dc01e7d3 | 634 | { |
229740c6 JR |
635 | /* Sets 'skb->inner_mac_header' since we are always called with |
636 | * 'skb->encapsulation' set. | |
637 | */ | |
9b174d88 | 638 | return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr)); |
dc01e7d3 OG |
639 | } |
640 | ||
d342894c | 641 | /* Add new entry to forwarding table -- assumes lock held */ |
642 | static int vxlan_fdb_create(struct vxlan_dev *vxlan, | |
e4c7ed41 | 643 | const u8 *mac, union vxlan_addr *ip, |
6681712d | 644 | __u16 state, __u16 flags, |
3ad7a4b1 RP |
645 | __be16 port, __be32 src_vni, __be32 vni, |
646 | __u32 ifindex, __u8 ndm_flags) | |
d342894c | 647 | { |
9e4b93f9 | 648 | struct vxlan_rdst *rd = NULL; |
d342894c | 649 | struct vxlan_fdb *f; |
650 | int notify = 0; | |
17b46365 | 651 | int rc; |
d342894c | 652 | |
3ad7a4b1 | 653 | f = __vxlan_find_mac(vxlan, mac, src_vni); |
d342894c | 654 | if (f) { |
655 | if (flags & NLM_F_EXCL) { | |
656 | netdev_dbg(vxlan->dev, | |
657 | "lost race to create %pM\n", mac); | |
658 | return -EEXIST; | |
659 | } | |
660 | if (f->state != state) { | |
661 | f->state = state; | |
662 | f->updated = jiffies; | |
663 | notify = 1; | |
664 | } | |
ae884082 DS |
665 | if (f->flags != ndm_flags) { |
666 | f->flags = ndm_flags; | |
667 | f->updated = jiffies; | |
668 | notify = 1; | |
669 | } | |
906dc186 TR |
670 | if ((flags & NLM_F_REPLACE)) { |
671 | /* Only change unicasts */ | |
672 | if (!(is_multicast_ether_addr(f->eth_addr) || | |
673 | is_zero_ether_addr(f->eth_addr))) { | |
60840429 | 674 | notify |= vxlan_fdb_replace(f, ip, port, vni, |
906dc186 | 675 | ifindex); |
906dc186 TR |
676 | } else |
677 | return -EOPNOTSUPP; | |
678 | } | |
6681712d | 679 | if ((flags & NLM_F_APPEND) && |
58e4c767 MR |
680 | (is_multicast_ether_addr(f->eth_addr) || |
681 | is_zero_ether_addr(f->eth_addr))) { | |
17b46365 | 682 | rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd); |
6681712d DS |
683 | |
684 | if (rc < 0) | |
685 | return rc; | |
686 | notify |= rc; | |
687 | } | |
d342894c | 688 | } else { |
689 | if (!(flags & NLM_F_CREATE)) | |
690 | return -ENOENT; | |
691 | ||
0dfbdf41 TG |
692 | if (vxlan->cfg.addrmax && |
693 | vxlan->addrcnt >= vxlan->cfg.addrmax) | |
d342894c | 694 | return -ENOSPC; |
695 | ||
906dc186 TR |
696 | /* Disallow replace to add a multicast entry */ |
697 | if ((flags & NLM_F_REPLACE) && | |
698 | (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac))) | |
699 | return -EOPNOTSUPP; | |
700 | ||
e4c7ed41 | 701 | netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip); |
d342894c | 702 | f = kmalloc(sizeof(*f), GFP_ATOMIC); |
703 | if (!f) | |
704 | return -ENOMEM; | |
705 | ||
706 | notify = 1; | |
d342894c | 707 | f->state = state; |
ae884082 | 708 | f->flags = ndm_flags; |
d342894c | 709 | f->updated = f->used = jiffies; |
3ad7a4b1 | 710 | f->vni = src_vni; |
3e61aa8f | 711 | INIT_LIST_HEAD(&f->remotes); |
d342894c | 712 | memcpy(f->eth_addr, mac, ETH_ALEN); |
713 | ||
17b46365 HY |
714 | rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd); |
715 | if (rc < 0) { | |
716 | kfree(f); | |
717 | return rc; | |
718 | } | |
3e61aa8f | 719 | |
d342894c | 720 | ++vxlan->addrcnt; |
721 | hlist_add_head_rcu(&f->hlist, | |
3ad7a4b1 | 722 | vxlan_fdb_head(vxlan, mac, src_vni)); |
d342894c | 723 | } |
724 | ||
9e4b93f9 ND |
725 | if (notify) { |
726 | if (rd == NULL) | |
727 | rd = first_remote_rtnl(f); | |
728 | vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH); | |
729 | } | |
d342894c | 730 | |
731 | return 0; | |
732 | } | |
733 | ||
6706c82e | 734 | static void vxlan_fdb_free(struct rcu_head *head) |
6681712d DS |
735 | { |
736 | struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu); | |
3e61aa8f | 737 | struct vxlan_rdst *rd, *nd; |
6681712d | 738 | |
0c1d70af PA |
739 | list_for_each_entry_safe(rd, nd, &f->remotes, list) { |
740 | dst_cache_destroy(&rd->dst_cache); | |
6681712d | 741 | kfree(rd); |
0c1d70af | 742 | } |
6681712d DS |
743 | kfree(f); |
744 | } | |
745 | ||
d342894c | 746 | static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f) |
747 | { | |
748 | netdev_dbg(vxlan->dev, | |
749 | "delete %pM\n", f->eth_addr); | |
750 | ||
751 | --vxlan->addrcnt; | |
9e4b93f9 | 752 | vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_DELNEIGH); |
d342894c | 753 | |
754 | hlist_del_rcu(&f->hlist); | |
6681712d | 755 | call_rcu(&f->rcu, vxlan_fdb_free); |
d342894c | 756 | } |
757 | ||
35cf2845 LR |
758 | static void vxlan_dst_free(struct rcu_head *head) |
759 | { | |
760 | struct vxlan_rdst *rd = container_of(head, struct vxlan_rdst, rcu); | |
761 | ||
762 | dst_cache_destroy(&rd->dst_cache); | |
763 | kfree(rd); | |
764 | } | |
765 | ||
766 | static void vxlan_fdb_dst_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f, | |
767 | struct vxlan_rdst *rd) | |
768 | { | |
769 | list_del_rcu(&rd->list); | |
770 | vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH); | |
771 | call_rcu(&rd->rcu, vxlan_dst_free); | |
772 | } | |
773 | ||
f0b074be | 774 | static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan, |
3ad7a4b1 RP |
775 | union vxlan_addr *ip, __be16 *port, __be32 *src_vni, |
776 | __be32 *vni, u32 *ifindex) | |
d342894c | 777 | { |
6681712d | 778 | struct net *net = dev_net(vxlan->dev); |
e4c7ed41 | 779 | int err; |
d342894c | 780 | |
f0b074be | 781 | if (tb[NDA_DST]) { |
e4c7ed41 CW |
782 | err = vxlan_nla_get_addr(ip, tb[NDA_DST]); |
783 | if (err) | |
784 | return err; | |
f0b074be | 785 | } else { |
e4c7ed41 CW |
786 | union vxlan_addr *remote = &vxlan->default_dst.remote_ip; |
787 | if (remote->sa.sa_family == AF_INET) { | |
788 | ip->sin.sin_addr.s_addr = htonl(INADDR_ANY); | |
789 | ip->sa.sa_family = AF_INET; | |
790 | #if IS_ENABLED(CONFIG_IPV6) | |
791 | } else { | |
792 | ip->sin6.sin6_addr = in6addr_any; | |
793 | ip->sa.sa_family = AF_INET6; | |
794 | #endif | |
795 | } | |
f0b074be | 796 | } |
d342894c | 797 | |
6681712d | 798 | if (tb[NDA_PORT]) { |
73cf3317 | 799 | if (nla_len(tb[NDA_PORT]) != sizeof(__be16)) |
6681712d | 800 | return -EINVAL; |
f0b074be MR |
801 | *port = nla_get_be16(tb[NDA_PORT]); |
802 | } else { | |
0dfbdf41 | 803 | *port = vxlan->cfg.dst_port; |
f0b074be | 804 | } |
6681712d DS |
805 | |
806 | if (tb[NDA_VNI]) { | |
807 | if (nla_len(tb[NDA_VNI]) != sizeof(u32)) | |
808 | return -EINVAL; | |
54bfd872 | 809 | *vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI])); |
f0b074be MR |
810 | } else { |
811 | *vni = vxlan->default_dst.remote_vni; | |
812 | } | |
6681712d | 813 | |
3ad7a4b1 RP |
814 | if (tb[NDA_SRC_VNI]) { |
815 | if (nla_len(tb[NDA_SRC_VNI]) != sizeof(u32)) | |
816 | return -EINVAL; | |
817 | *src_vni = cpu_to_be32(nla_get_u32(tb[NDA_SRC_VNI])); | |
818 | } else { | |
819 | *src_vni = vxlan->default_dst.remote_vni; | |
820 | } | |
821 | ||
6681712d | 822 | if (tb[NDA_IFINDEX]) { |
5abb0029 | 823 | struct net_device *tdev; |
6681712d DS |
824 | |
825 | if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32)) | |
826 | return -EINVAL; | |
f0b074be | 827 | *ifindex = nla_get_u32(tb[NDA_IFINDEX]); |
73763949 | 828 | tdev = __dev_get_by_index(net, *ifindex); |
5abb0029 | 829 | if (!tdev) |
6681712d | 830 | return -EADDRNOTAVAIL; |
f0b074be MR |
831 | } else { |
832 | *ifindex = 0; | |
833 | } | |
834 | ||
835 | return 0; | |
836 | } | |
837 | ||
838 | /* Add static entry (via netlink) */ | |
839 | static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[], | |
840 | struct net_device *dev, | |
f6f6424b | 841 | const unsigned char *addr, u16 vid, u16 flags) |
f0b074be MR |
842 | { |
843 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
844 | /* struct net *net = dev_net(vxlan->dev); */ | |
e4c7ed41 | 845 | union vxlan_addr ip; |
f0b074be | 846 | __be16 port; |
3ad7a4b1 | 847 | __be32 src_vni, vni; |
54bfd872 | 848 | u32 ifindex; |
f0b074be MR |
849 | int err; |
850 | ||
851 | if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) { | |
852 | pr_info("RTM_NEWNEIGH with invalid state %#x\n", | |
853 | ndm->ndm_state); | |
854 | return -EINVAL; | |
855 | } | |
856 | ||
857 | if (tb[NDA_DST] == NULL) | |
858 | return -EINVAL; | |
859 | ||
3ad7a4b1 | 860 | err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &src_vni, &vni, &ifindex); |
f0b074be MR |
861 | if (err) |
862 | return err; | |
6681712d | 863 | |
5933a7bb MR |
864 | if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family) |
865 | return -EAFNOSUPPORT; | |
866 | ||
d342894c | 867 | spin_lock_bh(&vxlan->hash_lock); |
e4c7ed41 | 868 | err = vxlan_fdb_create(vxlan, addr, &ip, ndm->ndm_state, flags, |
3ad7a4b1 | 869 | port, src_vni, vni, ifindex, ndm->ndm_flags); |
d342894c | 870 | spin_unlock_bh(&vxlan->hash_lock); |
871 | ||
872 | return err; | |
873 | } | |
874 | ||
3ad7a4b1 RP |
875 | static int __vxlan_fdb_delete(struct vxlan_dev *vxlan, |
876 | const unsigned char *addr, union vxlan_addr ip, | |
877 | __be16 port, __be32 src_vni, u32 vni, u32 ifindex, | |
878 | u16 vid) | |
d342894c | 879 | { |
d342894c | 880 | struct vxlan_fdb *f; |
bc7892ba | 881 | struct vxlan_rdst *rd = NULL; |
3ad7a4b1 | 882 | int err = -ENOENT; |
bc7892ba | 883 | |
3ad7a4b1 | 884 | f = vxlan_find_mac(vxlan, addr, src_vni); |
bc7892ba | 885 | if (!f) |
3ad7a4b1 | 886 | return err; |
bc7892ba | 887 | |
e4c7ed41 CW |
888 | if (!vxlan_addr_any(&ip)) { |
889 | rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex); | |
bc7892ba MR |
890 | if (!rd) |
891 | goto out; | |
892 | } | |
893 | ||
bc7892ba MR |
894 | /* remove a destination if it's not the only one on the list, |
895 | * otherwise destroy the fdb entry | |
896 | */ | |
897 | if (rd && !list_is_singular(&f->remotes)) { | |
35cf2845 | 898 | vxlan_fdb_dst_destroy(vxlan, f, rd); |
bc7892ba | 899 | goto out; |
d342894c | 900 | } |
bc7892ba MR |
901 | |
902 | vxlan_fdb_destroy(vxlan, f); | |
903 | ||
904 | out: | |
3ad7a4b1 RP |
905 | return 0; |
906 | } | |
907 | ||
908 | /* Delete entry (via netlink) */ | |
909 | static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[], | |
910 | struct net_device *dev, | |
911 | const unsigned char *addr, u16 vid) | |
912 | { | |
913 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
914 | union vxlan_addr ip; | |
915 | __be32 src_vni, vni; | |
916 | __be16 port; | |
917 | u32 ifindex; | |
918 | int err; | |
919 | ||
920 | err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &src_vni, &vni, &ifindex); | |
921 | if (err) | |
922 | return err; | |
923 | ||
924 | spin_lock_bh(&vxlan->hash_lock); | |
925 | err = __vxlan_fdb_delete(vxlan, addr, ip, port, src_vni, vni, ifindex, | |
926 | vid); | |
d342894c | 927 | spin_unlock_bh(&vxlan->hash_lock); |
928 | ||
929 | return err; | |
930 | } | |
931 | ||
932 | /* Dump forwarding table */ | |
933 | static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb, | |
5d5eacb3 | 934 | struct net_device *dev, |
d297653d | 935 | struct net_device *filter_dev, int *idx) |
d342894c | 936 | { |
937 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
938 | unsigned int h; | |
d297653d | 939 | int err = 0; |
d342894c | 940 | |
941 | for (h = 0; h < FDB_HASH_SIZE; ++h) { | |
942 | struct vxlan_fdb *f; | |
d342894c | 943 | |
b67bfe0d | 944 | hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) { |
6681712d | 945 | struct vxlan_rdst *rd; |
6681712d | 946 | |
3e61aa8f | 947 | list_for_each_entry_rcu(rd, &f->remotes, list) { |
d297653d | 948 | if (*idx < cb->args[2]) |
07a51cd3 AW |
949 | goto skip; |
950 | ||
6681712d DS |
951 | err = vxlan_fdb_info(skb, vxlan, f, |
952 | NETLINK_CB(cb->skb).portid, | |
953 | cb->nlh->nlmsg_seq, | |
954 | RTM_NEWNEIGH, | |
955 | NLM_F_MULTI, rd); | |
d297653d | 956 | if (err < 0) |
3e61aa8f | 957 | goto out; |
3e61aa8f | 958 | skip: |
d297653d | 959 | *idx += 1; |
07a51cd3 | 960 | } |
d342894c | 961 | } |
962 | } | |
3e61aa8f | 963 | out: |
d297653d | 964 | return err; |
d342894c | 965 | } |
966 | ||
967 | /* Watch incoming packets to learn mapping between Ethernet address | |
968 | * and Tunnel endpoint. | |
c4b49512 | 969 | * Return true if packet is bogus and should be dropped. |
d342894c | 970 | */ |
26a41ae6 | 971 | static bool vxlan_snoop(struct net_device *dev, |
3ad7a4b1 | 972 | union vxlan_addr *src_ip, const u8 *src_mac, |
87613de9 | 973 | u32 src_ifindex, __be32 vni) |
d342894c | 974 | { |
975 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
976 | struct vxlan_fdb *f; | |
87613de9 MS |
977 | u32 ifindex = 0; |
978 | ||
979 | #if IS_ENABLED(CONFIG_IPV6) | |
980 | if (src_ip->sa.sa_family == AF_INET6 && | |
981 | (ipv6_addr_type(&src_ip->sin6.sin6_addr) & IPV6_ADDR_LINKLOCAL)) | |
982 | ifindex = src_ifindex; | |
983 | #endif | |
d342894c | 984 | |
3ad7a4b1 | 985 | f = vxlan_find_mac(vxlan, src_mac, vni); |
d342894c | 986 | if (likely(f)) { |
5ca5461c | 987 | struct vxlan_rdst *rdst = first_remote_rcu(f); |
3e61aa8f | 988 | |
87613de9 MS |
989 | if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip) && |
990 | rdst->remote_ifindex == ifindex)) | |
26a41ae6 | 991 | return false; |
992 | ||
993 | /* Don't migrate static entries, drop packets */ | |
e0090a9e | 994 | if (f->state & (NUD_PERMANENT | NUD_NOARP)) |
26a41ae6 | 995 | return true; |
d342894c | 996 | |
997 | if (net_ratelimit()) | |
998 | netdev_info(dev, | |
e4c7ed41 | 999 | "%pM migrated from %pIS to %pIS\n", |
a4870f79 | 1000 | src_mac, &rdst->remote_ip.sa, &src_ip->sa); |
d342894c | 1001 | |
e4c7ed41 | 1002 | rdst->remote_ip = *src_ip; |
d342894c | 1003 | f->updated = jiffies; |
9e4b93f9 | 1004 | vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH); |
d342894c | 1005 | } else { |
1006 | /* learned new entry */ | |
1007 | spin_lock(&vxlan->hash_lock); | |
3bf74b1a | 1008 | |
1009 | /* close off race between vxlan_flush and incoming packets */ | |
1010 | if (netif_running(dev)) | |
1011 | vxlan_fdb_create(vxlan, src_mac, src_ip, | |
1012 | NUD_REACHABLE, | |
1013 | NLM_F_EXCL|NLM_F_CREATE, | |
0dfbdf41 | 1014 | vxlan->cfg.dst_port, |
3ad7a4b1 | 1015 | vni, |
3bf74b1a | 1016 | vxlan->default_dst.remote_vni, |
87613de9 | 1017 | ifindex, NTF_SELF); |
d342894c | 1018 | spin_unlock(&vxlan->hash_lock); |
1019 | } | |
26a41ae6 | 1020 | |
1021 | return false; | |
d342894c | 1022 | } |
1023 | ||
d342894c | 1024 | /* See if multicast group is already in use by other ID */ |
95ab0991 | 1025 | static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev) |
d342894c | 1026 | { |
553675fb | 1027 | struct vxlan_dev *vxlan; |
c6fcc4fc | 1028 | struct vxlan_sock *sock4; |
4053ab1b AB |
1029 | #if IS_ENABLED(CONFIG_IPV6) |
1030 | struct vxlan_sock *sock6; | |
1031 | #endif | |
b1be00a6 | 1032 | unsigned short family = dev->default_dst.remote_ip.sa.sa_family; |
d342894c | 1033 | |
c6fcc4fc | 1034 | sock4 = rtnl_dereference(dev->vn4_sock); |
1035 | ||
95ab0991 G |
1036 | /* The vxlan_sock is only used by dev, leaving group has |
1037 | * no effect on other vxlan devices. | |
1038 | */ | |
66af846f | 1039 | if (family == AF_INET && sock4 && refcount_read(&sock4->refcnt) == 1) |
95ab0991 | 1040 | return false; |
b1be00a6 | 1041 | #if IS_ENABLED(CONFIG_IPV6) |
c6fcc4fc | 1042 | sock6 = rtnl_dereference(dev->vn6_sock); |
66af846f | 1043 | if (family == AF_INET6 && sock6 && refcount_read(&sock6->refcnt) == 1) |
b1be00a6 JB |
1044 | return false; |
1045 | #endif | |
95ab0991 | 1046 | |
553675fb | 1047 | list_for_each_entry(vxlan, &vn->vxlan_list, next) { |
95ab0991 | 1048 | if (!netif_running(vxlan->dev) || vxlan == dev) |
553675fb | 1049 | continue; |
d342894c | 1050 | |
c6fcc4fc | 1051 | if (family == AF_INET && |
1052 | rtnl_dereference(vxlan->vn4_sock) != sock4) | |
95ab0991 | 1053 | continue; |
b1be00a6 | 1054 | #if IS_ENABLED(CONFIG_IPV6) |
c6fcc4fc | 1055 | if (family == AF_INET6 && |
1056 | rtnl_dereference(vxlan->vn6_sock) != sock6) | |
b1be00a6 JB |
1057 | continue; |
1058 | #endif | |
95ab0991 G |
1059 | |
1060 | if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip, | |
1061 | &dev->default_dst.remote_ip)) | |
1062 | continue; | |
1063 | ||
1064 | if (vxlan->default_dst.remote_ifindex != | |
1065 | dev->default_dst.remote_ifindex) | |
1066 | continue; | |
1067 | ||
1068 | return true; | |
553675fb | 1069 | } |
d342894c | 1070 | |
1071 | return false; | |
1072 | } | |
1073 | ||
544a773a | 1074 | static bool __vxlan_sock_release_prep(struct vxlan_sock *vs) |
7c47cedf | 1075 | { |
b1be00a6 | 1076 | struct vxlan_net *vn; |
012a5729 | 1077 | |
b1be00a6 | 1078 | if (!vs) |
544a773a | 1079 | return false; |
66af846f | 1080 | if (!refcount_dec_and_test(&vs->refcnt)) |
544a773a | 1081 | return false; |
d342894c | 1082 | |
b1be00a6 | 1083 | vn = net_generic(sock_net(vs->sock->sk), vxlan_net_id); |
1c51a915 | 1084 | spin_lock(&vn->sock_lock); |
7c47cedf | 1085 | hlist_del_rcu(&vs->hlist); |
e7b3db5e | 1086 | udp_tunnel_notify_del_rx_port(vs->sock, |
b9adcd69 AD |
1087 | (vs->flags & VXLAN_F_GPE) ? |
1088 | UDP_TUNNEL_TYPE_VXLAN_GPE : | |
e7b3db5e | 1089 | UDP_TUNNEL_TYPE_VXLAN); |
1c51a915 SH |
1090 | spin_unlock(&vn->sock_lock); |
1091 | ||
544a773a | 1092 | return true; |
d342894c | 1093 | } |
1094 | ||
b1be00a6 JB |
1095 | static void vxlan_sock_release(struct vxlan_dev *vxlan) |
1096 | { | |
c6fcc4fc | 1097 | struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock); |
b1be00a6 | 1098 | #if IS_ENABLED(CONFIG_IPV6) |
c6fcc4fc | 1099 | struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock); |
1100 | ||
57d88182 | 1101 | RCU_INIT_POINTER(vxlan->vn6_sock, NULL); |
544a773a HFS |
1102 | #endif |
1103 | ||
57d88182 | 1104 | RCU_INIT_POINTER(vxlan->vn4_sock, NULL); |
544a773a HFS |
1105 | synchronize_net(); |
1106 | ||
a53cb29b MB |
1107 | vxlan_vs_del_dev(vxlan); |
1108 | ||
c6fcc4fc | 1109 | if (__vxlan_sock_release_prep(sock4)) { |
1110 | udp_tunnel_sock_release(sock4->sock); | |
1111 | kfree(sock4); | |
544a773a HFS |
1112 | } |
1113 | ||
1114 | #if IS_ENABLED(CONFIG_IPV6) | |
c6fcc4fc | 1115 | if (__vxlan_sock_release_prep(sock6)) { |
1116 | udp_tunnel_sock_release(sock6->sock); | |
1117 | kfree(sock6); | |
544a773a | 1118 | } |
b1be00a6 JB |
1119 | #endif |
1120 | } | |
1121 | ||
56ef9c90 | 1122 | /* Update multicast group membership when first VNI on |
c4b49512 | 1123 | * multicast address is brought up |
7c47cedf | 1124 | */ |
56ef9c90 | 1125 | static int vxlan_igmp_join(struct vxlan_dev *vxlan) |
d342894c | 1126 | { |
b1be00a6 | 1127 | struct sock *sk; |
e4c7ed41 CW |
1128 | union vxlan_addr *ip = &vxlan->default_dst.remote_ip; |
1129 | int ifindex = vxlan->default_dst.remote_ifindex; | |
149d7549 | 1130 | int ret = -EINVAL; |
d342894c | 1131 | |
e4c7ed41 | 1132 | if (ip->sa.sa_family == AF_INET) { |
c6fcc4fc | 1133 | struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock); |
e4c7ed41 CW |
1134 | struct ip_mreqn mreq = { |
1135 | .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr, | |
1136 | .imr_ifindex = ifindex, | |
1137 | }; | |
1138 | ||
c6fcc4fc | 1139 | sk = sock4->sock->sk; |
b1be00a6 | 1140 | lock_sock(sk); |
56ef9c90 | 1141 | ret = ip_mc_join_group(sk, &mreq); |
b1be00a6 | 1142 | release_sock(sk); |
e4c7ed41 CW |
1143 | #if IS_ENABLED(CONFIG_IPV6) |
1144 | } else { | |
c6fcc4fc | 1145 | struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock); |
1146 | ||
1147 | sk = sock6->sock->sk; | |
b1be00a6 | 1148 | lock_sock(sk); |
56ef9c90 MRL |
1149 | ret = ipv6_stub->ipv6_sock_mc_join(sk, ifindex, |
1150 | &ip->sin6.sin6_addr); | |
b1be00a6 | 1151 | release_sock(sk); |
e4c7ed41 CW |
1152 | #endif |
1153 | } | |
3fc2de2f | 1154 | |
56ef9c90 | 1155 | return ret; |
3fc2de2f | 1156 | } |
1157 | ||
1158 | /* Inverse of vxlan_igmp_join when last VNI is brought down */ | |
56ef9c90 | 1159 | static int vxlan_igmp_leave(struct vxlan_dev *vxlan) |
3fc2de2f | 1160 | { |
b1be00a6 | 1161 | struct sock *sk; |
e4c7ed41 CW |
1162 | union vxlan_addr *ip = &vxlan->default_dst.remote_ip; |
1163 | int ifindex = vxlan->default_dst.remote_ifindex; | |
149d7549 | 1164 | int ret = -EINVAL; |
3fc2de2f | 1165 | |
e4c7ed41 | 1166 | if (ip->sa.sa_family == AF_INET) { |
c6fcc4fc | 1167 | struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock); |
e4c7ed41 CW |
1168 | struct ip_mreqn mreq = { |
1169 | .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr, | |
1170 | .imr_ifindex = ifindex, | |
1171 | }; | |
1172 | ||
c6fcc4fc | 1173 | sk = sock4->sock->sk; |
b1be00a6 | 1174 | lock_sock(sk); |
56ef9c90 | 1175 | ret = ip_mc_leave_group(sk, &mreq); |
b1be00a6 | 1176 | release_sock(sk); |
e4c7ed41 CW |
1177 | #if IS_ENABLED(CONFIG_IPV6) |
1178 | } else { | |
c6fcc4fc | 1179 | struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock); |
1180 | ||
1181 | sk = sock6->sock->sk; | |
b1be00a6 | 1182 | lock_sock(sk); |
56ef9c90 MRL |
1183 | ret = ipv6_stub->ipv6_sock_mc_drop(sk, ifindex, |
1184 | &ip->sin6.sin6_addr); | |
b1be00a6 | 1185 | release_sock(sk); |
e4c7ed41 CW |
1186 | #endif |
1187 | } | |
d342894c | 1188 | |
56ef9c90 | 1189 | return ret; |
d342894c | 1190 | } |
1191 | ||
f14ecebb JB |
1192 | static bool vxlan_remcsum(struct vxlanhdr *unparsed, |
1193 | struct sk_buff *skb, u32 vxflags) | |
dfd8645e | 1194 | { |
7d34fa75 | 1195 | size_t start, offset; |
dfd8645e | 1196 | |
f14ecebb JB |
1197 | if (!(unparsed->vx_flags & VXLAN_HF_RCO) || skb->remcsum_offload) |
1198 | goto out; | |
b7fe10e5 | 1199 | |
f14ecebb JB |
1200 | start = vxlan_rco_start(unparsed->vx_vni); |
1201 | offset = start + vxlan_rco_offset(unparsed->vx_vni); | |
dfd8645e | 1202 | |
7d34fa75 | 1203 | if (!pskb_may_pull(skb, offset + sizeof(u16))) |
be5cfeab | 1204 | return false; |
dfd8645e | 1205 | |
be5cfeab JB |
1206 | skb_remcsum_process(skb, (void *)(vxlan_hdr(skb) + 1), start, offset, |
1207 | !!(vxflags & VXLAN_F_REMCSUM_NOPARTIAL)); | |
f14ecebb JB |
1208 | out: |
1209 | unparsed->vx_flags &= ~VXLAN_HF_RCO; | |
1210 | unparsed->vx_vni &= VXLAN_VNI_MASK; | |
be5cfeab | 1211 | return true; |
dfd8645e TH |
1212 | } |
1213 | ||
f14ecebb | 1214 | static void vxlan_parse_gbp_hdr(struct vxlanhdr *unparsed, |
64f87d36 | 1215 | struct sk_buff *skb, u32 vxflags, |
10a5af23 | 1216 | struct vxlan_metadata *md) |
3288af08 | 1217 | { |
f14ecebb | 1218 | struct vxlanhdr_gbp *gbp = (struct vxlanhdr_gbp *)unparsed; |
10a5af23 | 1219 | struct metadata_dst *tun_dst; |
f14ecebb JB |
1220 | |
1221 | if (!(unparsed->vx_flags & VXLAN_HF_GBP)) | |
1222 | goto out; | |
3288af08 | 1223 | |
3288af08 JB |
1224 | md->gbp = ntohs(gbp->policy_id); |
1225 | ||
10a5af23 | 1226 | tun_dst = (struct metadata_dst *)skb_dst(skb); |
810813c4 | 1227 | if (tun_dst) { |
3288af08 | 1228 | tun_dst->u.tun_info.key.tun_flags |= TUNNEL_VXLAN_OPT; |
810813c4 DM |
1229 | tun_dst->u.tun_info.options_len = sizeof(*md); |
1230 | } | |
3288af08 JB |
1231 | if (gbp->dont_learn) |
1232 | md->gbp |= VXLAN_GBP_DONT_LEARN; | |
1233 | ||
1234 | if (gbp->policy_applied) | |
1235 | md->gbp |= VXLAN_GBP_POLICY_APPLIED; | |
f14ecebb | 1236 | |
64f87d36 JB |
1237 | /* In flow-based mode, GBP is carried in dst_metadata */ |
1238 | if (!(vxflags & VXLAN_F_COLLECT_METADATA)) | |
1239 | skb->mark = md->gbp; | |
f14ecebb JB |
1240 | out: |
1241 | unparsed->vx_flags &= ~VXLAN_GBP_USED_BITS; | |
3288af08 JB |
1242 | } |
1243 | ||
e1e5314d | 1244 | static bool vxlan_parse_gpe_hdr(struct vxlanhdr *unparsed, |
61618eea | 1245 | __be16 *protocol, |
e1e5314d JB |
1246 | struct sk_buff *skb, u32 vxflags) |
1247 | { | |
1248 | struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)unparsed; | |
1249 | ||
1250 | /* Need to have Next Protocol set for interfaces in GPE mode. */ | |
1251 | if (!gpe->np_applied) | |
1252 | return false; | |
1253 | /* "The initial version is 0. If a receiver does not support the | |
1254 | * version indicated it MUST drop the packet. | |
1255 | */ | |
1256 | if (gpe->version != 0) | |
1257 | return false; | |
1258 | /* "When the O bit is set to 1, the packet is an OAM packet and OAM | |
1259 | * processing MUST occur." However, we don't implement OAM | |
1260 | * processing, thus drop the packet. | |
1261 | */ | |
1262 | if (gpe->oam_flag) | |
1263 | return false; | |
1264 | ||
fa20e0e3 JB |
1265 | *protocol = tun_p_to_eth_p(gpe->next_protocol); |
1266 | if (!*protocol) | |
e1e5314d | 1267 | return false; |
e1e5314d JB |
1268 | |
1269 | unparsed->vx_flags &= ~VXLAN_GPE_USED_BITS; | |
1270 | return true; | |
1271 | } | |
1272 | ||
1ab016e2 JB |
1273 | static bool vxlan_set_mac(struct vxlan_dev *vxlan, |
1274 | struct vxlan_sock *vs, | |
3ad7a4b1 | 1275 | struct sk_buff *skb, __be32 vni) |
614732ea | 1276 | { |
614732ea | 1277 | union vxlan_addr saddr; |
87613de9 | 1278 | u32 ifindex = skb->dev->ifindex; |
614732ea | 1279 | |
614732ea | 1280 | skb_reset_mac_header(skb); |
614732ea TG |
1281 | skb->protocol = eth_type_trans(skb, vxlan->dev); |
1282 | skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); | |
1283 | ||
1284 | /* Ignore packet loops (and multicast echo) */ | |
1285 | if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr)) | |
1ab016e2 | 1286 | return false; |
614732ea | 1287 | |
760c6805 | 1288 | /* Get address from the outer IP header */ |
ce212d0f | 1289 | if (vxlan_get_sk_family(vs) == AF_INET) { |
1ab016e2 | 1290 | saddr.sin.sin_addr.s_addr = ip_hdr(skb)->saddr; |
614732ea TG |
1291 | saddr.sa.sa_family = AF_INET; |
1292 | #if IS_ENABLED(CONFIG_IPV6) | |
1293 | } else { | |
1ab016e2 | 1294 | saddr.sin6.sin6_addr = ipv6_hdr(skb)->saddr; |
614732ea TG |
1295 | saddr.sa.sa_family = AF_INET6; |
1296 | #endif | |
1297 | } | |
1298 | ||
dc5321d7 | 1299 | if ((vxlan->cfg.flags & VXLAN_F_LEARN) && |
87613de9 | 1300 | vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source, ifindex, vni)) |
1ab016e2 JB |
1301 | return false; |
1302 | ||
1303 | return true; | |
1304 | } | |
1305 | ||
760c6805 JB |
1306 | static bool vxlan_ecn_decapsulate(struct vxlan_sock *vs, void *oiph, |
1307 | struct sk_buff *skb) | |
1308 | { | |
1309 | int err = 0; | |
1310 | ||
1311 | if (vxlan_get_sk_family(vs) == AF_INET) | |
1312 | err = IP_ECN_decapsulate(oiph, skb); | |
1313 | #if IS_ENABLED(CONFIG_IPV6) | |
1314 | else | |
1315 | err = IP6_ECN_decapsulate(oiph, skb); | |
1316 | #endif | |
1317 | ||
1318 | if (unlikely(err) && log_ecn_error) { | |
1319 | if (vxlan_get_sk_family(vs) == AF_INET) | |
1320 | net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n", | |
1321 | &((struct iphdr *)oiph)->saddr, | |
1322 | ((struct iphdr *)oiph)->tos); | |
1323 | else | |
1324 | net_info_ratelimited("non-ECT from %pI6\n", | |
1325 | &((struct ipv6hdr *)oiph)->saddr); | |
1326 | } | |
1327 | return err <= 1; | |
1328 | } | |
1329 | ||
d342894c | 1330 | /* Callback from net/ipv4/udp.c to receive packets */ |
f2d1968e | 1331 | static int vxlan_rcv(struct sock *sk, struct sk_buff *skb) |
d342894c | 1332 | { |
f2d1968e | 1333 | struct pcpu_sw_netstats *stats; |
c9e78efb | 1334 | struct vxlan_dev *vxlan; |
5cfccc5a | 1335 | struct vxlan_sock *vs; |
f14ecebb | 1336 | struct vxlanhdr unparsed; |
ee122c79 TG |
1337 | struct vxlan_metadata _md; |
1338 | struct vxlan_metadata *md = &_md; | |
61618eea | 1339 | __be16 protocol = htons(ETH_P_TEB); |
e1e5314d | 1340 | bool raw_proto = false; |
f2d1968e | 1341 | void *oiph; |
3ad7a4b1 | 1342 | __be32 vni = 0; |
d342894c | 1343 | |
e1e5314d | 1344 | /* Need UDP and VXLAN header to be present */ |
7ce04758 | 1345 | if (!pskb_may_pull(skb, VXLAN_HLEN)) |
e5aed006 | 1346 | goto drop; |
d342894c | 1347 | |
f14ecebb | 1348 | unparsed = *vxlan_hdr(skb); |
288b01c8 JB |
1349 | /* VNI flag always required to be set */ |
1350 | if (!(unparsed.vx_flags & VXLAN_HF_VNI)) { | |
1351 | netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n", | |
1352 | ntohl(vxlan_hdr(skb)->vx_flags), | |
1353 | ntohl(vxlan_hdr(skb)->vx_vni)); | |
1354 | /* Return non vxlan pkt */ | |
e5aed006 | 1355 | goto drop; |
d342894c | 1356 | } |
288b01c8 JB |
1357 | unparsed.vx_flags &= ~VXLAN_HF_VNI; |
1358 | unparsed.vx_vni &= ~VXLAN_VNI_MASK; | |
d342894c | 1359 | |
559835ea | 1360 | vs = rcu_dereference_sk_user_data(sk); |
5cfccc5a | 1361 | if (!vs) |
d342894c | 1362 | goto drop; |
d342894c | 1363 | |
3ad7a4b1 RP |
1364 | vni = vxlan_vni(vxlan_hdr(skb)->vx_vni); |
1365 | ||
49f810f0 | 1366 | vxlan = vxlan_vs_find_vni(vs, skb->dev->ifindex, vni); |
c9e78efb JB |
1367 | if (!vxlan) |
1368 | goto drop; | |
1369 | ||
e1e5314d JB |
1370 | /* For backwards compatibility, only allow reserved fields to be |
1371 | * used by VXLAN extensions if explicitly requested. | |
1372 | */ | |
1373 | if (vs->flags & VXLAN_F_GPE) { | |
1374 | if (!vxlan_parse_gpe_hdr(&unparsed, &protocol, skb, vs->flags)) | |
1375 | goto drop; | |
1376 | raw_proto = true; | |
1377 | } | |
1378 | ||
1379 | if (__iptunnel_pull_header(skb, VXLAN_HLEN, protocol, raw_proto, | |
1380 | !net_eq(vxlan->net, dev_net(vxlan->dev)))) | |
1381 | goto drop; | |
c9e78efb | 1382 | |
ee122c79 | 1383 | if (vxlan_collect_metadata(vs)) { |
10a5af23 | 1384 | struct metadata_dst *tun_dst; |
07dabf20 | 1385 | |
c29a70d2 | 1386 | tun_dst = udp_tun_rx_dst(skb, vxlan_get_sk_family(vs), TUNNEL_KEY, |
d817f432 | 1387 | key32_to_tunnel_id(vni), sizeof(*md)); |
c29a70d2 | 1388 | |
ee122c79 TG |
1389 | if (!tun_dst) |
1390 | goto drop; | |
1391 | ||
0f1b7354 | 1392 | md = ip_tunnel_info_opts(&tun_dst->u.tun_info); |
10a5af23 JB |
1393 | |
1394 | skb_dst_set(skb, (struct dst_entry *)tun_dst); | |
ee122c79 TG |
1395 | } else { |
1396 | memset(md, 0, sizeof(*md)); | |
1397 | } | |
1398 | ||
f14ecebb JB |
1399 | if (vs->flags & VXLAN_F_REMCSUM_RX) |
1400 | if (!vxlan_remcsum(&unparsed, skb, vs->flags)) | |
1401 | goto drop; | |
1402 | if (vs->flags & VXLAN_F_GBP) | |
10a5af23 | 1403 | vxlan_parse_gbp_hdr(&unparsed, skb, vs->flags, md); |
e1e5314d JB |
1404 | /* Note that GBP and GPE can never be active together. This is |
1405 | * ensured in vxlan_dev_configure. | |
1406 | */ | |
3511494c | 1407 | |
f14ecebb | 1408 | if (unparsed.vx_flags || unparsed.vx_vni) { |
3bf39475 TH |
1409 | /* If there are any unprocessed flags remaining treat |
1410 | * this as a malformed packet. This behavior diverges from | |
1411 | * VXLAN RFC (RFC7348) which stipulates that bits in reserved | |
1412 | * in reserved fields are to be ignored. The approach here | |
c4b49512 | 1413 | * maintains compatibility with previous stack code, and also |
3bf39475 TH |
1414 | * is more robust and provides a little more security in |
1415 | * adding extensions to VXLAN. | |
1416 | */ | |
288b01c8 | 1417 | goto drop; |
3bf39475 TH |
1418 | } |
1419 | ||
e1e5314d | 1420 | if (!raw_proto) { |
3ad7a4b1 | 1421 | if (!vxlan_set_mac(vxlan, vs, skb, vni)) |
e1e5314d JB |
1422 | goto drop; |
1423 | } else { | |
8be0cfa4 | 1424 | skb_reset_mac_header(skb); |
e1e5314d JB |
1425 | skb->dev = vxlan->dev; |
1426 | skb->pkt_type = PACKET_HOST; | |
1427 | } | |
f2d1968e | 1428 | |
f2d1968e JB |
1429 | oiph = skb_network_header(skb); |
1430 | skb_reset_network_header(skb); | |
1431 | ||
1432 | if (!vxlan_ecn_decapsulate(vs, oiph, skb)) { | |
1433 | ++vxlan->dev->stats.rx_frame_errors; | |
1434 | ++vxlan->dev->stats.rx_errors; | |
1435 | goto drop; | |
1436 | } | |
1437 | ||
1438 | stats = this_cpu_ptr(vxlan->dev->tstats); | |
1439 | u64_stats_update_begin(&stats->syncp); | |
1440 | stats->rx_packets++; | |
1441 | stats->rx_bytes += skb->len; | |
1442 | u64_stats_update_end(&stats->syncp); | |
1443 | ||
1444 | gro_cells_receive(&vxlan->gro_cells, skb); | |
5cfccc5a PS |
1445 | return 0; |
1446 | ||
1447 | drop: | |
288b01c8 JB |
1448 | /* Consume bad packet */ |
1449 | kfree_skb(skb); | |
1450 | return 0; | |
5cfccc5a PS |
1451 | } |
1452 | ||
3ad7a4b1 | 1453 | static int arp_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni) |
e4f67add DS |
1454 | { |
1455 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
1456 | struct arphdr *parp; | |
1457 | u8 *arpptr, *sha; | |
1458 | __be32 sip, tip; | |
1459 | struct neighbour *n; | |
1460 | ||
1461 | if (dev->flags & IFF_NOARP) | |
1462 | goto out; | |
1463 | ||
1464 | if (!pskb_may_pull(skb, arp_hdr_len(dev))) { | |
1465 | dev->stats.tx_dropped++; | |
1466 | goto out; | |
1467 | } | |
1468 | parp = arp_hdr(skb); | |
1469 | ||
1470 | if ((parp->ar_hrd != htons(ARPHRD_ETHER) && | |
1471 | parp->ar_hrd != htons(ARPHRD_IEEE802)) || | |
1472 | parp->ar_pro != htons(ETH_P_IP) || | |
1473 | parp->ar_op != htons(ARPOP_REQUEST) || | |
1474 | parp->ar_hln != dev->addr_len || | |
1475 | parp->ar_pln != 4) | |
1476 | goto out; | |
1477 | arpptr = (u8 *)parp + sizeof(struct arphdr); | |
1478 | sha = arpptr; | |
1479 | arpptr += dev->addr_len; /* sha */ | |
1480 | memcpy(&sip, arpptr, sizeof(sip)); | |
1481 | arpptr += sizeof(sip); | |
1482 | arpptr += dev->addr_len; /* tha */ | |
1483 | memcpy(&tip, arpptr, sizeof(tip)); | |
1484 | ||
1485 | if (ipv4_is_loopback(tip) || | |
1486 | ipv4_is_multicast(tip)) | |
1487 | goto out; | |
1488 | ||
1489 | n = neigh_lookup(&arp_tbl, &tip, dev); | |
1490 | ||
1491 | if (n) { | |
e4f67add DS |
1492 | struct vxlan_fdb *f; |
1493 | struct sk_buff *reply; | |
1494 | ||
1495 | if (!(n->nud_state & NUD_CONNECTED)) { | |
1496 | neigh_release(n); | |
1497 | goto out; | |
1498 | } | |
1499 | ||
3ad7a4b1 | 1500 | f = vxlan_find_mac(vxlan, n->ha, vni); |
e4c7ed41 | 1501 | if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) { |
e4f67add DS |
1502 | /* bridge-local neighbor */ |
1503 | neigh_release(n); | |
1504 | goto out; | |
1505 | } | |
1506 | ||
1507 | reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha, | |
1508 | n->ha, sha); | |
1509 | ||
1510 | neigh_release(n); | |
1511 | ||
7346135d DS |
1512 | if (reply == NULL) |
1513 | goto out; | |
1514 | ||
e4f67add DS |
1515 | skb_reset_mac_header(reply); |
1516 | __skb_pull(reply, skb_network_offset(reply)); | |
1517 | reply->ip_summed = CHECKSUM_UNNECESSARY; | |
1518 | reply->pkt_type = PACKET_HOST; | |
1519 | ||
1520 | if (netif_rx_ni(reply) == NET_RX_DROP) | |
1521 | dev->stats.rx_dropped++; | |
dc5321d7 | 1522 | } else if (vxlan->cfg.flags & VXLAN_F_L3MISS) { |
e4c7ed41 CW |
1523 | union vxlan_addr ipa = { |
1524 | .sin.sin_addr.s_addr = tip, | |
a45e92a5 | 1525 | .sin.sin_family = AF_INET, |
e4c7ed41 CW |
1526 | }; |
1527 | ||
1528 | vxlan_ip_miss(dev, &ipa); | |
1529 | } | |
e4f67add DS |
1530 | out: |
1531 | consume_skb(skb); | |
1532 | return NETDEV_TX_OK; | |
1533 | } | |
1534 | ||
f564f45c | 1535 | #if IS_ENABLED(CONFIG_IPV6) |
4b29dba9 DS |
1536 | static struct sk_buff *vxlan_na_create(struct sk_buff *request, |
1537 | struct neighbour *n, bool isrouter) | |
1538 | { | |
1539 | struct net_device *dev = request->dev; | |
1540 | struct sk_buff *reply; | |
1541 | struct nd_msg *ns, *na; | |
1542 | struct ipv6hdr *pip6; | |
1543 | u8 *daddr; | |
1544 | int na_olen = 8; /* opt hdr + ETH_ALEN for target */ | |
1545 | int ns_olen; | |
1546 | int i, len; | |
1547 | ||
f1fb08f6 | 1548 | if (dev == NULL || !pskb_may_pull(request, request->len)) |
4b29dba9 DS |
1549 | return NULL; |
1550 | ||
1551 | len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) + | |
1552 | sizeof(*na) + na_olen + dev->needed_tailroom; | |
1553 | reply = alloc_skb(len, GFP_ATOMIC); | |
1554 | if (reply == NULL) | |
1555 | return NULL; | |
1556 | ||
1557 | reply->protocol = htons(ETH_P_IPV6); | |
1558 | reply->dev = dev; | |
1559 | skb_reserve(reply, LL_RESERVED_SPACE(request->dev)); | |
1560 | skb_push(reply, sizeof(struct ethhdr)); | |
6297b91c | 1561 | skb_reset_mac_header(reply); |
4b29dba9 | 1562 | |
f1fb08f6 | 1563 | ns = (struct nd_msg *)(ipv6_hdr(request) + 1); |
4b29dba9 DS |
1564 | |
1565 | daddr = eth_hdr(request)->h_source; | |
f1fb08f6 VB |
1566 | ns_olen = request->len - skb_network_offset(request) - |
1567 | sizeof(struct ipv6hdr) - sizeof(*ns); | |
4b29dba9 DS |
1568 | for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) { |
1569 | if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) { | |
1570 | daddr = ns->opt + i + sizeof(struct nd_opt_hdr); | |
1571 | break; | |
1572 | } | |
1573 | } | |
1574 | ||
1575 | /* Ethernet header */ | |
1576 | ether_addr_copy(eth_hdr(reply)->h_dest, daddr); | |
1577 | ether_addr_copy(eth_hdr(reply)->h_source, n->ha); | |
1578 | eth_hdr(reply)->h_proto = htons(ETH_P_IPV6); | |
1579 | reply->protocol = htons(ETH_P_IPV6); | |
1580 | ||
1581 | skb_pull(reply, sizeof(struct ethhdr)); | |
6297b91c | 1582 | skb_reset_network_header(reply); |
4b29dba9 DS |
1583 | skb_put(reply, sizeof(struct ipv6hdr)); |
1584 | ||
1585 | /* IPv6 header */ | |
1586 | ||
1587 | pip6 = ipv6_hdr(reply); | |
1588 | memset(pip6, 0, sizeof(struct ipv6hdr)); | |
1589 | pip6->version = 6; | |
1590 | pip6->priority = ipv6_hdr(request)->priority; | |
1591 | pip6->nexthdr = IPPROTO_ICMPV6; | |
1592 | pip6->hop_limit = 255; | |
1593 | pip6->daddr = ipv6_hdr(request)->saddr; | |
1594 | pip6->saddr = *(struct in6_addr *)n->primary_key; | |
1595 | ||
1596 | skb_pull(reply, sizeof(struct ipv6hdr)); | |
6297b91c | 1597 | skb_reset_transport_header(reply); |
4b29dba9 | 1598 | |
4b29dba9 | 1599 | /* Neighbor Advertisement */ |
b080db58 | 1600 | na = skb_put_zero(reply, sizeof(*na) + na_olen); |
4b29dba9 DS |
1601 | na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT; |
1602 | na->icmph.icmp6_router = isrouter; | |
1603 | na->icmph.icmp6_override = 1; | |
1604 | na->icmph.icmp6_solicited = 1; | |
1605 | na->target = ns->target; | |
1606 | ether_addr_copy(&na->opt[2], n->ha); | |
1607 | na->opt[0] = ND_OPT_TARGET_LL_ADDR; | |
1608 | na->opt[1] = na_olen >> 3; | |
1609 | ||
1610 | na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr, | |
1611 | &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6, | |
1612 | csum_partial(na, sizeof(*na)+na_olen, 0)); | |
1613 | ||
1614 | pip6->payload_len = htons(sizeof(*na)+na_olen); | |
1615 | ||
1616 | skb_push(reply, sizeof(struct ipv6hdr)); | |
1617 | ||
1618 | reply->ip_summed = CHECKSUM_UNNECESSARY; | |
1619 | ||
1620 | return reply; | |
1621 | } | |
1622 | ||
3ad7a4b1 | 1623 | static int neigh_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni) |
f564f45c CW |
1624 | { |
1625 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
8dcd81a9 | 1626 | const struct in6_addr *daddr; |
829af2cd | 1627 | const struct ipv6hdr *iphdr; |
4b29dba9 | 1628 | struct inet6_dev *in6_dev; |
829af2cd XL |
1629 | struct neighbour *n; |
1630 | struct nd_msg *msg; | |
f564f45c CW |
1631 | |
1632 | in6_dev = __in6_dev_get(dev); | |
1633 | if (!in6_dev) | |
1634 | goto out; | |
1635 | ||
f564f45c | 1636 | iphdr = ipv6_hdr(skb); |
f564f45c | 1637 | daddr = &iphdr->daddr; |
f1fb08f6 | 1638 | msg = (struct nd_msg *)(iphdr + 1); |
f564f45c | 1639 | |
4b29dba9 DS |
1640 | if (ipv6_addr_loopback(daddr) || |
1641 | ipv6_addr_is_multicast(&msg->target)) | |
1642 | goto out; | |
1643 | ||
1644 | n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev); | |
f564f45c CW |
1645 | |
1646 | if (n) { | |
1647 | struct vxlan_fdb *f; | |
4b29dba9 | 1648 | struct sk_buff *reply; |
f564f45c CW |
1649 | |
1650 | if (!(n->nud_state & NUD_CONNECTED)) { | |
1651 | neigh_release(n); | |
1652 | goto out; | |
1653 | } | |
1654 | ||
3ad7a4b1 | 1655 | f = vxlan_find_mac(vxlan, n->ha, vni); |
f564f45c CW |
1656 | if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) { |
1657 | /* bridge-local neighbor */ | |
1658 | neigh_release(n); | |
1659 | goto out; | |
1660 | } | |
1661 | ||
4b29dba9 DS |
1662 | reply = vxlan_na_create(skb, n, |
1663 | !!(f ? f->flags & NTF_ROUTER : 0)); | |
1664 | ||
f564f45c | 1665 | neigh_release(n); |
4b29dba9 DS |
1666 | |
1667 | if (reply == NULL) | |
1668 | goto out; | |
1669 | ||
1670 | if (netif_rx_ni(reply) == NET_RX_DROP) | |
1671 | dev->stats.rx_dropped++; | |
1672 | ||
dc5321d7 | 1673 | } else if (vxlan->cfg.flags & VXLAN_F_L3MISS) { |
4b29dba9 DS |
1674 | union vxlan_addr ipa = { |
1675 | .sin6.sin6_addr = msg->target, | |
a45e92a5 | 1676 | .sin6.sin6_family = AF_INET6, |
4b29dba9 DS |
1677 | }; |
1678 | ||
f564f45c CW |
1679 | vxlan_ip_miss(dev, &ipa); |
1680 | } | |
1681 | ||
1682 | out: | |
1683 | consume_skb(skb); | |
1684 | return NETDEV_TX_OK; | |
1685 | } | |
1686 | #endif | |
1687 | ||
e4f67add DS |
1688 | static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb) |
1689 | { | |
1690 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
1691 | struct neighbour *n; | |
e4f67add DS |
1692 | |
1693 | if (is_multicast_ether_addr(eth_hdr(skb)->h_dest)) | |
1694 | return false; | |
1695 | ||
1696 | n = NULL; | |
1697 | switch (ntohs(eth_hdr(skb)->h_proto)) { | |
1698 | case ETH_P_IP: | |
e15a00aa CW |
1699 | { |
1700 | struct iphdr *pip; | |
1701 | ||
e4f67add DS |
1702 | if (!pskb_may_pull(skb, sizeof(struct iphdr))) |
1703 | return false; | |
1704 | pip = ip_hdr(skb); | |
1705 | n = neigh_lookup(&arp_tbl, &pip->daddr, dev); | |
dc5321d7 | 1706 | if (!n && (vxlan->cfg.flags & VXLAN_F_L3MISS)) { |
e4c7ed41 CW |
1707 | union vxlan_addr ipa = { |
1708 | .sin.sin_addr.s_addr = pip->daddr, | |
a45e92a5 | 1709 | .sin.sin_family = AF_INET, |
e4c7ed41 CW |
1710 | }; |
1711 | ||
1712 | vxlan_ip_miss(dev, &ipa); | |
1713 | return false; | |
1714 | } | |
1715 | ||
e4f67add | 1716 | break; |
e15a00aa CW |
1717 | } |
1718 | #if IS_ENABLED(CONFIG_IPV6) | |
1719 | case ETH_P_IPV6: | |
1720 | { | |
1721 | struct ipv6hdr *pip6; | |
1722 | ||
1723 | if (!pskb_may_pull(skb, sizeof(struct ipv6hdr))) | |
1724 | return false; | |
1725 | pip6 = ipv6_hdr(skb); | |
1726 | n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev); | |
dc5321d7 | 1727 | if (!n && (vxlan->cfg.flags & VXLAN_F_L3MISS)) { |
e15a00aa CW |
1728 | union vxlan_addr ipa = { |
1729 | .sin6.sin6_addr = pip6->daddr, | |
a45e92a5 | 1730 | .sin6.sin6_family = AF_INET6, |
e15a00aa CW |
1731 | }; |
1732 | ||
1733 | vxlan_ip_miss(dev, &ipa); | |
1734 | return false; | |
1735 | } | |
1736 | ||
1737 | break; | |
1738 | } | |
1739 | #endif | |
e4f67add DS |
1740 | default: |
1741 | return false; | |
1742 | } | |
1743 | ||
1744 | if (n) { | |
1745 | bool diff; | |
1746 | ||
7367d0b5 | 1747 | diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha); |
e4f67add DS |
1748 | if (diff) { |
1749 | memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest, | |
1750 | dev->addr_len); | |
1751 | memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len); | |
1752 | } | |
1753 | neigh_release(n); | |
1754 | return diff; | |
e4c7ed41 CW |
1755 | } |
1756 | ||
e4f67add DS |
1757 | return false; |
1758 | } | |
1759 | ||
af33c1ad | 1760 | static void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, u32 vxflags, |
3511494c TG |
1761 | struct vxlan_metadata *md) |
1762 | { | |
1763 | struct vxlanhdr_gbp *gbp; | |
1764 | ||
db79a621 TG |
1765 | if (!md->gbp) |
1766 | return; | |
1767 | ||
3511494c | 1768 | gbp = (struct vxlanhdr_gbp *)vxh; |
54bfd872 | 1769 | vxh->vx_flags |= VXLAN_HF_GBP; |
3511494c TG |
1770 | |
1771 | if (md->gbp & VXLAN_GBP_DONT_LEARN) | |
1772 | gbp->dont_learn = 1; | |
1773 | ||
1774 | if (md->gbp & VXLAN_GBP_POLICY_APPLIED) | |
1775 | gbp->policy_applied = 1; | |
1776 | ||
1777 | gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK); | |
1778 | } | |
1779 | ||
e1e5314d JB |
1780 | static int vxlan_build_gpe_hdr(struct vxlanhdr *vxh, u32 vxflags, |
1781 | __be16 protocol) | |
1782 | { | |
1783 | struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)vxh; | |
1784 | ||
1785 | gpe->np_applied = 1; | |
fa20e0e3 JB |
1786 | gpe->next_protocol = tun_p_from_eth_p(protocol); |
1787 | if (!gpe->next_protocol) | |
1788 | return -EPFNOSUPPORT; | |
1789 | return 0; | |
e1e5314d JB |
1790 | } |
1791 | ||
f491e56d JB |
1792 | static int vxlan_build_skb(struct sk_buff *skb, struct dst_entry *dst, |
1793 | int iphdr_len, __be32 vni, | |
1794 | struct vxlan_metadata *md, u32 vxflags, | |
b4ed5cad | 1795 | bool udp_sum) |
e4c7ed41 | 1796 | { |
e4c7ed41 | 1797 | struct vxlanhdr *vxh; |
e4c7ed41 CW |
1798 | int min_headroom; |
1799 | int err; | |
dfd8645e | 1800 | int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL; |
e1e5314d | 1801 | __be16 inner_protocol = htons(ETH_P_TEB); |
dfd8645e | 1802 | |
af33c1ad | 1803 | if ((vxflags & VXLAN_F_REMCSUM_TX) && |
dfd8645e TH |
1804 | skb->ip_summed == CHECKSUM_PARTIAL) { |
1805 | int csum_start = skb_checksum_start_offset(skb); | |
1806 | ||
1807 | if (csum_start <= VXLAN_MAX_REMCSUM_START && | |
1808 | !(csum_start & VXLAN_RCO_SHIFT_MASK) && | |
1809 | (skb->csum_offset == offsetof(struct udphdr, check) || | |
b5708501 | 1810 | skb->csum_offset == offsetof(struct tcphdr, check))) |
dfd8645e | 1811 | type |= SKB_GSO_TUNNEL_REMCSUM; |
dfd8645e | 1812 | } |
e4c7ed41 | 1813 | |
e4c7ed41 | 1814 | min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len |
4a4f86cc | 1815 | + VXLAN_HLEN + iphdr_len; |
649c5b8b PS |
1816 | |
1817 | /* Need space for new headers (invalidates iph ptr) */ | |
1818 | err = skb_cow_head(skb, min_headroom); | |
e1e5314d | 1819 | if (unlikely(err)) |
c46b7897 | 1820 | return err; |
649c5b8b | 1821 | |
aed069df AD |
1822 | err = iptunnel_handle_offloads(skb, type); |
1823 | if (err) | |
c46b7897 | 1824 | return err; |
b736a623 | 1825 | |
d58ff351 | 1826 | vxh = __skb_push(skb, sizeof(*vxh)); |
54bfd872 JB |
1827 | vxh->vx_flags = VXLAN_HF_VNI; |
1828 | vxh->vx_vni = vxlan_vni_field(vni); | |
49560532 | 1829 | |
dfd8645e | 1830 | if (type & SKB_GSO_TUNNEL_REMCSUM) { |
54bfd872 | 1831 | unsigned int start; |
dfd8645e | 1832 | |
54bfd872 JB |
1833 | start = skb_checksum_start_offset(skb) - sizeof(struct vxlanhdr); |
1834 | vxh->vx_vni |= vxlan_compute_rco(start, skb->csum_offset); | |
1835 | vxh->vx_flags |= VXLAN_HF_RCO; | |
dfd8645e TH |
1836 | |
1837 | if (!skb_is_gso(skb)) { | |
1838 | skb->ip_summed = CHECKSUM_NONE; | |
1839 | skb->encapsulation = 0; | |
1840 | } | |
1841 | } | |
1842 | ||
af33c1ad TH |
1843 | if (vxflags & VXLAN_F_GBP) |
1844 | vxlan_build_gbp_hdr(vxh, vxflags, md); | |
e1e5314d JB |
1845 | if (vxflags & VXLAN_F_GPE) { |
1846 | err = vxlan_build_gpe_hdr(vxh, vxflags, skb->protocol); | |
1847 | if (err < 0) | |
c46b7897 | 1848 | return err; |
e1e5314d JB |
1849 | inner_protocol = skb->protocol; |
1850 | } | |
3511494c | 1851 | |
e1e5314d | 1852 | skb_set_inner_protocol(skb, inner_protocol); |
039f5062 | 1853 | return 0; |
49560532 | 1854 | } |
49560532 | 1855 | |
655c3de1 | 1856 | static struct rtable *vxlan_get_route(struct vxlan_dev *vxlan, struct net_device *dev, |
1857 | struct vxlan_sock *sock4, | |
1a8496ba | 1858 | struct sk_buff *skb, int oif, u8 tos, |
4ecb1d83 | 1859 | __be32 daddr, __be32 *saddr, __be16 dport, __be16 sport, |
0c1d70af | 1860 | struct dst_cache *dst_cache, |
db3c6139 | 1861 | const struct ip_tunnel_info *info) |
1a8496ba | 1862 | { |
db3c6139 | 1863 | bool use_cache = ip_tunnel_dst_cache_usable(skb, info); |
1a8496ba JB |
1864 | struct rtable *rt = NULL; |
1865 | struct flowi4 fl4; | |
1866 | ||
655c3de1 | 1867 | if (!sock4) |
1868 | return ERR_PTR(-EIO); | |
1869 | ||
db3c6139 DB |
1870 | if (tos && !info) |
1871 | use_cache = false; | |
1872 | if (use_cache) { | |
0c1d70af PA |
1873 | rt = dst_cache_get_ip4(dst_cache, saddr); |
1874 | if (rt) | |
1875 | return rt; | |
1876 | } | |
1877 | ||
1a8496ba JB |
1878 | memset(&fl4, 0, sizeof(fl4)); |
1879 | fl4.flowi4_oif = oif; | |
1880 | fl4.flowi4_tos = RT_TOS(tos); | |
1881 | fl4.flowi4_mark = skb->mark; | |
1882 | fl4.flowi4_proto = IPPROTO_UDP; | |
1883 | fl4.daddr = daddr; | |
272d96a5 | 1884 | fl4.saddr = *saddr; |
4ecb1d83 MP |
1885 | fl4.fl4_dport = dport; |
1886 | fl4.fl4_sport = sport; | |
1a8496ba JB |
1887 | |
1888 | rt = ip_route_output_key(vxlan->net, &fl4); | |
655c3de1 | 1889 | if (likely(!IS_ERR(rt))) { |
1890 | if (rt->dst.dev == dev) { | |
1891 | netdev_dbg(dev, "circular route to %pI4\n", &daddr); | |
1892 | ip_rt_put(rt); | |
1893 | return ERR_PTR(-ELOOP); | |
1894 | } | |
1895 | ||
1a8496ba | 1896 | *saddr = fl4.saddr; |
0c1d70af PA |
1897 | if (use_cache) |
1898 | dst_cache_set_ip4(dst_cache, &rt->dst, fl4.saddr); | |
655c3de1 | 1899 | } else { |
1900 | netdev_dbg(dev, "no route to %pI4\n", &daddr); | |
1901 | return ERR_PTR(-ENETUNREACH); | |
0c1d70af | 1902 | } |
1a8496ba JB |
1903 | return rt; |
1904 | } | |
1905 | ||
e5d4b29f JB |
1906 | #if IS_ENABLED(CONFIG_IPV6) |
1907 | static struct dst_entry *vxlan6_get_route(struct vxlan_dev *vxlan, | |
655c3de1 | 1908 | struct net_device *dev, |
03dc52a8 | 1909 | struct vxlan_sock *sock6, |
1400615d | 1910 | struct sk_buff *skb, int oif, u8 tos, |
e7f70af1 | 1911 | __be32 label, |
e5d4b29f | 1912 | const struct in6_addr *daddr, |
0c1d70af | 1913 | struct in6_addr *saddr, |
4ecb1d83 | 1914 | __be16 dport, __be16 sport, |
db3c6139 DB |
1915 | struct dst_cache *dst_cache, |
1916 | const struct ip_tunnel_info *info) | |
e5d4b29f | 1917 | { |
db3c6139 | 1918 | bool use_cache = ip_tunnel_dst_cache_usable(skb, info); |
e5d4b29f JB |
1919 | struct dst_entry *ndst; |
1920 | struct flowi6 fl6; | |
1921 | int err; | |
1922 | ||
c6fcc4fc | 1923 | if (!sock6) |
1924 | return ERR_PTR(-EIO); | |
1925 | ||
1400615d DB |
1926 | if (tos && !info) |
1927 | use_cache = false; | |
db3c6139 | 1928 | if (use_cache) { |
0c1d70af PA |
1929 | ndst = dst_cache_get_ip6(dst_cache, saddr); |
1930 | if (ndst) | |
1931 | return ndst; | |
1932 | } | |
1933 | ||
e5d4b29f JB |
1934 | memset(&fl6, 0, sizeof(fl6)); |
1935 | fl6.flowi6_oif = oif; | |
1936 | fl6.daddr = *daddr; | |
272d96a5 | 1937 | fl6.saddr = *saddr; |
eaa93bf4 | 1938 | fl6.flowlabel = ip6_make_flowinfo(RT_TOS(tos), label); |
e5d4b29f JB |
1939 | fl6.flowi6_mark = skb->mark; |
1940 | fl6.flowi6_proto = IPPROTO_UDP; | |
4ecb1d83 MP |
1941 | fl6.fl6_dport = dport; |
1942 | fl6.fl6_sport = sport; | |
e5d4b29f JB |
1943 | |
1944 | err = ipv6_stub->ipv6_dst_lookup(vxlan->net, | |
c6fcc4fc | 1945 | sock6->sock->sk, |
e5d4b29f | 1946 | &ndst, &fl6); |
655c3de1 | 1947 | if (unlikely(err < 0)) { |
1948 | netdev_dbg(dev, "no route to %pI6\n", daddr); | |
1949 | return ERR_PTR(-ENETUNREACH); | |
1950 | } | |
1951 | ||
1952 | if (unlikely(ndst->dev == dev)) { | |
1953 | netdev_dbg(dev, "circular route to %pI6\n", daddr); | |
1954 | dst_release(ndst); | |
1955 | return ERR_PTR(-ELOOP); | |
1956 | } | |
e5d4b29f JB |
1957 | |
1958 | *saddr = fl6.saddr; | |
db3c6139 | 1959 | if (use_cache) |
0c1d70af | 1960 | dst_cache_set_ip6(dst_cache, ndst, saddr); |
e5d4b29f JB |
1961 | return ndst; |
1962 | } | |
1963 | #endif | |
1964 | ||
9dcc71e1 SS |
1965 | /* Bypass encapsulation if the destination is local */ |
1966 | static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan, | |
3ad7a4b1 | 1967 | struct vxlan_dev *dst_vxlan, __be32 vni) |
9dcc71e1 | 1968 | { |
8f84985f | 1969 | struct pcpu_sw_netstats *tx_stats, *rx_stats; |
e4c7ed41 CW |
1970 | union vxlan_addr loopback; |
1971 | union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip; | |
ce6502a8 LR |
1972 | struct net_device *dev = skb->dev; |
1973 | int len = skb->len; | |
9dcc71e1 | 1974 | |
8f84985f LR |
1975 | tx_stats = this_cpu_ptr(src_vxlan->dev->tstats); |
1976 | rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats); | |
9dcc71e1 SS |
1977 | skb->pkt_type = PACKET_HOST; |
1978 | skb->encapsulation = 0; | |
1979 | skb->dev = dst_vxlan->dev; | |
1980 | __skb_pull(skb, skb_network_offset(skb)); | |
1981 | ||
e4c7ed41 CW |
1982 | if (remote_ip->sa.sa_family == AF_INET) { |
1983 | loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK); | |
1984 | loopback.sa.sa_family = AF_INET; | |
1985 | #if IS_ENABLED(CONFIG_IPV6) | |
1986 | } else { | |
1987 | loopback.sin6.sin6_addr = in6addr_loopback; | |
1988 | loopback.sa.sa_family = AF_INET6; | |
1989 | #endif | |
1990 | } | |
1991 | ||
dc5321d7 | 1992 | if (dst_vxlan->cfg.flags & VXLAN_F_LEARN) |
87613de9 MS |
1993 | vxlan_snoop(skb->dev, &loopback, eth_hdr(skb)->h_source, 0, |
1994 | vni); | |
9dcc71e1 SS |
1995 | |
1996 | u64_stats_update_begin(&tx_stats->syncp); | |
1997 | tx_stats->tx_packets++; | |
ce6502a8 | 1998 | tx_stats->tx_bytes += len; |
9dcc71e1 SS |
1999 | u64_stats_update_end(&tx_stats->syncp); |
2000 | ||
2001 | if (netif_rx(skb) == NET_RX_SUCCESS) { | |
2002 | u64_stats_update_begin(&rx_stats->syncp); | |
2003 | rx_stats->rx_packets++; | |
ce6502a8 | 2004 | rx_stats->rx_bytes += len; |
9dcc71e1 SS |
2005 | u64_stats_update_end(&rx_stats->syncp); |
2006 | } else { | |
ce6502a8 | 2007 | dev->stats.rx_dropped++; |
9dcc71e1 SS |
2008 | } |
2009 | } | |
2010 | ||
fee1fad7 | 2011 | static int encap_bypass_if_local(struct sk_buff *skb, struct net_device *dev, |
49f810f0 MS |
2012 | struct vxlan_dev *vxlan, |
2013 | union vxlan_addr *daddr, | |
2014 | __be16 dst_port, int dst_ifindex, __be32 vni, | |
2015 | struct dst_entry *dst, | |
fee1fad7 | 2016 | u32 rt_flags) |
2017 | { | |
2018 | #if IS_ENABLED(CONFIG_IPV6) | |
2019 | /* IPv6 rt-flags are checked against RTF_LOCAL, but the value of | |
2020 | * RTF_LOCAL is equal to RTCF_LOCAL. So to keep code simple | |
2021 | * we can use RTCF_LOCAL which works for ipv4 and ipv6 route entry. | |
2022 | */ | |
2023 | BUILD_BUG_ON(RTCF_LOCAL != RTF_LOCAL); | |
2024 | #endif | |
2025 | /* Bypass encapsulation if the destination is local */ | |
2026 | if (rt_flags & RTCF_LOCAL && | |
2027 | !(rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) { | |
2028 | struct vxlan_dev *dst_vxlan; | |
2029 | ||
2030 | dst_release(dst); | |
49f810f0 | 2031 | dst_vxlan = vxlan_find_vni(vxlan->net, dst_ifindex, vni, |
fee1fad7 | 2032 | daddr->sa.sa_family, dst_port, |
dc5321d7 | 2033 | vxlan->cfg.flags); |
fee1fad7 | 2034 | if (!dst_vxlan) { |
2035 | dev->stats.tx_errors++; | |
2036 | kfree_skb(skb); | |
2037 | ||
2038 | return -ENOENT; | |
2039 | } | |
3ad7a4b1 | 2040 | vxlan_encap_bypass(skb, vxlan, dst_vxlan, vni); |
fee1fad7 | 2041 | return 1; |
2042 | } | |
2043 | ||
2044 | return 0; | |
2045 | } | |
2046 | ||
4ad16930 | 2047 | static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev, |
3ad7a4b1 RP |
2048 | __be32 default_vni, struct vxlan_rdst *rdst, |
2049 | bool did_rsc) | |
d342894c | 2050 | { |
d71785ff | 2051 | struct dst_cache *dst_cache; |
3093fbe7 | 2052 | struct ip_tunnel_info *info; |
d342894c | 2053 | struct vxlan_dev *vxlan = netdev_priv(dev); |
0770b53b | 2054 | const struct iphdr *old_iph = ip_hdr(skb); |
e4c7ed41 | 2055 | union vxlan_addr *dst; |
272d96a5 | 2056 | union vxlan_addr remote_ip, local_ip; |
ee122c79 TG |
2057 | struct vxlan_metadata _md; |
2058 | struct vxlan_metadata *md = &_md; | |
e4c7ed41 | 2059 | __be16 src_port = 0, dst_port; |
655c3de1 | 2060 | struct dst_entry *ndst = NULL; |
e7f70af1 | 2061 | __be32 vni, label; |
d342894c | 2062 | __u8 tos, ttl; |
49f810f0 | 2063 | int ifindex; |
0e6fbc5b | 2064 | int err; |
dc5321d7 | 2065 | u32 flags = vxlan->cfg.flags; |
b4ed5cad | 2066 | bool udp_sum = false; |
f491e56d | 2067 | bool xnet = !net_eq(vxlan->net, dev_net(vxlan->dev)); |
d342894c | 2068 | |
61adedf3 | 2069 | info = skb_tunnel_info(skb); |
3093fbe7 | 2070 | |
ee122c79 | 2071 | if (rdst) { |
0770b53b | 2072 | dst = &rdst->remote_ip; |
2073 | if (vxlan_addr_any(dst)) { | |
2074 | if (did_rsc) { | |
2075 | /* short-circuited back to local bridge */ | |
3ad7a4b1 | 2076 | vxlan_encap_bypass(skb, vxlan, vxlan, default_vni); |
0770b53b | 2077 | return; |
2078 | } | |
2079 | goto drop; | |
2080 | } | |
2081 | ||
0dfbdf41 | 2082 | dst_port = rdst->remote_port ? rdst->remote_port : vxlan->cfg.dst_port; |
3ad7a4b1 | 2083 | vni = (rdst->remote_vni) ? : default_vni; |
49f810f0 | 2084 | ifindex = rdst->remote_ifindex; |
1158632b | 2085 | local_ip = vxlan->cfg.saddr; |
d71785ff | 2086 | dst_cache = &rdst->dst_cache; |
0770b53b | 2087 | md->gbp = skb->mark; |
2088 | ttl = vxlan->cfg.ttl; | |
2089 | if (!ttl && vxlan_addr_multicast(dst)) | |
2090 | ttl = 1; | |
2091 | ||
2092 | tos = vxlan->cfg.tos; | |
2093 | if (tos == 1) | |
2094 | tos = ip_tunnel_get_dsfield(old_iph, skb); | |
2095 | ||
2096 | if (dst->sa.sa_family == AF_INET) | |
2097 | udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM_TX); | |
2098 | else | |
2099 | udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM6_TX); | |
2100 | label = vxlan->cfg.label; | |
ee122c79 TG |
2101 | } else { |
2102 | if (!info) { | |
2103 | WARN_ONCE(1, "%s: Missing encapsulation instructions\n", | |
2104 | dev->name); | |
2105 | goto drop; | |
2106 | } | |
b1be00a6 | 2107 | remote_ip.sa.sa_family = ip_tunnel_info_af(info); |
272d96a5 | 2108 | if (remote_ip.sa.sa_family == AF_INET) { |
a725e514 | 2109 | remote_ip.sin.sin_addr.s_addr = info->key.u.ipv4.dst; |
272d96a5 | 2110 | local_ip.sin.sin_addr.s_addr = info->key.u.ipv4.src; |
2111 | } else { | |
a725e514 | 2112 | remote_ip.sin6.sin6_addr = info->key.u.ipv6.dst; |
272d96a5 | 2113 | local_ip.sin6.sin6_addr = info->key.u.ipv6.src; |
2114 | } | |
ee122c79 | 2115 | dst = &remote_ip; |
0770b53b | 2116 | dst_port = info->key.tp_dst ? : vxlan->cfg.dst_port; |
2117 | vni = tunnel_id_to_key32(info->key.tun_id); | |
49f810f0 | 2118 | ifindex = 0; |
d71785ff | 2119 | dst_cache = &info->dst_cache; |
0770b53b | 2120 | if (info->options_len) |
2121 | md = ip_tunnel_info_opts(info); | |
a725e514 JB |
2122 | ttl = info->key.ttl; |
2123 | tos = info->key.tos; | |
e7f70af1 | 2124 | label = info->key.label; |
b4ed5cad | 2125 | udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM); |
a725e514 | 2126 | } |
0770b53b | 2127 | src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min, |
2128 | vxlan->cfg.port_max, true); | |
a725e514 | 2129 | |
56de859e | 2130 | rcu_read_lock(); |
e4c7ed41 | 2131 | if (dst->sa.sa_family == AF_INET) { |
c6fcc4fc | 2132 | struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock); |
c46b7897 | 2133 | struct rtable *rt; |
0770b53b | 2134 | __be16 df = 0; |
c6fcc4fc | 2135 | |
49f810f0 | 2136 | rt = vxlan_get_route(vxlan, dev, sock4, skb, ifindex, tos, |
272d96a5 | 2137 | dst->sin.sin_addr.s_addr, |
1158632b | 2138 | &local_ip.sin.sin_addr.s_addr, |
4ecb1d83 | 2139 | dst_port, src_port, |
d71785ff | 2140 | dst_cache, info); |
8ebd115b DM |
2141 | if (IS_ERR(rt)) { |
2142 | err = PTR_ERR(rt); | |
c46b7897 | 2143 | goto tx_error; |
8ebd115b | 2144 | } |
e4c7ed41 CW |
2145 | |
2146 | /* Bypass encapsulation if the destination is local */ | |
fee1fad7 | 2147 | if (!info) { |
2148 | err = encap_bypass_if_local(skb, dev, vxlan, dst, | |
49f810f0 MS |
2149 | dst_port, ifindex, vni, |
2150 | &rt->dst, rt->rt_flags); | |
fee1fad7 | 2151 | if (err) |
56de859e | 2152 | goto out_unlock; |
fee1fad7 | 2153 | } else if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT) { |
6ceb31ca | 2154 | df = htons(IP_DF); |
fee1fad7 | 2155 | } |
6ceb31ca | 2156 | |
c46b7897 | 2157 | ndst = &rt->dst; |
4699beb7 XL |
2158 | if (skb_dst(skb)) { |
2159 | int mtu = dst_mtu(ndst) - VXLAN_HEADROOM; | |
2160 | ||
2161 | skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, | |
2162 | skb, mtu); | |
2163 | } | |
2164 | ||
e4c7ed41 CW |
2165 | tos = ip_tunnel_ecn_encap(tos, old_iph, skb); |
2166 | ttl = ttl ? : ip4_dst_hoplimit(&rt->dst); | |
c46b7897 | 2167 | err = vxlan_build_skb(skb, ndst, sizeof(struct iphdr), |
54bfd872 | 2168 | vni, md, flags, udp_sum); |
f491e56d | 2169 | if (err < 0) |
c46b7897 | 2170 | goto tx_error; |
f491e56d | 2171 | |
1158632b | 2172 | udp_tunnel_xmit_skb(rt, sock4->sock->sk, skb, local_ip.sin.sin_addr.s_addr, |
f491e56d JB |
2173 | dst->sin.sin_addr.s_addr, tos, ttl, df, |
2174 | src_port, dst_port, xnet, !udp_sum); | |
e4c7ed41 CW |
2175 | #if IS_ENABLED(CONFIG_IPV6) |
2176 | } else { | |
c6fcc4fc | 2177 | struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock); |
e4c7ed41 | 2178 | |
49f810f0 | 2179 | ndst = vxlan6_get_route(vxlan, dev, sock6, skb, ifindex, tos, |
272d96a5 | 2180 | label, &dst->sin6.sin6_addr, |
1158632b | 2181 | &local_ip.sin6.sin6_addr, |
4ecb1d83 | 2182 | dst_port, src_port, |
db3c6139 | 2183 | dst_cache, info); |
e5d4b29f | 2184 | if (IS_ERR(ndst)) { |
8ebd115b | 2185 | err = PTR_ERR(ndst); |
c46b7897 | 2186 | ndst = NULL; |
9dcc71e1 | 2187 | goto tx_error; |
e4c7ed41 | 2188 | } |
655c3de1 | 2189 | |
fee1fad7 | 2190 | if (!info) { |
2191 | u32 rt6i_flags = ((struct rt6_info *)ndst)->rt6i_flags; | |
49560532 | 2192 | |
fee1fad7 | 2193 | err = encap_bypass_if_local(skb, dev, vxlan, dst, |
49f810f0 MS |
2194 | dst_port, ifindex, vni, |
2195 | ndst, rt6i_flags); | |
fee1fad7 | 2196 | if (err) |
56de859e | 2197 | goto out_unlock; |
fee1fad7 | 2198 | } |
35e2d115 | 2199 | |
4699beb7 XL |
2200 | if (skb_dst(skb)) { |
2201 | int mtu = dst_mtu(ndst) - VXLAN6_HEADROOM; | |
2202 | ||
2203 | skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, | |
2204 | skb, mtu); | |
2205 | } | |
2206 | ||
1400615d | 2207 | tos = ip_tunnel_ecn_encap(tos, old_iph, skb); |
e4c7ed41 | 2208 | ttl = ttl ? : ip6_dst_hoplimit(ndst); |
f491e56d JB |
2209 | skb_scrub_packet(skb, xnet); |
2210 | err = vxlan_build_skb(skb, ndst, sizeof(struct ipv6hdr), | |
54bfd872 | 2211 | vni, md, flags, udp_sum); |
c46b7897 | 2212 | if (err < 0) |
2213 | goto tx_error; | |
2214 | ||
0770b53b | 2215 | udp_tunnel6_xmit_skb(ndst, sock6->sock->sk, skb, dev, |
1158632b | 2216 | &local_ip.sin6.sin6_addr, |
272d96a5 | 2217 | &dst->sin6.sin6_addr, tos, ttl, |
e7f70af1 | 2218 | label, src_port, dst_port, !udp_sum); |
e4c7ed41 CW |
2219 | #endif |
2220 | } | |
56de859e JK |
2221 | out_unlock: |
2222 | rcu_read_unlock(); | |
4ad16930 | 2223 | return; |
d342894c | 2224 | |
2225 | drop: | |
2226 | dev->stats.tx_dropped++; | |
c46b7897 | 2227 | dev_kfree_skb(skb); |
2228 | return; | |
d342894c | 2229 | |
2230 | tx_error: | |
56de859e | 2231 | rcu_read_unlock(); |
655c3de1 | 2232 | if (err == -ELOOP) |
2233 | dev->stats.collisions++; | |
2234 | else if (err == -ENETUNREACH) | |
2235 | dev->stats.tx_carrier_errors++; | |
c46b7897 | 2236 | dst_release(ndst); |
d342894c | 2237 | dev->stats.tx_errors++; |
c46b7897 | 2238 | kfree_skb(skb); |
d342894c | 2239 | } |
2240 | ||
6681712d DS |
2241 | /* Transmit local packets over Vxlan |
2242 | * | |
2243 | * Outer IP header inherits ECN and DF from inner header. | |
2244 | * Outer UDP destination is the VXLAN assigned port. | |
2245 | * source port is based on hash of flow | |
2246 | */ | |
2247 | static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev) | |
2248 | { | |
2249 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
829af2cd | 2250 | struct vxlan_rdst *rdst, *fdst = NULL; |
3093fbe7 | 2251 | const struct ip_tunnel_info *info; |
6681712d | 2252 | bool did_rsc = false; |
6681712d | 2253 | struct vxlan_fdb *f; |
829af2cd | 2254 | struct ethhdr *eth; |
3ad7a4b1 | 2255 | __be32 vni = 0; |
6681712d | 2256 | |
61adedf3 | 2257 | info = skb_tunnel_info(skb); |
3093fbe7 | 2258 | |
6681712d | 2259 | skb_reset_mac_header(skb); |
6681712d | 2260 | |
dc5321d7 | 2261 | if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) { |
3ad7a4b1 RP |
2262 | if (info && info->mode & IP_TUNNEL_INFO_BRIDGE && |
2263 | info->mode & IP_TUNNEL_INFO_TX) { | |
2264 | vni = tunnel_id_to_key32(info->key.tun_id); | |
2265 | } else { | |
2266 | if (info && info->mode & IP_TUNNEL_INFO_TX) | |
2267 | vxlan_xmit_one(skb, dev, vni, NULL, false); | |
2268 | else | |
2269 | kfree_skb(skb); | |
2270 | return NETDEV_TX_OK; | |
2271 | } | |
47e5d1b0 JB |
2272 | } |
2273 | ||
dc5321d7 | 2274 | if (vxlan->cfg.flags & VXLAN_F_PROXY) { |
47e5d1b0 | 2275 | eth = eth_hdr(skb); |
f564f45c | 2276 | if (ntohs(eth->h_proto) == ETH_P_ARP) |
3ad7a4b1 | 2277 | return arp_reduce(dev, skb, vni); |
f564f45c | 2278 | #if IS_ENABLED(CONFIG_IPV6) |
829af2cd XL |
2279 | else if (ntohs(eth->h_proto) == ETH_P_IPV6 && |
2280 | pskb_may_pull(skb, sizeof(struct ipv6hdr) + | |
2281 | sizeof(struct nd_msg)) && | |
2282 | ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) { | |
2283 | struct nd_msg *m = (struct nd_msg *)(ipv6_hdr(skb) + 1); | |
2284 | ||
2285 | if (m->icmph.icmp6_code == 0 && | |
2286 | m->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION) | |
f1fb08f6 | 2287 | return neigh_reduce(dev, skb, vni); |
f564f45c CW |
2288 | } |
2289 | #endif | |
2290 | } | |
6681712d | 2291 | |
47e5d1b0 | 2292 | eth = eth_hdr(skb); |
3ad7a4b1 | 2293 | f = vxlan_find_mac(vxlan, eth->h_dest, vni); |
ae884082 DS |
2294 | did_rsc = false; |
2295 | ||
dc5321d7 | 2296 | if (f && (f->flags & NTF_ROUTER) && (vxlan->cfg.flags & VXLAN_F_RSC) && |
e15a00aa CW |
2297 | (ntohs(eth->h_proto) == ETH_P_IP || |
2298 | ntohs(eth->h_proto) == ETH_P_IPV6)) { | |
ae884082 DS |
2299 | did_rsc = route_shortcircuit(dev, skb); |
2300 | if (did_rsc) | |
3ad7a4b1 | 2301 | f = vxlan_find_mac(vxlan, eth->h_dest, vni); |
ae884082 DS |
2302 | } |
2303 | ||
6681712d | 2304 | if (f == NULL) { |
3ad7a4b1 | 2305 | f = vxlan_find_mac(vxlan, all_zeros_mac, vni); |
afbd8bae | 2306 | if (f == NULL) { |
dc5321d7 | 2307 | if ((vxlan->cfg.flags & VXLAN_F_L2MISS) && |
afbd8bae MR |
2308 | !is_multicast_ether_addr(eth->h_dest)) |
2309 | vxlan_fdb_miss(vxlan, eth->h_dest); | |
2310 | ||
2311 | dev->stats.tx_dropped++; | |
8f646c92 | 2312 | kfree_skb(skb); |
afbd8bae MR |
2313 | return NETDEV_TX_OK; |
2314 | } | |
2315 | } | |
6681712d | 2316 | |
afbd8bae MR |
2317 | list_for_each_entry_rcu(rdst, &f->remotes, list) { |
2318 | struct sk_buff *skb1; | |
6681712d | 2319 | |
8f646c92 ED |
2320 | if (!fdst) { |
2321 | fdst = rdst; | |
2322 | continue; | |
2323 | } | |
afbd8bae MR |
2324 | skb1 = skb_clone(skb, GFP_ATOMIC); |
2325 | if (skb1) | |
3ad7a4b1 | 2326 | vxlan_xmit_one(skb1, dev, vni, rdst, did_rsc); |
6681712d DS |
2327 | } |
2328 | ||
8f646c92 | 2329 | if (fdst) |
3ad7a4b1 | 2330 | vxlan_xmit_one(skb, dev, vni, fdst, did_rsc); |
8f646c92 ED |
2331 | else |
2332 | kfree_skb(skb); | |
4ad16930 | 2333 | return NETDEV_TX_OK; |
6681712d DS |
2334 | } |
2335 | ||
d342894c | 2336 | /* Walk the forwarding table and purge stale entries */ |
2337 | static void vxlan_cleanup(unsigned long arg) | |
2338 | { | |
2339 | struct vxlan_dev *vxlan = (struct vxlan_dev *) arg; | |
2340 | unsigned long next_timer = jiffies + FDB_AGE_INTERVAL; | |
2341 | unsigned int h; | |
2342 | ||
2343 | if (!netif_running(vxlan->dev)) | |
2344 | return; | |
2345 | ||
d342894c | 2346 | for (h = 0; h < FDB_HASH_SIZE; ++h) { |
2347 | struct hlist_node *p, *n; | |
14e1d0fa SD |
2348 | |
2349 | spin_lock_bh(&vxlan->hash_lock); | |
d342894c | 2350 | hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) { |
2351 | struct vxlan_fdb *f | |
2352 | = container_of(p, struct vxlan_fdb, hlist); | |
2353 | unsigned long timeout; | |
2354 | ||
efb5f68f | 2355 | if (f->state & (NUD_PERMANENT | NUD_NOARP)) |
d342894c | 2356 | continue; |
2357 | ||
def499c9 RP |
2358 | if (f->flags & NTF_EXT_LEARNED) |
2359 | continue; | |
2360 | ||
0dfbdf41 | 2361 | timeout = f->used + vxlan->cfg.age_interval * HZ; |
d342894c | 2362 | if (time_before_eq(timeout, jiffies)) { |
2363 | netdev_dbg(vxlan->dev, | |
2364 | "garbage collect %pM\n", | |
2365 | f->eth_addr); | |
2366 | f->state = NUD_STALE; | |
2367 | vxlan_fdb_destroy(vxlan, f); | |
2368 | } else if (time_before(timeout, next_timer)) | |
2369 | next_timer = timeout; | |
2370 | } | |
14e1d0fa | 2371 | spin_unlock_bh(&vxlan->hash_lock); |
d342894c | 2372 | } |
d342894c | 2373 | |
2374 | mod_timer(&vxlan->age_timer, next_timer); | |
2375 | } | |
2376 | ||
a53cb29b MB |
2377 | static void vxlan_vs_del_dev(struct vxlan_dev *vxlan) |
2378 | { | |
2379 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); | |
2380 | ||
2381 | spin_lock(&vn->sock_lock); | |
69e76661 JB |
2382 | hlist_del_init_rcu(&vxlan->hlist4.hlist); |
2383 | #if IS_ENABLED(CONFIG_IPV6) | |
2384 | hlist_del_init_rcu(&vxlan->hlist6.hlist); | |
2385 | #endif | |
a53cb29b MB |
2386 | spin_unlock(&vn->sock_lock); |
2387 | } | |
2388 | ||
69e76661 JB |
2389 | static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan, |
2390 | struct vxlan_dev_node *node) | |
9c2e24e1 | 2391 | { |
56ef9c90 | 2392 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
54bfd872 | 2393 | __be32 vni = vxlan->default_dst.remote_vni; |
9c2e24e1 | 2394 | |
69e76661 | 2395 | node->vxlan = vxlan; |
56ef9c90 | 2396 | spin_lock(&vn->sock_lock); |
69e76661 | 2397 | hlist_add_head_rcu(&node->hlist, vni_head(vs, vni)); |
56ef9c90 | 2398 | spin_unlock(&vn->sock_lock); |
9c2e24e1 PS |
2399 | } |
2400 | ||
d342894c | 2401 | /* Setup stats when device is created */ |
2402 | static int vxlan_init(struct net_device *dev) | |
2403 | { | |
1c213bd2 | 2404 | dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); |
e8171045 | 2405 | if (!dev->tstats) |
d342894c | 2406 | return -ENOMEM; |
2407 | ||
2408 | return 0; | |
2409 | } | |
2410 | ||
3ad7a4b1 | 2411 | static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan, __be32 vni) |
afbd8bae MR |
2412 | { |
2413 | struct vxlan_fdb *f; | |
2414 | ||
2415 | spin_lock_bh(&vxlan->hash_lock); | |
3ad7a4b1 | 2416 | f = __vxlan_find_mac(vxlan, all_zeros_mac, vni); |
afbd8bae MR |
2417 | if (f) |
2418 | vxlan_fdb_destroy(vxlan, f); | |
2419 | spin_unlock_bh(&vxlan->hash_lock); | |
2420 | } | |
2421 | ||
ebf4063e SH |
2422 | static void vxlan_uninit(struct net_device *dev) |
2423 | { | |
2424 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
ebf4063e | 2425 | |
3ad7a4b1 | 2426 | vxlan_fdb_delete_default(vxlan, vxlan->cfg.vni); |
afbd8bae | 2427 | |
ebf4063e SH |
2428 | free_percpu(dev->tstats); |
2429 | } | |
2430 | ||
d342894c | 2431 | /* Start ageing timer and join group when device is brought up */ |
2432 | static int vxlan_open(struct net_device *dev) | |
2433 | { | |
2434 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
205f356d | 2435 | int ret; |
56ef9c90 | 2436 | |
205f356d JB |
2437 | ret = vxlan_sock_add(vxlan); |
2438 | if (ret < 0) | |
2439 | return ret; | |
d342894c | 2440 | |
79d4a94f | 2441 | if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) { |
56ef9c90 | 2442 | ret = vxlan_igmp_join(vxlan); |
bef0057b MRL |
2443 | if (ret == -EADDRINUSE) |
2444 | ret = 0; | |
56ef9c90 | 2445 | if (ret) { |
205f356d | 2446 | vxlan_sock_release(vxlan); |
56ef9c90 MRL |
2447 | return ret; |
2448 | } | |
d342894c | 2449 | } |
2450 | ||
0dfbdf41 | 2451 | if (vxlan->cfg.age_interval) |
d342894c | 2452 | mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL); |
2453 | ||
56ef9c90 | 2454 | return ret; |
d342894c | 2455 | } |
2456 | ||
2457 | /* Purge the forwarding table */ | |
8b3f9337 | 2458 | static void vxlan_flush(struct vxlan_dev *vxlan, bool do_all) |
d342894c | 2459 | { |
31fec5aa | 2460 | unsigned int h; |
d342894c | 2461 | |
2462 | spin_lock_bh(&vxlan->hash_lock); | |
2463 | for (h = 0; h < FDB_HASH_SIZE; ++h) { | |
2464 | struct hlist_node *p, *n; | |
2465 | hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) { | |
2466 | struct vxlan_fdb *f | |
2467 | = container_of(p, struct vxlan_fdb, hlist); | |
8b3f9337 RP |
2468 | if (!do_all && (f->state & (NUD_PERMANENT | NUD_NOARP))) |
2469 | continue; | |
afbd8bae MR |
2470 | /* the all_zeros_mac entry is deleted at vxlan_uninit */ |
2471 | if (!is_zero_ether_addr(f->eth_addr)) | |
2472 | vxlan_fdb_destroy(vxlan, f); | |
d342894c | 2473 | } |
2474 | } | |
2475 | spin_unlock_bh(&vxlan->hash_lock); | |
2476 | } | |
2477 | ||
2478 | /* Cleanup timer and forwarding table on shutdown */ | |
2479 | static int vxlan_stop(struct net_device *dev) | |
2480 | { | |
2481 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
f01ec1c0 | 2482 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
56ef9c90 | 2483 | int ret = 0; |
d342894c | 2484 | |
24c0e683 | 2485 | if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip) && |
f13b1689 | 2486 | !vxlan_group_used(vn, vxlan)) |
56ef9c90 | 2487 | ret = vxlan_igmp_leave(vxlan); |
d342894c | 2488 | |
2489 | del_timer_sync(&vxlan->age_timer); | |
2490 | ||
8b3f9337 | 2491 | vxlan_flush(vxlan, false); |
205f356d | 2492 | vxlan_sock_release(vxlan); |
d342894c | 2493 | |
56ef9c90 | 2494 | return ret; |
d342894c | 2495 | } |
2496 | ||
d342894c | 2497 | /* Stub, nothing needs to be done. */ |
2498 | static void vxlan_set_multicast_list(struct net_device *dev) | |
2499 | { | |
2500 | } | |
2501 | ||
91572088 | 2502 | static int vxlan_change_mtu(struct net_device *dev, int new_mtu) |
345010b5 | 2503 | { |
91572088 JW |
2504 | struct vxlan_dev *vxlan = netdev_priv(dev); |
2505 | struct vxlan_rdst *dst = &vxlan->default_dst; | |
2506 | struct net_device *lowerdev = __dev_get_by_index(vxlan->net, | |
2507 | dst->remote_ifindex); | |
ce44a4ae | 2508 | bool use_ipv6 = !!(vxlan->cfg.flags & VXLAN_F_IPV6); |
345010b5 | 2509 | |
91572088 JW |
2510 | /* This check is different than dev->max_mtu, because it looks at |
2511 | * the lowerdev->mtu, rather than the static dev->max_mtu | |
2512 | */ | |
2513 | if (lowerdev) { | |
2514 | int max_mtu = lowerdev->mtu - | |
2515 | (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM); | |
2516 | if (new_mtu > max_mtu) | |
72564b59 | 2517 | return -EINVAL; |
72564b59 DW |
2518 | } |
2519 | ||
345010b5 DB |
2520 | dev->mtu = new_mtu; |
2521 | return 0; | |
2522 | } | |
2523 | ||
fc4099f1 PS |
2524 | static int vxlan_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) |
2525 | { | |
2526 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
2527 | struct ip_tunnel_info *info = skb_tunnel_info(skb); | |
2528 | __be16 sport, dport; | |
2529 | ||
2530 | sport = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min, | |
2531 | vxlan->cfg.port_max, true); | |
2532 | dport = info->key.tp_dst ? : vxlan->cfg.dst_port; | |
2533 | ||
239e944f | 2534 | if (ip_tunnel_info_af(info) == AF_INET) { |
c6fcc4fc | 2535 | struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock); |
1a8496ba JB |
2536 | struct rtable *rt; |
2537 | ||
655c3de1 | 2538 | rt = vxlan_get_route(vxlan, dev, sock4, skb, 0, info->key.tos, |
1a8496ba | 2539 | info->key.u.ipv4.dst, |
22f0708a PA |
2540 | &info->key.u.ipv4.src, dport, sport, |
2541 | &info->dst_cache, info); | |
1a8496ba JB |
2542 | if (IS_ERR(rt)) |
2543 | return PTR_ERR(rt); | |
2544 | ip_rt_put(rt); | |
239e944f JB |
2545 | } else { |
2546 | #if IS_ENABLED(CONFIG_IPV6) | |
03dc52a8 | 2547 | struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock); |
239e944f JB |
2548 | struct dst_entry *ndst; |
2549 | ||
655c3de1 | 2550 | ndst = vxlan6_get_route(vxlan, dev, sock6, skb, 0, info->key.tos, |
e7f70af1 | 2551 | info->key.label, &info->key.u.ipv6.dst, |
22f0708a PA |
2552 | &info->key.u.ipv6.src, dport, sport, |
2553 | &info->dst_cache, info); | |
239e944f JB |
2554 | if (IS_ERR(ndst)) |
2555 | return PTR_ERR(ndst); | |
2556 | dst_release(ndst); | |
239e944f JB |
2557 | #else /* !CONFIG_IPV6 */ |
2558 | return -EPFNOSUPPORT; | |
2559 | #endif | |
2560 | } | |
1a8496ba JB |
2561 | info->key.tp_src = sport; |
2562 | info->key.tp_dst = dport; | |
239e944f | 2563 | return 0; |
fc4099f1 PS |
2564 | } |
2565 | ||
0c867c9b | 2566 | static const struct net_device_ops vxlan_netdev_ether_ops = { |
d342894c | 2567 | .ndo_init = vxlan_init, |
ebf4063e | 2568 | .ndo_uninit = vxlan_uninit, |
d342894c | 2569 | .ndo_open = vxlan_open, |
2570 | .ndo_stop = vxlan_stop, | |
2571 | .ndo_start_xmit = vxlan_xmit, | |
e8171045 | 2572 | .ndo_get_stats64 = ip_tunnel_get_stats64, |
d342894c | 2573 | .ndo_set_rx_mode = vxlan_set_multicast_list, |
345010b5 | 2574 | .ndo_change_mtu = vxlan_change_mtu, |
d342894c | 2575 | .ndo_validate_addr = eth_validate_addr, |
2576 | .ndo_set_mac_address = eth_mac_addr, | |
2577 | .ndo_fdb_add = vxlan_fdb_add, | |
2578 | .ndo_fdb_del = vxlan_fdb_delete, | |
2579 | .ndo_fdb_dump = vxlan_fdb_dump, | |
fc4099f1 | 2580 | .ndo_fill_metadata_dst = vxlan_fill_metadata_dst, |
d342894c | 2581 | }; |
2582 | ||
e1e5314d JB |
2583 | static const struct net_device_ops vxlan_netdev_raw_ops = { |
2584 | .ndo_init = vxlan_init, | |
2585 | .ndo_uninit = vxlan_uninit, | |
2586 | .ndo_open = vxlan_open, | |
2587 | .ndo_stop = vxlan_stop, | |
2588 | .ndo_start_xmit = vxlan_xmit, | |
2589 | .ndo_get_stats64 = ip_tunnel_get_stats64, | |
2590 | .ndo_change_mtu = vxlan_change_mtu, | |
2591 | .ndo_fill_metadata_dst = vxlan_fill_metadata_dst, | |
2592 | }; | |
2593 | ||
d342894c | 2594 | /* Info for udev, that this is a virtual tunnel endpoint */ |
2595 | static struct device_type vxlan_type = { | |
2596 | .name = "vxlan", | |
2597 | }; | |
2598 | ||
e5de25dc | 2599 | /* Calls the ndo_udp_tunnel_add of the caller in order to |
35e42379 | 2600 | * supply the listening VXLAN udp ports. Callers are expected |
e5de25dc | 2601 | * to implement the ndo_udp_tunnel_add. |
53cf5275 | 2602 | */ |
2d2b13fc | 2603 | static void vxlan_offload_rx_ports(struct net_device *dev, bool push) |
53cf5275 JG |
2604 | { |
2605 | struct vxlan_sock *vs; | |
2606 | struct net *net = dev_net(dev); | |
2607 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
35e42379 | 2608 | unsigned int i; |
53cf5275 JG |
2609 | |
2610 | spin_lock(&vn->sock_lock); | |
2611 | for (i = 0; i < PORT_HASH_SIZE; ++i) { | |
2d2b13fc SD |
2612 | hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) { |
2613 | unsigned short type; | |
2614 | ||
2615 | if (vs->flags & VXLAN_F_GPE) | |
2616 | type = UDP_TUNNEL_TYPE_VXLAN_GPE; | |
2617 | else | |
2618 | type = UDP_TUNNEL_TYPE_VXLAN; | |
2619 | ||
2620 | if (push) | |
2621 | udp_tunnel_push_rx_port(dev, vs->sock, type); | |
2622 | else | |
2623 | udp_tunnel_drop_rx_port(dev, vs->sock, type); | |
2624 | } | |
53cf5275 JG |
2625 | } |
2626 | spin_unlock(&vn->sock_lock); | |
2627 | } | |
53cf5275 | 2628 | |
d342894c | 2629 | /* Initialize the device structure. */ |
2630 | static void vxlan_setup(struct net_device *dev) | |
2631 | { | |
2632 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
31fec5aa | 2633 | unsigned int h; |
d342894c | 2634 | |
65226ef8 JB |
2635 | eth_hw_addr_random(dev); |
2636 | ether_setup(dev); | |
2637 | ||
cf124db5 | 2638 | dev->needs_free_netdev = true; |
d342894c | 2639 | SET_NETDEV_DEVTYPE(dev, &vxlan_type); |
2640 | ||
d342894c | 2641 | dev->features |= NETIF_F_LLTX; |
d6727fe3 | 2642 | dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM; |
0afb1666 | 2643 | dev->features |= NETIF_F_RXCSUM; |
05c0db08 | 2644 | dev->features |= NETIF_F_GSO_SOFTWARE; |
0afb1666 | 2645 | |
1eaa8178 | 2646 | dev->vlan_features = dev->features; |
0afb1666 | 2647 | dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM; |
05c0db08 | 2648 | dev->hw_features |= NETIF_F_GSO_SOFTWARE; |
02875878 | 2649 | netif_keep_dst(dev); |
0c867c9b | 2650 | dev->priv_flags |= IFF_NO_QUEUE; |
d342894c | 2651 | |
a985343b MS |
2652 | /* MTU range: 68 - 65535 */ |
2653 | dev->min_mtu = ETH_MIN_MTU; | |
2654 | dev->max_mtu = ETH_MAX_MTU; | |
2655 | ||
553675fb | 2656 | INIT_LIST_HEAD(&vxlan->next); |
d342894c | 2657 | spin_lock_init(&vxlan->hash_lock); |
2658 | ||
2659 | init_timer_deferrable(&vxlan->age_timer); | |
2660 | vxlan->age_timer.function = vxlan_cleanup; | |
2661 | vxlan->age_timer.data = (unsigned long) vxlan; | |
2662 | ||
2663 | vxlan->dev = dev; | |
2664 | ||
58ce31cc TH |
2665 | gro_cells_init(&vxlan->gro_cells, dev); |
2666 | ||
d342894c | 2667 | for (h = 0; h < FDB_HASH_SIZE; ++h) |
2668 | INIT_HLIST_HEAD(&vxlan->fdb_head[h]); | |
2669 | } | |
2670 | ||
0c867c9b JB |
2671 | static void vxlan_ether_setup(struct net_device *dev) |
2672 | { | |
0c867c9b JB |
2673 | dev->priv_flags &= ~IFF_TX_SKB_SHARING; |
2674 | dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; | |
2675 | dev->netdev_ops = &vxlan_netdev_ether_ops; | |
2676 | } | |
2677 | ||
e1e5314d JB |
2678 | static void vxlan_raw_setup(struct net_device *dev) |
2679 | { | |
65226ef8 | 2680 | dev->header_ops = NULL; |
e1e5314d JB |
2681 | dev->type = ARPHRD_NONE; |
2682 | dev->hard_header_len = 0; | |
2683 | dev->addr_len = 0; | |
e1e5314d JB |
2684 | dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST; |
2685 | dev->netdev_ops = &vxlan_netdev_raw_ops; | |
2686 | } | |
2687 | ||
d342894c | 2688 | static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = { |
2689 | [IFLA_VXLAN_ID] = { .type = NLA_U32 }, | |
5d174dd8 | 2690 | [IFLA_VXLAN_GROUP] = { .len = FIELD_SIZEOF(struct iphdr, daddr) }, |
e4c7ed41 | 2691 | [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) }, |
d342894c | 2692 | [IFLA_VXLAN_LINK] = { .type = NLA_U32 }, |
2693 | [IFLA_VXLAN_LOCAL] = { .len = FIELD_SIZEOF(struct iphdr, saddr) }, | |
e4c7ed41 | 2694 | [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) }, |
d342894c | 2695 | [IFLA_VXLAN_TOS] = { .type = NLA_U8 }, |
2696 | [IFLA_VXLAN_TTL] = { .type = NLA_U8 }, | |
e7f70af1 | 2697 | [IFLA_VXLAN_LABEL] = { .type = NLA_U32 }, |
d342894c | 2698 | [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 }, |
2699 | [IFLA_VXLAN_AGEING] = { .type = NLA_U32 }, | |
2700 | [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 }, | |
05f47d69 | 2701 | [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) }, |
e4f67add DS |
2702 | [IFLA_VXLAN_PROXY] = { .type = NLA_U8 }, |
2703 | [IFLA_VXLAN_RSC] = { .type = NLA_U8 }, | |
2704 | [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 }, | |
2705 | [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 }, | |
f8a9b1bc | 2706 | [IFLA_VXLAN_COLLECT_METADATA] = { .type = NLA_U8 }, |
823aa873 | 2707 | [IFLA_VXLAN_PORT] = { .type = NLA_U16 }, |
5c91ae08 TH |
2708 | [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 }, |
2709 | [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 }, | |
2710 | [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 }, | |
dfd8645e TH |
2711 | [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 }, |
2712 | [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 }, | |
3511494c | 2713 | [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, }, |
e1e5314d | 2714 | [IFLA_VXLAN_GPE] = { .type = NLA_FLAG, }, |
0ace2ca8 | 2715 | [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG }, |
d342894c | 2716 | }; |
2717 | ||
a8b8a889 MS |
2718 | static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[], |
2719 | struct netlink_ext_ack *extack) | |
d342894c | 2720 | { |
2721 | if (tb[IFLA_ADDRESS]) { | |
2722 | if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) { | |
653ef6a3 GM |
2723 | NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_ADDRESS], |
2724 | "Provided link layer address is not Ethernet"); | |
d342894c | 2725 | return -EINVAL; |
2726 | } | |
2727 | ||
2728 | if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) { | |
653ef6a3 GM |
2729 | NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_ADDRESS], |
2730 | "Provided Ethernet address is not unicast"); | |
d342894c | 2731 | return -EADDRNOTAVAIL; |
2732 | } | |
2733 | } | |
2734 | ||
a985343b | 2735 | if (tb[IFLA_MTU]) { |
019b13ae | 2736 | u32 mtu = nla_get_u32(tb[IFLA_MTU]); |
a985343b | 2737 | |
653ef6a3 GM |
2738 | if (mtu < ETH_MIN_MTU || mtu > ETH_MAX_MTU) { |
2739 | NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_MTU], | |
2740 | "MTU must be between 68 and 65535"); | |
a985343b | 2741 | return -EINVAL; |
653ef6a3 | 2742 | } |
a985343b MS |
2743 | } |
2744 | ||
653ef6a3 GM |
2745 | if (!data) { |
2746 | NL_SET_ERR_MSG(extack, | |
2747 | "Required attributes not provided to perform the operation"); | |
d342894c | 2748 | return -EINVAL; |
653ef6a3 | 2749 | } |
d342894c | 2750 | |
2751 | if (data[IFLA_VXLAN_ID]) { | |
a985343b MS |
2752 | u32 id = nla_get_u32(data[IFLA_VXLAN_ID]); |
2753 | ||
653ef6a3 GM |
2754 | if (id >= VXLAN_N_VID) { |
2755 | NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_ID], | |
2756 | "VXLAN ID must be lower than 16777216"); | |
d342894c | 2757 | return -ERANGE; |
653ef6a3 | 2758 | } |
d342894c | 2759 | } |
2760 | ||
05f47d69 | 2761 | if (data[IFLA_VXLAN_PORT_RANGE]) { |
2762 | const struct ifla_vxlan_port_range *p | |
2763 | = nla_data(data[IFLA_VXLAN_PORT_RANGE]); | |
2764 | ||
2765 | if (ntohs(p->high) < ntohs(p->low)) { | |
653ef6a3 GM |
2766 | NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_PORT_RANGE], |
2767 | "Invalid source port range"); | |
05f47d69 | 2768 | return -EINVAL; |
2769 | } | |
2770 | } | |
2771 | ||
d342894c | 2772 | return 0; |
2773 | } | |
2774 | ||
1b13c97f YB |
2775 | static void vxlan_get_drvinfo(struct net_device *netdev, |
2776 | struct ethtool_drvinfo *drvinfo) | |
2777 | { | |
2778 | strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version)); | |
2779 | strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver)); | |
2780 | } | |
2781 | ||
2782 | static const struct ethtool_ops vxlan_ethtool_ops = { | |
2783 | .get_drvinfo = vxlan_get_drvinfo, | |
2784 | .get_link = ethtool_op_get_link, | |
2785 | }; | |
2786 | ||
3ee64f39 TH |
2787 | static struct socket *vxlan_create_sock(struct net *net, bool ipv6, |
2788 | __be16 port, u32 flags) | |
553675fb | 2789 | { |
e4c7ed41 | 2790 | struct socket *sock; |
3ee64f39 TH |
2791 | struct udp_port_cfg udp_conf; |
2792 | int err; | |
e4c7ed41 | 2793 | |
3ee64f39 | 2794 | memset(&udp_conf, 0, sizeof(udp_conf)); |
e4c7ed41 | 2795 | |
3ee64f39 TH |
2796 | if (ipv6) { |
2797 | udp_conf.family = AF_INET6; | |
3ee64f39 | 2798 | udp_conf.use_udp6_rx_checksums = |
3dc2b6a8 | 2799 | !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX); |
a43a9ef6 | 2800 | udp_conf.ipv6_v6only = 1; |
3ee64f39 TH |
2801 | } else { |
2802 | udp_conf.family = AF_INET; | |
e4c7ed41 | 2803 | } |
e4c7ed41 | 2804 | |
3ee64f39 | 2805 | udp_conf.local_udp_port = port; |
553675fb | 2806 | |
3ee64f39 TH |
2807 | /* Open UDP socket */ |
2808 | err = udp_sock_create(net, &udp_conf, &sock); | |
2809 | if (err < 0) | |
2810 | return ERR_PTR(err); | |
e4c7ed41 | 2811 | |
39deb2c7 | 2812 | return sock; |
e4c7ed41 CW |
2813 | } |
2814 | ||
2815 | /* Create new listen socket if needed */ | |
b1be00a6 JB |
2816 | static struct vxlan_sock *vxlan_socket_create(struct net *net, bool ipv6, |
2817 | __be16 port, u32 flags) | |
e4c7ed41 CW |
2818 | { |
2819 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
2820 | struct vxlan_sock *vs; | |
2821 | struct socket *sock; | |
e4c7ed41 | 2822 | unsigned int h; |
acbf74a7 | 2823 | struct udp_tunnel_sock_cfg tunnel_cfg; |
e4c7ed41 | 2824 | |
dc01e7d3 | 2825 | vs = kzalloc(sizeof(*vs), GFP_KERNEL); |
e4c7ed41 CW |
2826 | if (!vs) |
2827 | return ERR_PTR(-ENOMEM); | |
2828 | ||
2829 | for (h = 0; h < VNI_HASH_SIZE; ++h) | |
2830 | INIT_HLIST_HEAD(&vs->vni_list[h]); | |
2831 | ||
3ee64f39 | 2832 | sock = vxlan_create_sock(net, ipv6, port, flags); |
39deb2c7 | 2833 | if (IS_ERR(sock)) { |
553675fb | 2834 | kfree(vs); |
e50fddc8 | 2835 | return ERR_CAST(sock); |
553675fb | 2836 | } |
e4c7ed41 CW |
2837 | |
2838 | vs->sock = sock; | |
66af846f | 2839 | refcount_set(&vs->refcnt, 1); |
af33c1ad | 2840 | vs->flags = (flags & VXLAN_F_RCV_FLAGS); |
553675fb | 2841 | |
9c2e24e1 PS |
2842 | spin_lock(&vn->sock_lock); |
2843 | hlist_add_head_rcu(&vs->hlist, vs_head(net, port)); | |
e7b3db5e | 2844 | udp_tunnel_notify_add_rx_port(sock, |
b9adcd69 AD |
2845 | (vs->flags & VXLAN_F_GPE) ? |
2846 | UDP_TUNNEL_TYPE_VXLAN_GPE : | |
e7b3db5e | 2847 | UDP_TUNNEL_TYPE_VXLAN); |
9c2e24e1 | 2848 | spin_unlock(&vn->sock_lock); |
553675fb | 2849 | |
2850 | /* Mark socket as an encapsulation socket. */ | |
5602c48c | 2851 | memset(&tunnel_cfg, 0, sizeof(tunnel_cfg)); |
acbf74a7 AZ |
2852 | tunnel_cfg.sk_user_data = vs; |
2853 | tunnel_cfg.encap_type = 1; | |
f2d1968e | 2854 | tunnel_cfg.encap_rcv = vxlan_rcv; |
acbf74a7 | 2855 | tunnel_cfg.encap_destroy = NULL; |
5602c48c TH |
2856 | tunnel_cfg.gro_receive = vxlan_gro_receive; |
2857 | tunnel_cfg.gro_complete = vxlan_gro_complete; | |
acbf74a7 AZ |
2858 | |
2859 | setup_udp_tunnel_sock(net, sock, &tunnel_cfg); | |
e4c7ed41 | 2860 | |
9c2e24e1 PS |
2861 | return vs; |
2862 | } | |
2863 | ||
b1be00a6 | 2864 | static int __vxlan_sock_add(struct vxlan_dev *vxlan, bool ipv6) |
9c2e24e1 | 2865 | { |
205f356d JB |
2866 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
2867 | struct vxlan_sock *vs = NULL; | |
69e76661 | 2868 | struct vxlan_dev_node *node; |
9c2e24e1 | 2869 | |
205f356d | 2870 | if (!vxlan->cfg.no_share) { |
56ef9c90 | 2871 | spin_lock(&vn->sock_lock); |
205f356d | 2872 | vs = vxlan_find_sock(vxlan->net, ipv6 ? AF_INET6 : AF_INET, |
dc5321d7 | 2873 | vxlan->cfg.dst_port, vxlan->cfg.flags); |
66af846f | 2874 | if (vs && !refcount_inc_not_zero(&vs->refcnt)) { |
56ef9c90 | 2875 | spin_unlock(&vn->sock_lock); |
205f356d | 2876 | return -EBUSY; |
56ef9c90 MRL |
2877 | } |
2878 | spin_unlock(&vn->sock_lock); | |
2879 | } | |
205f356d | 2880 | if (!vs) |
b1be00a6 | 2881 | vs = vxlan_socket_create(vxlan->net, ipv6, |
dc5321d7 | 2882 | vxlan->cfg.dst_port, vxlan->cfg.flags); |
205f356d JB |
2883 | if (IS_ERR(vs)) |
2884 | return PTR_ERR(vs); | |
b1be00a6 | 2885 | #if IS_ENABLED(CONFIG_IPV6) |
69e76661 | 2886 | if (ipv6) { |
c6fcc4fc | 2887 | rcu_assign_pointer(vxlan->vn6_sock, vs); |
69e76661 JB |
2888 | node = &vxlan->hlist6; |
2889 | } else | |
b1be00a6 | 2890 | #endif |
69e76661 | 2891 | { |
c6fcc4fc | 2892 | rcu_assign_pointer(vxlan->vn4_sock, vs); |
69e76661 JB |
2893 | node = &vxlan->hlist4; |
2894 | } | |
2895 | vxlan_vs_add_dev(vs, vxlan, node); | |
205f356d | 2896 | return 0; |
553675fb | 2897 | } |
2898 | ||
b1be00a6 JB |
2899 | static int vxlan_sock_add(struct vxlan_dev *vxlan) |
2900 | { | |
dc5321d7 MS |
2901 | bool metadata = vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA; |
2902 | bool ipv6 = vxlan->cfg.flags & VXLAN_F_IPV6 || metadata; | |
d074bf96 | 2903 | bool ipv4 = !ipv6 || metadata; |
b1be00a6 JB |
2904 | int ret = 0; |
2905 | ||
c6fcc4fc | 2906 | RCU_INIT_POINTER(vxlan->vn4_sock, NULL); |
b1be00a6 | 2907 | #if IS_ENABLED(CONFIG_IPV6) |
c6fcc4fc | 2908 | RCU_INIT_POINTER(vxlan->vn6_sock, NULL); |
d074bf96 | 2909 | if (ipv6) { |
b1be00a6 | 2910 | ret = __vxlan_sock_add(vxlan, true); |
d074bf96 JB |
2911 | if (ret < 0 && ret != -EAFNOSUPPORT) |
2912 | ipv4 = false; | |
2913 | } | |
b1be00a6 | 2914 | #endif |
d074bf96 | 2915 | if (ipv4) |
b1be00a6 JB |
2916 | ret = __vxlan_sock_add(vxlan, false); |
2917 | if (ret < 0) | |
2918 | vxlan_sock_release(vxlan); | |
2919 | return ret; | |
2920 | } | |
2921 | ||
a985343b MS |
2922 | static int vxlan_config_validate(struct net *src_net, struct vxlan_config *conf, |
2923 | struct net_device **lower, | |
653ef6a3 GM |
2924 | struct vxlan_dev *old, |
2925 | struct netlink_ext_ack *extack) | |
d342894c | 2926 | { |
33564bbb | 2927 | struct vxlan_net *vn = net_generic(src_net, vxlan_net_id); |
a985343b | 2928 | struct vxlan_dev *tmp; |
e4c7ed41 | 2929 | bool use_ipv6 = false; |
d342894c | 2930 | |
a985343b MS |
2931 | if (conf->flags & VXLAN_F_GPE) { |
2932 | /* For now, allow GPE only together with | |
2933 | * COLLECT_METADATA. This can be relaxed later; in such | |
2934 | * case, the other side of the PtP link will have to be | |
2935 | * provided. | |
2936 | */ | |
2937 | if ((conf->flags & ~VXLAN_F_ALLOWED_GPE) || | |
2938 | !(conf->flags & VXLAN_F_COLLECT_METADATA)) { | |
653ef6a3 GM |
2939 | NL_SET_ERR_MSG(extack, |
2940 | "VXLAN GPE does not support this combination of attributes"); | |
a985343b | 2941 | return -EINVAL; |
3555621d | 2942 | } |
e1e5314d | 2943 | } |
0c867c9b | 2944 | |
ce44a4ae MS |
2945 | if (!conf->remote_ip.sa.sa_family && !conf->saddr.sa.sa_family) { |
2946 | /* Unless IPv6 is explicitly requested, assume IPv4 */ | |
a985343b | 2947 | conf->remote_ip.sa.sa_family = AF_INET; |
ce44a4ae MS |
2948 | conf->saddr.sa.sa_family = AF_INET; |
2949 | } else if (!conf->remote_ip.sa.sa_family) { | |
2950 | conf->remote_ip.sa.sa_family = conf->saddr.sa.sa_family; | |
2951 | } else if (!conf->saddr.sa.sa_family) { | |
2952 | conf->saddr.sa.sa_family = conf->remote_ip.sa.sa_family; | |
2953 | } | |
2954 | ||
653ef6a3 GM |
2955 | if (conf->saddr.sa.sa_family != conf->remote_ip.sa.sa_family) { |
2956 | NL_SET_ERR_MSG(extack, | |
2957 | "Local and remote address must be from the same family"); | |
ce44a4ae | 2958 | return -EINVAL; |
653ef6a3 | 2959 | } |
e4c7ed41 | 2960 | |
653ef6a3 GM |
2961 | if (vxlan_addr_multicast(&conf->saddr)) { |
2962 | NL_SET_ERR_MSG(extack, "Local address cannot be multicast"); | |
0f22a3c6 | 2963 | return -EINVAL; |
653ef6a3 | 2964 | } |
0f22a3c6 | 2965 | |
ce44a4ae | 2966 | if (conf->saddr.sa.sa_family == AF_INET6) { |
653ef6a3 GM |
2967 | if (!IS_ENABLED(CONFIG_IPV6)) { |
2968 | NL_SET_ERR_MSG(extack, | |
2969 | "IPv6 support not enabled in the kernel"); | |
057ba29b | 2970 | return -EPFNOSUPPORT; |
653ef6a3 | 2971 | } |
e4c7ed41 | 2972 | use_ipv6 = true; |
a985343b | 2973 | conf->flags |= VXLAN_F_IPV6; |
0f22a3c6 MS |
2974 | |
2975 | if (!(conf->flags & VXLAN_F_COLLECT_METADATA)) { | |
2976 | int local_type = | |
2977 | ipv6_addr_type(&conf->saddr.sin6.sin6_addr); | |
2978 | int remote_type = | |
2979 | ipv6_addr_type(&conf->remote_ip.sin6.sin6_addr); | |
2980 | ||
2981 | if (local_type & IPV6_ADDR_LINKLOCAL) { | |
2982 | if (!(remote_type & IPV6_ADDR_LINKLOCAL) && | |
653ef6a3 GM |
2983 | (remote_type != IPV6_ADDR_ANY)) { |
2984 | NL_SET_ERR_MSG(extack, | |
2985 | "Invalid combination of local and remote address scopes"); | |
0f22a3c6 | 2986 | return -EINVAL; |
653ef6a3 | 2987 | } |
0f22a3c6 MS |
2988 | |
2989 | conf->flags |= VXLAN_F_IPV6_LINKLOCAL; | |
2990 | } else { | |
2991 | if (remote_type == | |
653ef6a3 GM |
2992 | (IPV6_ADDR_UNICAST | IPV6_ADDR_LINKLOCAL)) { |
2993 | NL_SET_ERR_MSG(extack, | |
2994 | "Invalid combination of local and remote address scopes"); | |
0f22a3c6 | 2995 | return -EINVAL; |
653ef6a3 | 2996 | } |
0f22a3c6 MS |
2997 | |
2998 | conf->flags &= ~VXLAN_F_IPV6_LINKLOCAL; | |
2999 | } | |
3000 | } | |
057ba29b | 3001 | } |
d342894c | 3002 | |
653ef6a3 GM |
3003 | if (conf->label && !use_ipv6) { |
3004 | NL_SET_ERR_MSG(extack, | |
3005 | "Label attribute only applies to IPv6 VXLAN devices"); | |
e7f70af1 | 3006 | return -EINVAL; |
653ef6a3 | 3007 | } |
e7f70af1 | 3008 | |
a985343b MS |
3009 | if (conf->remote_ifindex) { |
3010 | struct net_device *lowerdev; | |
34e02aa1 | 3011 | |
a985343b | 3012 | lowerdev = __dev_get_by_index(src_net, conf->remote_ifindex); |
653ef6a3 GM |
3013 | if (!lowerdev) { |
3014 | NL_SET_ERR_MSG(extack, | |
3015 | "Invalid local interface, device not found"); | |
34e02aa1 | 3016 | return -ENODEV; |
653ef6a3 | 3017 | } |
d342894c | 3018 | |
e4c7ed41 CW |
3019 | #if IS_ENABLED(CONFIG_IPV6) |
3020 | if (use_ipv6) { | |
3021 | struct inet6_dev *idev = __in6_dev_get(lowerdev); | |
653ef6a3 GM |
3022 | if (idev && idev->cnf.disable_ipv6) { |
3023 | NL_SET_ERR_MSG(extack, | |
3024 | "IPv6 support disabled by administrator"); | |
e4c7ed41 | 3025 | return -EPERM; |
653ef6a3 | 3026 | } |
e4c7ed41 CW |
3027 | } |
3028 | #endif | |
3029 | ||
a985343b MS |
3030 | *lower = lowerdev; |
3031 | } else { | |
653ef6a3 GM |
3032 | if (vxlan_addr_multicast(&conf->remote_ip)) { |
3033 | NL_SET_ERR_MSG(extack, | |
3034 | "Local interface required for multicast remote destination"); | |
3035 | ||
a985343b | 3036 | return -EINVAL; |
653ef6a3 | 3037 | } |
1ba56fb4 | 3038 | |
0f22a3c6 | 3039 | #if IS_ENABLED(CONFIG_IPV6) |
653ef6a3 GM |
3040 | if (conf->flags & VXLAN_F_IPV6_LINKLOCAL) { |
3041 | NL_SET_ERR_MSG(extack, | |
3042 | "Local interface required for link-local local/remote addresses"); | |
0f22a3c6 | 3043 | return -EINVAL; |
653ef6a3 | 3044 | } |
0f22a3c6 MS |
3045 | #endif |
3046 | ||
a985343b | 3047 | *lower = NULL; |
9dc2ad10 | 3048 | } |
d342894c | 3049 | |
a985343b MS |
3050 | if (!conf->dst_port) { |
3051 | if (conf->flags & VXLAN_F_GPE) | |
3052 | conf->dst_port = htons(4790); /* IANA VXLAN-GPE port */ | |
3053 | else | |
3054 | conf->dst_port = htons(vxlan_port); | |
d6acfeb1 FM |
3055 | } |
3056 | ||
a985343b MS |
3057 | if (!conf->age_interval) |
3058 | conf->age_interval = FDB_AGE_DEFAULT; | |
91572088 | 3059 | |
a985343b MS |
3060 | list_for_each_entry(tmp, &vn->vxlan_list, next) { |
3061 | if (tmp == old) | |
3062 | continue; | |
91572088 | 3063 | |
49f810f0 MS |
3064 | if (tmp->cfg.vni != conf->vni) |
3065 | continue; | |
3066 | if (tmp->cfg.dst_port != conf->dst_port) | |
3067 | continue; | |
3068 | if ((tmp->cfg.flags & (VXLAN_F_RCV_FLAGS | VXLAN_F_IPV6)) != | |
ce44a4ae | 3069 | (conf->flags & (VXLAN_F_RCV_FLAGS | VXLAN_F_IPV6))) |
49f810f0 MS |
3070 | continue; |
3071 | ||
3072 | if ((conf->flags & VXLAN_F_IPV6_LINKLOCAL) && | |
3073 | tmp->cfg.remote_ifindex != conf->remote_ifindex) | |
3074 | continue; | |
3075 | ||
653ef6a3 GM |
3076 | NL_SET_ERR_MSG(extack, |
3077 | "A VXLAN device with the specified VNI already exists"); | |
49f810f0 | 3078 | return -EEXIST; |
a985343b | 3079 | } |
91572088 | 3080 | |
a985343b MS |
3081 | return 0; |
3082 | } | |
3083 | ||
3084 | static void vxlan_config_apply(struct net_device *dev, | |
3085 | struct vxlan_config *conf, | |
889ce937 SD |
3086 | struct net_device *lowerdev, |
3087 | struct net *src_net, | |
3088 | bool changelink) | |
a985343b MS |
3089 | { |
3090 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3091 | struct vxlan_rdst *dst = &vxlan->default_dst; | |
3092 | unsigned short needed_headroom = ETH_HLEN; | |
3093 | bool use_ipv6 = !!(conf->flags & VXLAN_F_IPV6); | |
3094 | int max_mtu = ETH_MAX_MTU; | |
3095 | ||
3096 | if (!changelink) { | |
3097 | if (conf->flags & VXLAN_F_GPE) | |
3098 | vxlan_raw_setup(dev); | |
3099 | else | |
3100 | vxlan_ether_setup(dev); | |
3101 | ||
3102 | if (conf->mtu) | |
3103 | dev->mtu = conf->mtu; | |
889ce937 SD |
3104 | |
3105 | vxlan->net = src_net; | |
a985343b MS |
3106 | } |
3107 | ||
3108 | dst->remote_vni = conf->vni; | |
3109 | ||
3110 | memcpy(&dst->remote_ip, &conf->remote_ip, sizeof(conf->remote_ip)); | |
3111 | ||
3112 | if (lowerdev) { | |
3113 | dst->remote_ifindex = conf->remote_ifindex; | |
3114 | ||
3115 | dev->gso_max_size = lowerdev->gso_max_size; | |
3116 | dev->gso_max_segs = lowerdev->gso_max_segs; | |
91572088 | 3117 | |
a985343b | 3118 | needed_headroom = lowerdev->hard_header_len; |
91572088 | 3119 | |
a985343b MS |
3120 | max_mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : |
3121 | VXLAN_HEADROOM); | |
215b69e2 AK |
3122 | if (max_mtu < ETH_MIN_MTU) |
3123 | max_mtu = ETH_MIN_MTU; | |
3124 | ||
3125 | if (!changelink && !conf->mtu) | |
3126 | dev->mtu = max_mtu; | |
7e059158 DW |
3127 | } |
3128 | ||
a985343b MS |
3129 | if (dev->mtu > max_mtu) |
3130 | dev->mtu = max_mtu; | |
3131 | ||
b1be00a6 JB |
3132 | if (use_ipv6 || conf->flags & VXLAN_F_COLLECT_METADATA) |
3133 | needed_headroom += VXLAN6_HEADROOM; | |
3134 | else | |
3135 | needed_headroom += VXLAN_HEADROOM; | |
3136 | dev->needed_headroom = needed_headroom; | |
3137 | ||
0dfbdf41 | 3138 | memcpy(&vxlan->cfg, conf, sizeof(*conf)); |
a985343b | 3139 | } |
0dfbdf41 | 3140 | |
a985343b | 3141 | static int vxlan_dev_configure(struct net *src_net, struct net_device *dev, |
653ef6a3 GM |
3142 | struct vxlan_config *conf, bool changelink, |
3143 | struct netlink_ext_ack *extack) | |
a985343b MS |
3144 | { |
3145 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3146 | struct net_device *lowerdev; | |
3147 | int ret; | |
0dfbdf41 | 3148 | |
653ef6a3 | 3149 | ret = vxlan_config_validate(src_net, conf, &lowerdev, vxlan, extack); |
a985343b MS |
3150 | if (ret) |
3151 | return ret; | |
8bcdc4f3 | 3152 | |
889ce937 | 3153 | vxlan_config_apply(dev, conf, lowerdev, src_net, changelink); |
0dfbdf41 | 3154 | |
0dfbdf41 TG |
3155 | return 0; |
3156 | } | |
3157 | ||
c80498e3 | 3158 | static int __vxlan_dev_create(struct net *net, struct net_device *dev, |
653ef6a3 GM |
3159 | struct vxlan_config *conf, |
3160 | struct netlink_ext_ack *extack) | |
c80498e3 ND |
3161 | { |
3162 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
3163 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3164 | int err; | |
3165 | ||
653ef6a3 | 3166 | err = vxlan_dev_configure(net, dev, conf, false, extack); |
c80498e3 ND |
3167 | if (err) |
3168 | return err; | |
3169 | ||
3170 | dev->ethtool_ops = &vxlan_ethtool_ops; | |
3171 | ||
3172 | /* create an fdb entry for a valid default destination */ | |
3173 | if (!vxlan_addr_any(&vxlan->default_dst.remote_ip)) { | |
3174 | err = vxlan_fdb_create(vxlan, all_zeros_mac, | |
3175 | &vxlan->default_dst.remote_ip, | |
3176 | NUD_REACHABLE | NUD_PERMANENT, | |
3177 | NLM_F_EXCL | NLM_F_CREATE, | |
3178 | vxlan->cfg.dst_port, | |
3179 | vxlan->default_dst.remote_vni, | |
3180 | vxlan->default_dst.remote_vni, | |
3181 | vxlan->default_dst.remote_ifindex, | |
3182 | NTF_SELF); | |
3183 | if (err) | |
3184 | return err; | |
3185 | } | |
3186 | ||
3187 | err = register_netdevice(dev); | |
3188 | if (err) { | |
3189 | vxlan_fdb_delete_default(vxlan, vxlan->default_dst.remote_vni); | |
3190 | return err; | |
3191 | } | |
3192 | ||
3193 | list_add(&vxlan->next, &vn->vxlan_list); | |
3194 | return 0; | |
3195 | } | |
3196 | ||
8bcdc4f3 RP |
3197 | static int vxlan_nl2conf(struct nlattr *tb[], struct nlattr *data[], |
3198 | struct net_device *dev, struct vxlan_config *conf, | |
3199 | bool changelink) | |
0dfbdf41 | 3200 | { |
8bcdc4f3 | 3201 | struct vxlan_dev *vxlan = netdev_priv(dev); |
0dfbdf41 | 3202 | |
8bcdc4f3 | 3203 | memset(conf, 0, sizeof(*conf)); |
e277de5f | 3204 | |
8bcdc4f3 RP |
3205 | /* if changelink operation, start with old existing cfg */ |
3206 | if (changelink) | |
3207 | memcpy(conf, &vxlan->cfg, sizeof(*conf)); | |
3208 | ||
3209 | if (data[IFLA_VXLAN_ID]) { | |
3210 | __be32 vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID])); | |
3211 | ||
3212 | if (changelink && (vni != conf->vni)) | |
3213 | return -EOPNOTSUPP; | |
3214 | conf->vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID])); | |
3215 | } | |
0dfbdf41 TG |
3216 | |
3217 | if (data[IFLA_VXLAN_GROUP]) { | |
ce44a4ae MS |
3218 | if (changelink && (conf->remote_ip.sa.sa_family != AF_INET)) |
3219 | return -EOPNOTSUPP; | |
3220 | ||
8bcdc4f3 | 3221 | conf->remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]); |
ce44a4ae | 3222 | conf->remote_ip.sa.sa_family = AF_INET; |
0dfbdf41 TG |
3223 | } else if (data[IFLA_VXLAN_GROUP6]) { |
3224 | if (!IS_ENABLED(CONFIG_IPV6)) | |
3225 | return -EPFNOSUPPORT; | |
3226 | ||
ce44a4ae MS |
3227 | if (changelink && (conf->remote_ip.sa.sa_family != AF_INET6)) |
3228 | return -EOPNOTSUPP; | |
3229 | ||
8bcdc4f3 RP |
3230 | conf->remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]); |
3231 | conf->remote_ip.sa.sa_family = AF_INET6; | |
0dfbdf41 TG |
3232 | } |
3233 | ||
3234 | if (data[IFLA_VXLAN_LOCAL]) { | |
ce44a4ae MS |
3235 | if (changelink && (conf->saddr.sa.sa_family != AF_INET)) |
3236 | return -EOPNOTSUPP; | |
3237 | ||
8bcdc4f3 RP |
3238 | conf->saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]); |
3239 | conf->saddr.sa.sa_family = AF_INET; | |
0dfbdf41 TG |
3240 | } else if (data[IFLA_VXLAN_LOCAL6]) { |
3241 | if (!IS_ENABLED(CONFIG_IPV6)) | |
3242 | return -EPFNOSUPPORT; | |
3243 | ||
ce44a4ae MS |
3244 | if (changelink && (conf->saddr.sa.sa_family != AF_INET6)) |
3245 | return -EOPNOTSUPP; | |
3246 | ||
0dfbdf41 | 3247 | /* TODO: respect scope id */ |
8bcdc4f3 RP |
3248 | conf->saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]); |
3249 | conf->saddr.sa.sa_family = AF_INET6; | |
0dfbdf41 TG |
3250 | } |
3251 | ||
3252 | if (data[IFLA_VXLAN_LINK]) | |
8bcdc4f3 | 3253 | conf->remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]); |
0dfbdf41 | 3254 | |
d342894c | 3255 | if (data[IFLA_VXLAN_TOS]) |
8bcdc4f3 | 3256 | conf->tos = nla_get_u8(data[IFLA_VXLAN_TOS]); |
d342894c | 3257 | |
afb97186 | 3258 | if (data[IFLA_VXLAN_TTL]) |
8bcdc4f3 | 3259 | conf->ttl = nla_get_u8(data[IFLA_VXLAN_TTL]); |
afb97186 | 3260 | |
e7f70af1 | 3261 | if (data[IFLA_VXLAN_LABEL]) |
8bcdc4f3 | 3262 | conf->label = nla_get_be32(data[IFLA_VXLAN_LABEL]) & |
e7f70af1 DB |
3263 | IPV6_FLOWLABEL_MASK; |
3264 | ||
8bcdc4f3 | 3265 | if (data[IFLA_VXLAN_LEARNING]) { |
dc5321d7 | 3266 | if (nla_get_u8(data[IFLA_VXLAN_LEARNING])) |
8bcdc4f3 | 3267 | conf->flags |= VXLAN_F_LEARN; |
dc5321d7 | 3268 | else |
8bcdc4f3 | 3269 | conf->flags &= ~VXLAN_F_LEARN; |
8bcdc4f3 RP |
3270 | } else if (!changelink) { |
3271 | /* default to learn on a new device */ | |
3272 | conf->flags |= VXLAN_F_LEARN; | |
3273 | } | |
d342894c | 3274 | |
8bcdc4f3 RP |
3275 | if (data[IFLA_VXLAN_AGEING]) { |
3276 | if (changelink) | |
3277 | return -EOPNOTSUPP; | |
3278 | conf->age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]); | |
3279 | } | |
d342894c | 3280 | |
8bcdc4f3 RP |
3281 | if (data[IFLA_VXLAN_PROXY]) { |
3282 | if (changelink) | |
3283 | return -EOPNOTSUPP; | |
3284 | if (nla_get_u8(data[IFLA_VXLAN_PROXY])) | |
3285 | conf->flags |= VXLAN_F_PROXY; | |
3286 | } | |
e4f67add | 3287 | |
8bcdc4f3 RP |
3288 | if (data[IFLA_VXLAN_RSC]) { |
3289 | if (changelink) | |
3290 | return -EOPNOTSUPP; | |
3291 | if (nla_get_u8(data[IFLA_VXLAN_RSC])) | |
3292 | conf->flags |= VXLAN_F_RSC; | |
3293 | } | |
e4f67add | 3294 | |
8bcdc4f3 RP |
3295 | if (data[IFLA_VXLAN_L2MISS]) { |
3296 | if (changelink) | |
3297 | return -EOPNOTSUPP; | |
3298 | if (nla_get_u8(data[IFLA_VXLAN_L2MISS])) | |
3299 | conf->flags |= VXLAN_F_L2MISS; | |
3300 | } | |
e4f67add | 3301 | |
8bcdc4f3 RP |
3302 | if (data[IFLA_VXLAN_L3MISS]) { |
3303 | if (changelink) | |
3304 | return -EOPNOTSUPP; | |
3305 | if (nla_get_u8(data[IFLA_VXLAN_L3MISS])) | |
3306 | conf->flags |= VXLAN_F_L3MISS; | |
3307 | } | |
e4f67add | 3308 | |
8bcdc4f3 RP |
3309 | if (data[IFLA_VXLAN_LIMIT]) { |
3310 | if (changelink) | |
3311 | return -EOPNOTSUPP; | |
3312 | conf->addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]); | |
3313 | } | |
d342894c | 3314 | |
8bcdc4f3 RP |
3315 | if (data[IFLA_VXLAN_COLLECT_METADATA]) { |
3316 | if (changelink) | |
3317 | return -EOPNOTSUPP; | |
3318 | if (nla_get_u8(data[IFLA_VXLAN_COLLECT_METADATA])) | |
3319 | conf->flags |= VXLAN_F_COLLECT_METADATA; | |
3320 | } | |
f8a9b1bc | 3321 | |
05f47d69 | 3322 | if (data[IFLA_VXLAN_PORT_RANGE]) { |
8bcdc4f3 RP |
3323 | if (!changelink) { |
3324 | const struct ifla_vxlan_port_range *p | |
3325 | = nla_data(data[IFLA_VXLAN_PORT_RANGE]); | |
3326 | conf->port_min = ntohs(p->low); | |
3327 | conf->port_max = ntohs(p->high); | |
3328 | } else { | |
3329 | return -EOPNOTSUPP; | |
3330 | } | |
05f47d69 | 3331 | } |
3332 | ||
8bcdc4f3 RP |
3333 | if (data[IFLA_VXLAN_PORT]) { |
3334 | if (changelink) | |
3335 | return -EOPNOTSUPP; | |
3336 | conf->dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]); | |
3337 | } | |
823aa873 | 3338 | |
8bcdc4f3 RP |
3339 | if (data[IFLA_VXLAN_UDP_CSUM]) { |
3340 | if (changelink) | |
3341 | return -EOPNOTSUPP; | |
3342 | if (!nla_get_u8(data[IFLA_VXLAN_UDP_CSUM])) | |
3343 | conf->flags |= VXLAN_F_UDP_ZERO_CSUM_TX; | |
3344 | } | |
359a0ea9 | 3345 | |
8bcdc4f3 RP |
3346 | if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX]) { |
3347 | if (changelink) | |
3348 | return -EOPNOTSUPP; | |
3349 | if (nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX])) | |
3350 | conf->flags |= VXLAN_F_UDP_ZERO_CSUM6_TX; | |
3351 | } | |
359a0ea9 | 3352 | |
8bcdc4f3 RP |
3353 | if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX]) { |
3354 | if (changelink) | |
3355 | return -EOPNOTSUPP; | |
3356 | if (nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX])) | |
3357 | conf->flags |= VXLAN_F_UDP_ZERO_CSUM6_RX; | |
3358 | } | |
359a0ea9 | 3359 | |
8bcdc4f3 RP |
3360 | if (data[IFLA_VXLAN_REMCSUM_TX]) { |
3361 | if (changelink) | |
3362 | return -EOPNOTSUPP; | |
3363 | if (nla_get_u8(data[IFLA_VXLAN_REMCSUM_TX])) | |
3364 | conf->flags |= VXLAN_F_REMCSUM_TX; | |
3365 | } | |
dfd8645e | 3366 | |
8bcdc4f3 RP |
3367 | if (data[IFLA_VXLAN_REMCSUM_RX]) { |
3368 | if (changelink) | |
3369 | return -EOPNOTSUPP; | |
3370 | if (nla_get_u8(data[IFLA_VXLAN_REMCSUM_RX])) | |
3371 | conf->flags |= VXLAN_F_REMCSUM_RX; | |
3372 | } | |
3373 | ||
3374 | if (data[IFLA_VXLAN_GBP]) { | |
3375 | if (changelink) | |
3376 | return -EOPNOTSUPP; | |
3377 | conf->flags |= VXLAN_F_GBP; | |
3378 | } | |
3379 | ||
3380 | if (data[IFLA_VXLAN_GPE]) { | |
3381 | if (changelink) | |
3382 | return -EOPNOTSUPP; | |
3383 | conf->flags |= VXLAN_F_GPE; | |
3384 | } | |
3385 | ||
3386 | if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL]) { | |
3387 | if (changelink) | |
3388 | return -EOPNOTSUPP; | |
3389 | conf->flags |= VXLAN_F_REMCSUM_NOPARTIAL; | |
3390 | } | |
3391 | ||
3392 | if (tb[IFLA_MTU]) { | |
3393 | if (changelink) | |
3394 | return -EOPNOTSUPP; | |
3395 | conf->mtu = nla_get_u32(tb[IFLA_MTU]); | |
3396 | } | |
3397 | ||
3398 | return 0; | |
3399 | } | |
3400 | ||
3401 | static int vxlan_newlink(struct net *src_net, struct net_device *dev, | |
7a3f4a18 MS |
3402 | struct nlattr *tb[], struct nlattr *data[], |
3403 | struct netlink_ext_ack *extack) | |
8bcdc4f3 | 3404 | { |
8bcdc4f3 RP |
3405 | struct vxlan_config conf; |
3406 | int err; | |
3407 | ||
3408 | err = vxlan_nl2conf(tb, data, dev, &conf, false); | |
3409 | if (err) | |
3410 | return err; | |
3411 | ||
653ef6a3 | 3412 | return __vxlan_dev_create(src_net, dev, &conf, extack); |
8bcdc4f3 | 3413 | } |
dfd8645e | 3414 | |
8bcdc4f3 | 3415 | static int vxlan_changelink(struct net_device *dev, struct nlattr *tb[], |
ad744b22 MS |
3416 | struct nlattr *data[], |
3417 | struct netlink_ext_ack *extack) | |
8bcdc4f3 RP |
3418 | { |
3419 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3420 | struct vxlan_rdst *dst = &vxlan->default_dst; | |
3421 | struct vxlan_rdst old_dst; | |
3422 | struct vxlan_config conf; | |
3423 | int err; | |
3424 | ||
3425 | err = vxlan_nl2conf(tb, data, | |
3426 | dev, &conf, true); | |
3427 | if (err) | |
3428 | return err; | |
3511494c | 3429 | |
8bcdc4f3 | 3430 | memcpy(&old_dst, dst, sizeof(struct vxlan_rdst)); |
e1e5314d | 3431 | |
653ef6a3 | 3432 | err = vxlan_dev_configure(vxlan->net, dev, &conf, true, extack); |
8bcdc4f3 RP |
3433 | if (err) |
3434 | return err; | |
0ace2ca8 | 3435 | |
8bcdc4f3 RP |
3436 | /* handle default dst entry */ |
3437 | if (!vxlan_addr_equal(&dst->remote_ip, &old_dst.remote_ip)) { | |
3438 | spin_lock_bh(&vxlan->hash_lock); | |
3439 | if (!vxlan_addr_any(&old_dst.remote_ip)) | |
3440 | __vxlan_fdb_delete(vxlan, all_zeros_mac, | |
3441 | old_dst.remote_ip, | |
3442 | vxlan->cfg.dst_port, | |
3443 | old_dst.remote_vni, | |
3444 | old_dst.remote_vni, | |
3445 | old_dst.remote_ifindex, 0); | |
3446 | ||
3447 | if (!vxlan_addr_any(&dst->remote_ip)) { | |
3448 | err = vxlan_fdb_create(vxlan, all_zeros_mac, | |
3449 | &dst->remote_ip, | |
3450 | NUD_REACHABLE | NUD_PERMANENT, | |
3451 | NLM_F_CREATE | NLM_F_APPEND, | |
3452 | vxlan->cfg.dst_port, | |
3453 | dst->remote_vni, | |
3454 | dst->remote_vni, | |
3455 | dst->remote_ifindex, | |
3456 | NTF_SELF); | |
3457 | if (err) { | |
3458 | spin_unlock_bh(&vxlan->hash_lock); | |
3459 | return err; | |
3460 | } | |
3461 | } | |
3462 | spin_unlock_bh(&vxlan->hash_lock); | |
3463 | } | |
ce577668 | 3464 | |
8bcdc4f3 | 3465 | return 0; |
d342894c | 3466 | } |
3467 | ||
3468 | static void vxlan_dellink(struct net_device *dev, struct list_head *head) | |
3469 | { | |
3470 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3471 | ||
8b3f9337 RP |
3472 | vxlan_flush(vxlan, true); |
3473 | ||
58ce31cc | 3474 | gro_cells_destroy(&vxlan->gro_cells); |
553675fb | 3475 | list_del(&vxlan->next); |
d342894c | 3476 | unregister_netdevice_queue(dev, head); |
3477 | } | |
3478 | ||
3479 | static size_t vxlan_get_size(const struct net_device *dev) | |
3480 | { | |
3481 | ||
3482 | return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */ | |
e4c7ed41 | 3483 | nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */ |
d342894c | 3484 | nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */ |
e4c7ed41 | 3485 | nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */ |
d342894c | 3486 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */ |
3487 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */ | |
e7f70af1 | 3488 | nla_total_size(sizeof(__be32)) + /* IFLA_VXLAN_LABEL */ |
d342894c | 3489 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */ |
e4f67add DS |
3490 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */ |
3491 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */ | |
3492 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */ | |
3493 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */ | |
da8b43c0 | 3494 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_COLLECT_METADATA */ |
d342894c | 3495 | nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */ |
3496 | nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */ | |
05f47d69 | 3497 | nla_total_size(sizeof(struct ifla_vxlan_port_range)) + |
359a0ea9 TH |
3498 | nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */ |
3499 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */ | |
3500 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */ | |
3501 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */ | |
dfd8645e TH |
3502 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */ |
3503 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */ | |
d342894c | 3504 | 0; |
3505 | } | |
3506 | ||
3507 | static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev) | |
3508 | { | |
3509 | const struct vxlan_dev *vxlan = netdev_priv(dev); | |
c7995c43 | 3510 | const struct vxlan_rdst *dst = &vxlan->default_dst; |
05f47d69 | 3511 | struct ifla_vxlan_port_range ports = { |
0dfbdf41 TG |
3512 | .low = htons(vxlan->cfg.port_min), |
3513 | .high = htons(vxlan->cfg.port_max), | |
05f47d69 | 3514 | }; |
d342894c | 3515 | |
54bfd872 | 3516 | if (nla_put_u32(skb, IFLA_VXLAN_ID, be32_to_cpu(dst->remote_vni))) |
d342894c | 3517 | goto nla_put_failure; |
3518 | ||
e4c7ed41 CW |
3519 | if (!vxlan_addr_any(&dst->remote_ip)) { |
3520 | if (dst->remote_ip.sa.sa_family == AF_INET) { | |
930345ea JB |
3521 | if (nla_put_in_addr(skb, IFLA_VXLAN_GROUP, |
3522 | dst->remote_ip.sin.sin_addr.s_addr)) | |
e4c7ed41 CW |
3523 | goto nla_put_failure; |
3524 | #if IS_ENABLED(CONFIG_IPV6) | |
3525 | } else { | |
930345ea JB |
3526 | if (nla_put_in6_addr(skb, IFLA_VXLAN_GROUP6, |
3527 | &dst->remote_ip.sin6.sin6_addr)) | |
e4c7ed41 CW |
3528 | goto nla_put_failure; |
3529 | #endif | |
3530 | } | |
3531 | } | |
d342894c | 3532 | |
c7995c43 | 3533 | if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex)) |
d342894c | 3534 | goto nla_put_failure; |
3535 | ||
0dfbdf41 TG |
3536 | if (!vxlan_addr_any(&vxlan->cfg.saddr)) { |
3537 | if (vxlan->cfg.saddr.sa.sa_family == AF_INET) { | |
930345ea | 3538 | if (nla_put_in_addr(skb, IFLA_VXLAN_LOCAL, |
0dfbdf41 | 3539 | vxlan->cfg.saddr.sin.sin_addr.s_addr)) |
e4c7ed41 CW |
3540 | goto nla_put_failure; |
3541 | #if IS_ENABLED(CONFIG_IPV6) | |
3542 | } else { | |
930345ea | 3543 | if (nla_put_in6_addr(skb, IFLA_VXLAN_LOCAL6, |
0dfbdf41 | 3544 | &vxlan->cfg.saddr.sin6.sin6_addr)) |
e4c7ed41 CW |
3545 | goto nla_put_failure; |
3546 | #endif | |
3547 | } | |
3548 | } | |
d342894c | 3549 | |
0dfbdf41 TG |
3550 | if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->cfg.ttl) || |
3551 | nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->cfg.tos) || | |
e7f70af1 | 3552 | nla_put_be32(skb, IFLA_VXLAN_LABEL, vxlan->cfg.label) || |
e4f67add | 3553 | nla_put_u8(skb, IFLA_VXLAN_LEARNING, |
dc5321d7 | 3554 | !!(vxlan->cfg.flags & VXLAN_F_LEARN)) || |
e4f67add | 3555 | nla_put_u8(skb, IFLA_VXLAN_PROXY, |
dc5321d7 MS |
3556 | !!(vxlan->cfg.flags & VXLAN_F_PROXY)) || |
3557 | nla_put_u8(skb, IFLA_VXLAN_RSC, | |
3558 | !!(vxlan->cfg.flags & VXLAN_F_RSC)) || | |
e4f67add | 3559 | nla_put_u8(skb, IFLA_VXLAN_L2MISS, |
dc5321d7 | 3560 | !!(vxlan->cfg.flags & VXLAN_F_L2MISS)) || |
e4f67add | 3561 | nla_put_u8(skb, IFLA_VXLAN_L3MISS, |
dc5321d7 | 3562 | !!(vxlan->cfg.flags & VXLAN_F_L3MISS)) || |
da8b43c0 | 3563 | nla_put_u8(skb, IFLA_VXLAN_COLLECT_METADATA, |
dc5321d7 | 3564 | !!(vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA)) || |
0dfbdf41 TG |
3565 | nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->cfg.age_interval) || |
3566 | nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->cfg.addrmax) || | |
3567 | nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->cfg.dst_port) || | |
359a0ea9 | 3568 | nla_put_u8(skb, IFLA_VXLAN_UDP_CSUM, |
dc5321d7 | 3569 | !(vxlan->cfg.flags & VXLAN_F_UDP_ZERO_CSUM_TX)) || |
359a0ea9 | 3570 | nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, |
dc5321d7 | 3571 | !!(vxlan->cfg.flags & VXLAN_F_UDP_ZERO_CSUM6_TX)) || |
359a0ea9 | 3572 | nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, |
dc5321d7 | 3573 | !!(vxlan->cfg.flags & VXLAN_F_UDP_ZERO_CSUM6_RX)) || |
dfd8645e | 3574 | nla_put_u8(skb, IFLA_VXLAN_REMCSUM_TX, |
dc5321d7 | 3575 | !!(vxlan->cfg.flags & VXLAN_F_REMCSUM_TX)) || |
dfd8645e | 3576 | nla_put_u8(skb, IFLA_VXLAN_REMCSUM_RX, |
dc5321d7 | 3577 | !!(vxlan->cfg.flags & VXLAN_F_REMCSUM_RX))) |
d342894c | 3578 | goto nla_put_failure; |
3579 | ||
05f47d69 | 3580 | if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports)) |
3581 | goto nla_put_failure; | |
3582 | ||
dc5321d7 | 3583 | if (vxlan->cfg.flags & VXLAN_F_GBP && |
3511494c TG |
3584 | nla_put_flag(skb, IFLA_VXLAN_GBP)) |
3585 | goto nla_put_failure; | |
3586 | ||
dc5321d7 | 3587 | if (vxlan->cfg.flags & VXLAN_F_GPE && |
e1e5314d JB |
3588 | nla_put_flag(skb, IFLA_VXLAN_GPE)) |
3589 | goto nla_put_failure; | |
3590 | ||
dc5321d7 | 3591 | if (vxlan->cfg.flags & VXLAN_F_REMCSUM_NOPARTIAL && |
0ace2ca8 TH |
3592 | nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL)) |
3593 | goto nla_put_failure; | |
3594 | ||
d342894c | 3595 | return 0; |
3596 | ||
3597 | nla_put_failure: | |
3598 | return -EMSGSIZE; | |
3599 | } | |
3600 | ||
1728d4fa ND |
3601 | static struct net *vxlan_get_link_net(const struct net_device *dev) |
3602 | { | |
3603 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3604 | ||
3605 | return vxlan->net; | |
3606 | } | |
3607 | ||
d342894c | 3608 | static struct rtnl_link_ops vxlan_link_ops __read_mostly = { |
3609 | .kind = "vxlan", | |
3610 | .maxtype = IFLA_VXLAN_MAX, | |
3611 | .policy = vxlan_policy, | |
3612 | .priv_size = sizeof(struct vxlan_dev), | |
3613 | .setup = vxlan_setup, | |
3614 | .validate = vxlan_validate, | |
3615 | .newlink = vxlan_newlink, | |
8bcdc4f3 | 3616 | .changelink = vxlan_changelink, |
d342894c | 3617 | .dellink = vxlan_dellink, |
3618 | .get_size = vxlan_get_size, | |
3619 | .fill_info = vxlan_fill_info, | |
1728d4fa | 3620 | .get_link_net = vxlan_get_link_net, |
d342894c | 3621 | }; |
3622 | ||
cf5da330 ND |
3623 | struct net_device *vxlan_dev_create(struct net *net, const char *name, |
3624 | u8 name_assign_type, | |
3625 | struct vxlan_config *conf) | |
3626 | { | |
3627 | struct nlattr *tb[IFLA_MAX + 1]; | |
3628 | struct net_device *dev; | |
3629 | int err; | |
3630 | ||
3631 | memset(&tb, 0, sizeof(tb)); | |
3632 | ||
3633 | dev = rtnl_create_link(net, name, name_assign_type, | |
3634 | &vxlan_link_ops, tb); | |
3635 | if (IS_ERR(dev)) | |
3636 | return dev; | |
3637 | ||
653ef6a3 | 3638 | err = __vxlan_dev_create(net, dev, conf, NULL); |
cf5da330 ND |
3639 | if (err < 0) { |
3640 | free_netdev(dev); | |
3641 | return ERR_PTR(err); | |
3642 | } | |
3643 | ||
3644 | err = rtnl_configure_link(dev, NULL); | |
3645 | if (err < 0) { | |
3646 | LIST_HEAD(list_kill); | |
3647 | ||
3648 | vxlan_dellink(dev, &list_kill); | |
3649 | unregister_netdevice_many(&list_kill); | |
3650 | return ERR_PTR(err); | |
3651 | } | |
3652 | ||
3653 | return dev; | |
3654 | } | |
3655 | EXPORT_SYMBOL_GPL(vxlan_dev_create); | |
3656 | ||
acaf4e70 DB |
3657 | static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn, |
3658 | struct net_device *dev) | |
3659 | { | |
3660 | struct vxlan_dev *vxlan, *next; | |
3661 | LIST_HEAD(list_kill); | |
3662 | ||
3663 | list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) { | |
3664 | struct vxlan_rdst *dst = &vxlan->default_dst; | |
3665 | ||
3666 | /* In case we created vxlan device with carrier | |
3667 | * and we loose the carrier due to module unload | |
3668 | * we also need to remove vxlan device. In other | |
3669 | * cases, it's not necessary and remote_ifindex | |
3670 | * is 0 here, so no matches. | |
3671 | */ | |
3672 | if (dst->remote_ifindex == dev->ifindex) | |
3673 | vxlan_dellink(vxlan->dev, &list_kill); | |
3674 | } | |
3675 | ||
3676 | unregister_netdevice_many(&list_kill); | |
3677 | } | |
3678 | ||
b7aade15 HFS |
3679 | static int vxlan_netdevice_event(struct notifier_block *unused, |
3680 | unsigned long event, void *ptr) | |
acaf4e70 DB |
3681 | { |
3682 | struct net_device *dev = netdev_notifier_info_to_dev(ptr); | |
783c1463 | 3683 | struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id); |
acaf4e70 | 3684 | |
04584957 SD |
3685 | if (event == NETDEV_UNREGISTER) { |
3686 | vxlan_offload_rx_ports(dev, false); | |
acaf4e70 | 3687 | vxlan_handle_lowerdev_unregister(vn, dev); |
04584957 SD |
3688 | } else if (event == NETDEV_REGISTER) { |
3689 | vxlan_offload_rx_ports(dev, true); | |
3690 | } else if (event == NETDEV_UDP_TUNNEL_PUSH_INFO || | |
3691 | event == NETDEV_UDP_TUNNEL_DROP_INFO) { | |
2d2b13fc | 3692 | vxlan_offload_rx_ports(dev, event == NETDEV_UDP_TUNNEL_PUSH_INFO); |
04584957 | 3693 | } |
acaf4e70 DB |
3694 | |
3695 | return NOTIFY_DONE; | |
3696 | } | |
3697 | ||
3698 | static struct notifier_block vxlan_notifier_block __read_mostly = { | |
b7aade15 | 3699 | .notifier_call = vxlan_netdevice_event, |
acaf4e70 DB |
3700 | }; |
3701 | ||
d342894c | 3702 | static __net_init int vxlan_init_net(struct net *net) |
3703 | { | |
3704 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
31fec5aa | 3705 | unsigned int h; |
d342894c | 3706 | |
553675fb | 3707 | INIT_LIST_HEAD(&vn->vxlan_list); |
1c51a915 | 3708 | spin_lock_init(&vn->sock_lock); |
d342894c | 3709 | |
553675fb | 3710 | for (h = 0; h < PORT_HASH_SIZE; ++h) |
3711 | INIT_HLIST_HEAD(&vn->sock_list[h]); | |
d342894c | 3712 | |
3713 | return 0; | |
3714 | } | |
3715 | ||
f01ec1c0 ND |
3716 | static void __net_exit vxlan_exit_net(struct net *net) |
3717 | { | |
3718 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
3719 | struct vxlan_dev *vxlan, *next; | |
3720 | struct net_device *dev, *aux; | |
3721 | LIST_HEAD(list); | |
3722 | ||
3723 | rtnl_lock(); | |
3724 | for_each_netdev_safe(net, dev, aux) | |
3725 | if (dev->rtnl_link_ops == &vxlan_link_ops) | |
3726 | unregister_netdevice_queue(dev, &list); | |
3727 | ||
3728 | list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) { | |
3729 | /* If vxlan->dev is in the same netns, it has already been added | |
3730 | * to the list by the previous loop. | |
3731 | */ | |
58ce31cc TH |
3732 | if (!net_eq(dev_net(vxlan->dev), net)) { |
3733 | gro_cells_destroy(&vxlan->gro_cells); | |
13c3ed6a | 3734 | unregister_netdevice_queue(vxlan->dev, &list); |
58ce31cc | 3735 | } |
f01ec1c0 ND |
3736 | } |
3737 | ||
3738 | unregister_netdevice_many(&list); | |
3739 | rtnl_unlock(); | |
3740 | } | |
3741 | ||
d342894c | 3742 | static struct pernet_operations vxlan_net_ops = { |
3743 | .init = vxlan_init_net, | |
f01ec1c0 | 3744 | .exit = vxlan_exit_net, |
d342894c | 3745 | .id = &vxlan_net_id, |
3746 | .size = sizeof(struct vxlan_net), | |
3747 | }; | |
3748 | ||
3749 | static int __init vxlan_init_module(void) | |
3750 | { | |
3751 | int rc; | |
3752 | ||
3753 | get_random_bytes(&vxlan_salt, sizeof(vxlan_salt)); | |
3754 | ||
783c1463 | 3755 | rc = register_pernet_subsys(&vxlan_net_ops); |
d342894c | 3756 | if (rc) |
3757 | goto out1; | |
3758 | ||
acaf4e70 | 3759 | rc = register_netdevice_notifier(&vxlan_notifier_block); |
d342894c | 3760 | if (rc) |
3761 | goto out2; | |
3762 | ||
acaf4e70 DB |
3763 | rc = rtnl_link_register(&vxlan_link_ops); |
3764 | if (rc) | |
3765 | goto out3; | |
d342894c | 3766 | |
acaf4e70 DB |
3767 | return 0; |
3768 | out3: | |
3769 | unregister_netdevice_notifier(&vxlan_notifier_block); | |
d342894c | 3770 | out2: |
783c1463 | 3771 | unregister_pernet_subsys(&vxlan_net_ops); |
d342894c | 3772 | out1: |
3773 | return rc; | |
3774 | } | |
7332a13b | 3775 | late_initcall(vxlan_init_module); |
d342894c | 3776 | |
3777 | static void __exit vxlan_cleanup_module(void) | |
3778 | { | |
b7153984 | 3779 | rtnl_link_unregister(&vxlan_link_ops); |
acaf4e70 | 3780 | unregister_netdevice_notifier(&vxlan_notifier_block); |
783c1463 DB |
3781 | unregister_pernet_subsys(&vxlan_net_ops); |
3782 | /* rcu_barrier() is called by netns */ | |
d342894c | 3783 | } |
3784 | module_exit(vxlan_cleanup_module); | |
3785 | ||
3786 | MODULE_LICENSE("GPL"); | |
3787 | MODULE_VERSION(VXLAN_VERSION); | |
3b8df3c6 | 3788 | MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>"); |
ead5139a | 3789 | MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic"); |
d342894c | 3790 | MODULE_ALIAS_RTNL_LINK("vxlan"); |