]>
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> | |
012a5729 | 29 | #include <net/vxlan.h> |
40d29af0 | 30 | |
e4c7ed41 | 31 | #if IS_ENABLED(CONFIG_IPV6) |
e4c7ed41 | 32 | #include <net/ip6_tunnel.h> |
660d98ca | 33 | #include <net/ip6_checksum.h> |
e4c7ed41 | 34 | #endif |
d342894c | 35 | |
36 | #define VXLAN_VERSION "0.1" | |
37 | ||
553675fb | 38 | #define PORT_HASH_BITS 8 |
39 | #define PORT_HASH_SIZE (1<<PORT_HASH_BITS) | |
d342894c | 40 | #define FDB_AGE_DEFAULT 300 /* 5 min */ |
41 | #define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */ | |
42 | ||
23c578bf | 43 | /* UDP port for VXLAN traffic. |
44 | * The IANA assigned port is 4789, but the Linux default is 8472 | |
234f5b73 | 45 | * for compatibility with early adopters. |
23c578bf | 46 | */ |
9daaa397 SH |
47 | static unsigned short vxlan_port __read_mostly = 8472; |
48 | module_param_named(udp_port, vxlan_port, ushort, 0444); | |
d342894c | 49 | MODULE_PARM_DESC(udp_port, "Destination UDP port"); |
50 | ||
51 | static bool log_ecn_error = true; | |
52 | module_param(log_ecn_error, bool, 0644); | |
53 | MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN"); | |
54 | ||
60d9d4c6 | 55 | static int vxlan_net_id; |
0dfbdf41 | 56 | static struct rtnl_link_ops vxlan_link_ops; |
553675fb | 57 | |
7256eac1 | 58 | static const u8 all_zeros_mac[ETH_ALEN + 2]; |
afbd8bae | 59 | |
205f356d | 60 | static int vxlan_sock_add(struct vxlan_dev *vxlan); |
614732ea | 61 | |
553675fb | 62 | /* per-network namespace private data for this module */ |
63 | struct vxlan_net { | |
64 | struct list_head vxlan_list; | |
65 | struct hlist_head sock_list[PORT_HASH_SIZE]; | |
1c51a915 | 66 | spinlock_t sock_lock; |
553675fb | 67 | }; |
68 | ||
d342894c | 69 | /* Forwarding table entry */ |
70 | struct vxlan_fdb { | |
71 | struct hlist_node hlist; /* linked list of entries */ | |
72 | struct rcu_head rcu; | |
73 | unsigned long updated; /* jiffies */ | |
74 | unsigned long used; | |
3e61aa8f | 75 | struct list_head remotes; |
7177a3b0 | 76 | u8 eth_addr[ETH_ALEN]; |
d342894c | 77 | u16 state; /* see ndm_state */ |
ae884082 | 78 | u8 flags; /* see ndm_flags */ |
d342894c | 79 | }; |
80 | ||
d342894c | 81 | /* salt for hash table */ |
82 | static u32 vxlan_salt __read_mostly; | |
83 | ||
ee122c79 TG |
84 | static inline bool vxlan_collect_metadata(struct vxlan_sock *vs) |
85 | { | |
e7030878 TG |
86 | return vs->flags & VXLAN_F_COLLECT_METADATA || |
87 | ip_tunnel_collect_metadata(); | |
ee122c79 TG |
88 | } |
89 | ||
e4c7ed41 CW |
90 | #if IS_ENABLED(CONFIG_IPV6) |
91 | static inline | |
92 | bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b) | |
93 | { | |
f0ef3126 JB |
94 | if (a->sa.sa_family != b->sa.sa_family) |
95 | return false; | |
96 | if (a->sa.sa_family == AF_INET6) | |
97 | return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr); | |
98 | else | |
99 | return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr; | |
e4c7ed41 CW |
100 | } |
101 | ||
102 | static inline bool vxlan_addr_any(const union vxlan_addr *ipa) | |
103 | { | |
f0ef3126 JB |
104 | if (ipa->sa.sa_family == AF_INET6) |
105 | return ipv6_addr_any(&ipa->sin6.sin6_addr); | |
106 | else | |
107 | return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY); | |
e4c7ed41 CW |
108 | } |
109 | ||
110 | static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa) | |
111 | { | |
f0ef3126 JB |
112 | if (ipa->sa.sa_family == AF_INET6) |
113 | return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr); | |
114 | else | |
115 | return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr)); | |
e4c7ed41 CW |
116 | } |
117 | ||
118 | static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla) | |
119 | { | |
f0ef3126 | 120 | if (nla_len(nla) >= sizeof(struct in6_addr)) { |
67b61f6c | 121 | ip->sin6.sin6_addr = nla_get_in6_addr(nla); |
f0ef3126 JB |
122 | ip->sa.sa_family = AF_INET6; |
123 | return 0; | |
124 | } else if (nla_len(nla) >= sizeof(__be32)) { | |
67b61f6c | 125 | ip->sin.sin_addr.s_addr = nla_get_in_addr(nla); |
f0ef3126 JB |
126 | ip->sa.sa_family = AF_INET; |
127 | return 0; | |
128 | } else { | |
129 | return -EAFNOSUPPORT; | |
130 | } | |
e4c7ed41 CW |
131 | } |
132 | ||
133 | static int vxlan_nla_put_addr(struct sk_buff *skb, int attr, | |
f0ef3126 | 134 | const union vxlan_addr *ip) |
e4c7ed41 | 135 | { |
f0ef3126 | 136 | if (ip->sa.sa_family == AF_INET6) |
930345ea | 137 | return nla_put_in6_addr(skb, attr, &ip->sin6.sin6_addr); |
f0ef3126 | 138 | else |
930345ea | 139 | return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr); |
e4c7ed41 CW |
140 | } |
141 | ||
142 | #else /* !CONFIG_IPV6 */ | |
143 | ||
144 | static inline | |
145 | bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b) | |
146 | { | |
f0ef3126 | 147 | return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr; |
e4c7ed41 CW |
148 | } |
149 | ||
150 | static inline bool vxlan_addr_any(const union vxlan_addr *ipa) | |
151 | { | |
f0ef3126 | 152 | return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY); |
e4c7ed41 CW |
153 | } |
154 | ||
155 | static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa) | |
156 | { | |
f0ef3126 | 157 | return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr)); |
e4c7ed41 CW |
158 | } |
159 | ||
160 | static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla) | |
161 | { | |
f0ef3126 JB |
162 | if (nla_len(nla) >= sizeof(struct in6_addr)) { |
163 | return -EAFNOSUPPORT; | |
164 | } else if (nla_len(nla) >= sizeof(__be32)) { | |
67b61f6c | 165 | ip->sin.sin_addr.s_addr = nla_get_in_addr(nla); |
f0ef3126 JB |
166 | ip->sa.sa_family = AF_INET; |
167 | return 0; | |
168 | } else { | |
169 | return -EAFNOSUPPORT; | |
170 | } | |
e4c7ed41 CW |
171 | } |
172 | ||
173 | static int vxlan_nla_put_addr(struct sk_buff *skb, int attr, | |
f0ef3126 | 174 | const union vxlan_addr *ip) |
e4c7ed41 | 175 | { |
930345ea | 176 | return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr); |
e4c7ed41 CW |
177 | } |
178 | #endif | |
179 | ||
553675fb | 180 | /* Virtual Network hash table head */ |
54bfd872 | 181 | static inline struct hlist_head *vni_head(struct vxlan_sock *vs, __be32 vni) |
553675fb | 182 | { |
54bfd872 | 183 | return &vs->vni_list[hash_32((__force u32)vni, VNI_HASH_BITS)]; |
553675fb | 184 | } |
185 | ||
186 | /* Socket hash table head */ | |
187 | static inline struct hlist_head *vs_head(struct net *net, __be16 port) | |
d342894c | 188 | { |
189 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
190 | ||
553675fb | 191 | return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)]; |
192 | } | |
193 | ||
3e61aa8f SH |
194 | /* First remote destination for a forwarding entry. |
195 | * Guaranteed to be non-NULL because remotes are never deleted. | |
196 | */ | |
5ca5461c | 197 | static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb) |
3e61aa8f | 198 | { |
5ca5461c | 199 | return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list); |
200 | } | |
201 | ||
202 | static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb) | |
203 | { | |
204 | return list_first_entry(&fdb->remotes, struct vxlan_rdst, list); | |
3e61aa8f SH |
205 | } |
206 | ||
ac5132d1 TG |
207 | /* Find VXLAN socket based on network namespace, address family and UDP port |
208 | * and enabled unshareable flags. | |
209 | */ | |
210 | static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family, | |
211 | __be16 port, u32 flags) | |
553675fb | 212 | { |
213 | struct vxlan_sock *vs; | |
af33c1ad TH |
214 | |
215 | flags &= VXLAN_F_RCV_FLAGS; | |
553675fb | 216 | |
217 | hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) { | |
19ca9fc1 | 218 | if (inet_sk(vs->sock->sk)->inet_sport == port && |
705cc62f | 219 | vxlan_get_sk_family(vs) == family && |
af33c1ad | 220 | vs->flags == flags) |
553675fb | 221 | return vs; |
222 | } | |
223 | return NULL; | |
d342894c | 224 | } |
225 | ||
54bfd872 | 226 | static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, __be32 vni) |
d342894c | 227 | { |
228 | struct vxlan_dev *vxlan; | |
d342894c | 229 | |
b9167b2e JB |
230 | /* For flow based devices, map all packets to VNI 0 */ |
231 | if (vs->flags & VXLAN_F_COLLECT_METADATA) | |
232 | vni = 0; | |
233 | ||
54bfd872 JB |
234 | hlist_for_each_entry_rcu(vxlan, vni_head(vs, vni), hlist) { |
235 | if (vxlan->default_dst.remote_vni == vni) | |
d342894c | 236 | return vxlan; |
237 | } | |
238 | ||
239 | return NULL; | |
240 | } | |
241 | ||
5cfccc5a | 242 | /* Look up VNI in a per net namespace table */ |
54bfd872 | 243 | static struct vxlan_dev *vxlan_find_vni(struct net *net, __be32 vni, |
ac5132d1 TG |
244 | sa_family_t family, __be16 port, |
245 | u32 flags) | |
5cfccc5a PS |
246 | { |
247 | struct vxlan_sock *vs; | |
248 | ||
ac5132d1 | 249 | vs = vxlan_find_sock(net, family, port, flags); |
5cfccc5a PS |
250 | if (!vs) |
251 | return NULL; | |
252 | ||
54bfd872 | 253 | return vxlan_vs_find_vni(vs, vni); |
5cfccc5a PS |
254 | } |
255 | ||
d342894c | 256 | /* Fill in neighbour message in skbuff. */ |
257 | static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan, | |
234f5b73 SH |
258 | const struct vxlan_fdb *fdb, |
259 | u32 portid, u32 seq, int type, unsigned int flags, | |
260 | const struct vxlan_rdst *rdst) | |
d342894c | 261 | { |
262 | unsigned long now = jiffies; | |
263 | struct nda_cacheinfo ci; | |
264 | struct nlmsghdr *nlh; | |
265 | struct ndmsg *ndm; | |
e4f67add | 266 | bool send_ip, send_eth; |
d342894c | 267 | |
268 | nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags); | |
269 | if (nlh == NULL) | |
270 | return -EMSGSIZE; | |
271 | ||
272 | ndm = nlmsg_data(nlh); | |
273 | memset(ndm, 0, sizeof(*ndm)); | |
e4f67add DS |
274 | |
275 | send_eth = send_ip = true; | |
276 | ||
277 | if (type == RTM_GETNEIGH) { | |
278 | ndm->ndm_family = AF_INET; | |
e4c7ed41 | 279 | send_ip = !vxlan_addr_any(&rdst->remote_ip); |
e4f67add DS |
280 | send_eth = !is_zero_ether_addr(fdb->eth_addr); |
281 | } else | |
282 | ndm->ndm_family = AF_BRIDGE; | |
d342894c | 283 | ndm->ndm_state = fdb->state; |
284 | ndm->ndm_ifindex = vxlan->dev->ifindex; | |
ae884082 | 285 | ndm->ndm_flags = fdb->flags; |
545469f7 | 286 | ndm->ndm_type = RTN_UNICAST; |
d342894c | 287 | |
193523bf | 288 | if (!net_eq(dev_net(vxlan->dev), vxlan->net) && |
4967082b | 289 | nla_put_s32(skb, NDA_LINK_NETNSID, |
38f507f1 | 290 | peernet2id(dev_net(vxlan->dev), vxlan->net))) |
193523bf ND |
291 | goto nla_put_failure; |
292 | ||
e4f67add | 293 | if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr)) |
d342894c | 294 | goto nla_put_failure; |
295 | ||
e4c7ed41 | 296 | if (send_ip && vxlan_nla_put_addr(skb, NDA_DST, &rdst->remote_ip)) |
6681712d DS |
297 | goto nla_put_failure; |
298 | ||
0dfbdf41 | 299 | if (rdst->remote_port && rdst->remote_port != vxlan->cfg.dst_port && |
6681712d DS |
300 | nla_put_be16(skb, NDA_PORT, rdst->remote_port)) |
301 | goto nla_put_failure; | |
c7995c43 | 302 | if (rdst->remote_vni != vxlan->default_dst.remote_vni && |
54bfd872 | 303 | nla_put_u32(skb, NDA_VNI, be32_to_cpu(rdst->remote_vni))) |
6681712d DS |
304 | goto nla_put_failure; |
305 | if (rdst->remote_ifindex && | |
306 | nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex)) | |
d342894c | 307 | goto nla_put_failure; |
308 | ||
309 | ci.ndm_used = jiffies_to_clock_t(now - fdb->used); | |
310 | ci.ndm_confirmed = 0; | |
311 | ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated); | |
312 | ci.ndm_refcnt = 0; | |
313 | ||
314 | if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci)) | |
315 | goto nla_put_failure; | |
316 | ||
053c095a JB |
317 | nlmsg_end(skb, nlh); |
318 | return 0; | |
d342894c | 319 | |
320 | nla_put_failure: | |
321 | nlmsg_cancel(skb, nlh); | |
322 | return -EMSGSIZE; | |
323 | } | |
324 | ||
325 | static inline size_t vxlan_nlmsg_size(void) | |
326 | { | |
327 | return NLMSG_ALIGN(sizeof(struct ndmsg)) | |
328 | + nla_total_size(ETH_ALEN) /* NDA_LLADDR */ | |
e4c7ed41 | 329 | + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */ |
73cf3317 | 330 | + nla_total_size(sizeof(__be16)) /* NDA_PORT */ |
6681712d DS |
331 | + nla_total_size(sizeof(__be32)) /* NDA_VNI */ |
332 | + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */ | |
4967082b | 333 | + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */ |
d342894c | 334 | + nla_total_size(sizeof(struct nda_cacheinfo)); |
335 | } | |
336 | ||
9e4b93f9 ND |
337 | static void vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb, |
338 | struct vxlan_rdst *rd, int type) | |
d342894c | 339 | { |
340 | struct net *net = dev_net(vxlan->dev); | |
341 | struct sk_buff *skb; | |
342 | int err = -ENOBUFS; | |
343 | ||
344 | skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC); | |
345 | if (skb == NULL) | |
346 | goto errout; | |
347 | ||
9e4b93f9 | 348 | err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd); |
d342894c | 349 | if (err < 0) { |
350 | /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */ | |
351 | WARN_ON(err == -EMSGSIZE); | |
352 | kfree_skb(skb); | |
353 | goto errout; | |
354 | } | |
355 | ||
356 | rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC); | |
357 | return; | |
358 | errout: | |
359 | if (err < 0) | |
360 | rtnl_set_sk_err(net, RTNLGRP_NEIGH, err); | |
361 | } | |
362 | ||
e4c7ed41 | 363 | static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa) |
e4f67add DS |
364 | { |
365 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
bb3fd687 SH |
366 | struct vxlan_fdb f = { |
367 | .state = NUD_STALE, | |
368 | }; | |
369 | struct vxlan_rdst remote = { | |
e4c7ed41 | 370 | .remote_ip = *ipa, /* goes to NDA_DST */ |
54bfd872 | 371 | .remote_vni = cpu_to_be32(VXLAN_N_VID), |
bb3fd687 | 372 | }; |
3e61aa8f | 373 | |
9e4b93f9 | 374 | vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH); |
e4f67add DS |
375 | } |
376 | ||
377 | static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN]) | |
378 | { | |
bb3fd687 SH |
379 | struct vxlan_fdb f = { |
380 | .state = NUD_STALE, | |
381 | }; | |
9e4b93f9 | 382 | struct vxlan_rdst remote = { }; |
e4f67add | 383 | |
e4f67add DS |
384 | memcpy(f.eth_addr, eth_addr, ETH_ALEN); |
385 | ||
9e4b93f9 | 386 | vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH); |
e4f67add DS |
387 | } |
388 | ||
d342894c | 389 | /* Hash Ethernet address */ |
390 | static u32 eth_hash(const unsigned char *addr) | |
391 | { | |
392 | u64 value = get_unaligned((u64 *)addr); | |
393 | ||
394 | /* only want 6 bytes */ | |
395 | #ifdef __BIG_ENDIAN | |
d342894c | 396 | value >>= 16; |
321fb991 | 397 | #else |
398 | value <<= 16; | |
d342894c | 399 | #endif |
400 | return hash_64(value, FDB_HASH_BITS); | |
401 | } | |
402 | ||
403 | /* Hash chain to use given mac address */ | |
404 | static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan, | |
405 | const u8 *mac) | |
406 | { | |
407 | return &vxlan->fdb_head[eth_hash(mac)]; | |
408 | } | |
409 | ||
410 | /* Look up Ethernet address in forwarding table */ | |
014be2c8 | 411 | static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan, |
d342894c | 412 | const u8 *mac) |
d342894c | 413 | { |
414 | struct hlist_head *head = vxlan_fdb_head(vxlan, mac); | |
415 | struct vxlan_fdb *f; | |
d342894c | 416 | |
b67bfe0d | 417 | hlist_for_each_entry_rcu(f, head, hlist) { |
7367d0b5 | 418 | if (ether_addr_equal(mac, f->eth_addr)) |
d342894c | 419 | return f; |
420 | } | |
421 | ||
422 | return NULL; | |
423 | } | |
424 | ||
014be2c8 SS |
425 | static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan, |
426 | const u8 *mac) | |
427 | { | |
428 | struct vxlan_fdb *f; | |
429 | ||
430 | f = __vxlan_find_mac(vxlan, mac); | |
431 | if (f) | |
432 | f->used = jiffies; | |
433 | ||
434 | return f; | |
435 | } | |
436 | ||
a5e7c10a MR |
437 | /* caller should hold vxlan->hash_lock */ |
438 | static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f, | |
e4c7ed41 | 439 | union vxlan_addr *ip, __be16 port, |
54bfd872 | 440 | __be32 vni, __u32 ifindex) |
6681712d | 441 | { |
3e61aa8f | 442 | struct vxlan_rdst *rd; |
6681712d | 443 | |
3e61aa8f | 444 | list_for_each_entry(rd, &f->remotes, list) { |
e4c7ed41 | 445 | if (vxlan_addr_equal(&rd->remote_ip, ip) && |
6681712d DS |
446 | rd->remote_port == port && |
447 | rd->remote_vni == vni && | |
448 | rd->remote_ifindex == ifindex) | |
a5e7c10a | 449 | return rd; |
6681712d | 450 | } |
3e61aa8f | 451 | |
a5e7c10a MR |
452 | return NULL; |
453 | } | |
454 | ||
906dc186 TR |
455 | /* Replace destination of unicast mac */ |
456 | static int vxlan_fdb_replace(struct vxlan_fdb *f, | |
54bfd872 JB |
457 | union vxlan_addr *ip, __be16 port, __be32 vni, |
458 | __u32 ifindex) | |
906dc186 TR |
459 | { |
460 | struct vxlan_rdst *rd; | |
461 | ||
462 | rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex); | |
463 | if (rd) | |
464 | return 0; | |
465 | ||
466 | rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list); | |
467 | if (!rd) | |
468 | return 0; | |
0c1d70af PA |
469 | |
470 | dst_cache_reset(&rd->dst_cache); | |
e4c7ed41 | 471 | rd->remote_ip = *ip; |
906dc186 TR |
472 | rd->remote_port = port; |
473 | rd->remote_vni = vni; | |
474 | rd->remote_ifindex = ifindex; | |
475 | return 1; | |
476 | } | |
477 | ||
a5e7c10a MR |
478 | /* Add/update destinations for multicast */ |
479 | static int vxlan_fdb_append(struct vxlan_fdb *f, | |
54bfd872 | 480 | union vxlan_addr *ip, __be16 port, __be32 vni, |
9e4b93f9 | 481 | __u32 ifindex, struct vxlan_rdst **rdp) |
a5e7c10a MR |
482 | { |
483 | struct vxlan_rdst *rd; | |
484 | ||
485 | rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex); | |
486 | if (rd) | |
487 | return 0; | |
488 | ||
6681712d DS |
489 | rd = kmalloc(sizeof(*rd), GFP_ATOMIC); |
490 | if (rd == NULL) | |
491 | return -ENOBUFS; | |
0c1d70af PA |
492 | |
493 | if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) { | |
494 | kfree(rd); | |
495 | return -ENOBUFS; | |
496 | } | |
497 | ||
e4c7ed41 | 498 | rd->remote_ip = *ip; |
6681712d DS |
499 | rd->remote_port = port; |
500 | rd->remote_vni = vni; | |
501 | rd->remote_ifindex = ifindex; | |
3e61aa8f SH |
502 | |
503 | list_add_tail_rcu(&rd->list, &f->remotes); | |
504 | ||
9e4b93f9 | 505 | *rdp = rd; |
6681712d DS |
506 | return 1; |
507 | } | |
508 | ||
dfd8645e TH |
509 | static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb, |
510 | unsigned int off, | |
511 | struct vxlanhdr *vh, size_t hdrlen, | |
54bfd872 JB |
512 | __be32 vni_field, |
513 | struct gro_remcsum *grc, | |
0ace2ca8 | 514 | bool nopartial) |
dfd8645e | 515 | { |
b7fe10e5 | 516 | size_t start, offset; |
dfd8645e TH |
517 | |
518 | if (skb->remcsum_offload) | |
b7fe10e5 | 519 | return vh; |
dfd8645e TH |
520 | |
521 | if (!NAPI_GRO_CB(skb)->csum_valid) | |
522 | return NULL; | |
523 | ||
54bfd872 JB |
524 | start = vxlan_rco_start(vni_field); |
525 | offset = start + vxlan_rco_offset(vni_field); | |
dfd8645e | 526 | |
b7fe10e5 TH |
527 | vh = skb_gro_remcsum_process(skb, (void *)vh, off, hdrlen, |
528 | start, offset, grc, nopartial); | |
dfd8645e TH |
529 | |
530 | skb->remcsum_offload = 1; | |
531 | ||
532 | return vh; | |
533 | } | |
534 | ||
5602c48c TH |
535 | static struct sk_buff **vxlan_gro_receive(struct sock *sk, |
536 | struct sk_buff **head, | |
537 | struct sk_buff *skb) | |
dc01e7d3 OG |
538 | { |
539 | struct sk_buff *p, **pp = NULL; | |
540 | struct vxlanhdr *vh, *vh2; | |
9b174d88 | 541 | unsigned int hlen, off_vx; |
dc01e7d3 | 542 | int flush = 1; |
5602c48c | 543 | struct vxlan_sock *vs = rcu_dereference_sk_user_data(sk); |
54bfd872 | 544 | __be32 flags; |
26c4f7da TH |
545 | struct gro_remcsum grc; |
546 | ||
547 | skb_gro_remcsum_init(&grc); | |
dc01e7d3 OG |
548 | |
549 | off_vx = skb_gro_offset(skb); | |
550 | hlen = off_vx + sizeof(*vh); | |
551 | vh = skb_gro_header_fast(skb, off_vx); | |
552 | if (skb_gro_header_hard(skb, hlen)) { | |
553 | vh = skb_gro_header_slow(skb, hlen, off_vx); | |
554 | if (unlikely(!vh)) | |
555 | goto out; | |
556 | } | |
dc01e7d3 | 557 | |
dfd8645e TH |
558 | skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr)); |
559 | ||
54bfd872 | 560 | flags = vh->vx_flags; |
dfd8645e TH |
561 | |
562 | if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) { | |
563 | vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr), | |
54bfd872 | 564 | vh->vx_vni, &grc, |
0ace2ca8 TH |
565 | !!(vs->flags & |
566 | VXLAN_F_REMCSUM_NOPARTIAL)); | |
dfd8645e TH |
567 | |
568 | if (!vh) | |
569 | goto out; | |
570 | } | |
571 | ||
b7fe10e5 TH |
572 | skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */ |
573 | ||
dc01e7d3 OG |
574 | for (p = *head; p; p = p->next) { |
575 | if (!NAPI_GRO_CB(p)->same_flow) | |
576 | continue; | |
577 | ||
578 | vh2 = (struct vxlanhdr *)(p->data + off_vx); | |
3511494c TG |
579 | if (vh->vx_flags != vh2->vx_flags || |
580 | vh->vx_vni != vh2->vx_vni) { | |
dc01e7d3 OG |
581 | NAPI_GRO_CB(p)->same_flow = 0; |
582 | continue; | |
583 | } | |
dc01e7d3 OG |
584 | } |
585 | ||
9b174d88 | 586 | pp = eth_gro_receive(head, skb); |
c194cf93 | 587 | flush = 0; |
dc01e7d3 | 588 | |
dc01e7d3 | 589 | out: |
26c4f7da | 590 | skb_gro_remcsum_cleanup(skb, &grc); |
dc01e7d3 OG |
591 | NAPI_GRO_CB(skb)->flush |= flush; |
592 | ||
593 | return pp; | |
594 | } | |
595 | ||
5602c48c | 596 | static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff) |
dc01e7d3 | 597 | { |
229740c6 JR |
598 | /* Sets 'skb->inner_mac_header' since we are always called with |
599 | * 'skb->encapsulation' set. | |
600 | */ | |
9b174d88 | 601 | return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr)); |
dc01e7d3 OG |
602 | } |
603 | ||
d342894c | 604 | /* Add new entry to forwarding table -- assumes lock held */ |
605 | static int vxlan_fdb_create(struct vxlan_dev *vxlan, | |
e4c7ed41 | 606 | const u8 *mac, union vxlan_addr *ip, |
6681712d | 607 | __u16 state, __u16 flags, |
54bfd872 | 608 | __be16 port, __be32 vni, __u32 ifindex, |
ae884082 | 609 | __u8 ndm_flags) |
d342894c | 610 | { |
9e4b93f9 | 611 | struct vxlan_rdst *rd = NULL; |
d342894c | 612 | struct vxlan_fdb *f; |
613 | int notify = 0; | |
614 | ||
014be2c8 | 615 | f = __vxlan_find_mac(vxlan, mac); |
d342894c | 616 | if (f) { |
617 | if (flags & NLM_F_EXCL) { | |
618 | netdev_dbg(vxlan->dev, | |
619 | "lost race to create %pM\n", mac); | |
620 | return -EEXIST; | |
621 | } | |
622 | if (f->state != state) { | |
623 | f->state = state; | |
624 | f->updated = jiffies; | |
625 | notify = 1; | |
626 | } | |
ae884082 DS |
627 | if (f->flags != ndm_flags) { |
628 | f->flags = ndm_flags; | |
629 | f->updated = jiffies; | |
630 | notify = 1; | |
631 | } | |
906dc186 TR |
632 | if ((flags & NLM_F_REPLACE)) { |
633 | /* Only change unicasts */ | |
634 | if (!(is_multicast_ether_addr(f->eth_addr) || | |
635 | is_zero_ether_addr(f->eth_addr))) { | |
60840429 | 636 | notify |= vxlan_fdb_replace(f, ip, port, vni, |
906dc186 | 637 | ifindex); |
906dc186 TR |
638 | } else |
639 | return -EOPNOTSUPP; | |
640 | } | |
6681712d | 641 | if ((flags & NLM_F_APPEND) && |
58e4c767 MR |
642 | (is_multicast_ether_addr(f->eth_addr) || |
643 | is_zero_ether_addr(f->eth_addr))) { | |
9e4b93f9 ND |
644 | int rc = vxlan_fdb_append(f, ip, port, vni, ifindex, |
645 | &rd); | |
6681712d DS |
646 | |
647 | if (rc < 0) | |
648 | return rc; | |
649 | notify |= rc; | |
650 | } | |
d342894c | 651 | } else { |
652 | if (!(flags & NLM_F_CREATE)) | |
653 | return -ENOENT; | |
654 | ||
0dfbdf41 TG |
655 | if (vxlan->cfg.addrmax && |
656 | vxlan->addrcnt >= vxlan->cfg.addrmax) | |
d342894c | 657 | return -ENOSPC; |
658 | ||
906dc186 TR |
659 | /* Disallow replace to add a multicast entry */ |
660 | if ((flags & NLM_F_REPLACE) && | |
661 | (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac))) | |
662 | return -EOPNOTSUPP; | |
663 | ||
e4c7ed41 | 664 | netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip); |
d342894c | 665 | f = kmalloc(sizeof(*f), GFP_ATOMIC); |
666 | if (!f) | |
667 | return -ENOMEM; | |
668 | ||
669 | notify = 1; | |
d342894c | 670 | f->state = state; |
ae884082 | 671 | f->flags = ndm_flags; |
d342894c | 672 | f->updated = f->used = jiffies; |
3e61aa8f | 673 | INIT_LIST_HEAD(&f->remotes); |
d342894c | 674 | memcpy(f->eth_addr, mac, ETH_ALEN); |
675 | ||
9e4b93f9 | 676 | vxlan_fdb_append(f, ip, port, vni, ifindex, &rd); |
3e61aa8f | 677 | |
d342894c | 678 | ++vxlan->addrcnt; |
679 | hlist_add_head_rcu(&f->hlist, | |
680 | vxlan_fdb_head(vxlan, mac)); | |
681 | } | |
682 | ||
9e4b93f9 ND |
683 | if (notify) { |
684 | if (rd == NULL) | |
685 | rd = first_remote_rtnl(f); | |
686 | vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH); | |
687 | } | |
d342894c | 688 | |
689 | return 0; | |
690 | } | |
691 | ||
6706c82e | 692 | static void vxlan_fdb_free(struct rcu_head *head) |
6681712d DS |
693 | { |
694 | struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu); | |
3e61aa8f | 695 | struct vxlan_rdst *rd, *nd; |
6681712d | 696 | |
0c1d70af PA |
697 | list_for_each_entry_safe(rd, nd, &f->remotes, list) { |
698 | dst_cache_destroy(&rd->dst_cache); | |
6681712d | 699 | kfree(rd); |
0c1d70af | 700 | } |
6681712d DS |
701 | kfree(f); |
702 | } | |
703 | ||
d342894c | 704 | static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f) |
705 | { | |
706 | netdev_dbg(vxlan->dev, | |
707 | "delete %pM\n", f->eth_addr); | |
708 | ||
709 | --vxlan->addrcnt; | |
9e4b93f9 | 710 | vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_DELNEIGH); |
d342894c | 711 | |
712 | hlist_del_rcu(&f->hlist); | |
6681712d | 713 | call_rcu(&f->rcu, vxlan_fdb_free); |
d342894c | 714 | } |
715 | ||
f0b074be | 716 | static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan, |
54bfd872 JB |
717 | union vxlan_addr *ip, __be16 *port, __be32 *vni, |
718 | u32 *ifindex) | |
d342894c | 719 | { |
6681712d | 720 | struct net *net = dev_net(vxlan->dev); |
e4c7ed41 | 721 | int err; |
d342894c | 722 | |
f0b074be | 723 | if (tb[NDA_DST]) { |
e4c7ed41 CW |
724 | err = vxlan_nla_get_addr(ip, tb[NDA_DST]); |
725 | if (err) | |
726 | return err; | |
f0b074be | 727 | } else { |
e4c7ed41 CW |
728 | union vxlan_addr *remote = &vxlan->default_dst.remote_ip; |
729 | if (remote->sa.sa_family == AF_INET) { | |
730 | ip->sin.sin_addr.s_addr = htonl(INADDR_ANY); | |
731 | ip->sa.sa_family = AF_INET; | |
732 | #if IS_ENABLED(CONFIG_IPV6) | |
733 | } else { | |
734 | ip->sin6.sin6_addr = in6addr_any; | |
735 | ip->sa.sa_family = AF_INET6; | |
736 | #endif | |
737 | } | |
f0b074be | 738 | } |
d342894c | 739 | |
6681712d | 740 | if (tb[NDA_PORT]) { |
73cf3317 | 741 | if (nla_len(tb[NDA_PORT]) != sizeof(__be16)) |
6681712d | 742 | return -EINVAL; |
f0b074be MR |
743 | *port = nla_get_be16(tb[NDA_PORT]); |
744 | } else { | |
0dfbdf41 | 745 | *port = vxlan->cfg.dst_port; |
f0b074be | 746 | } |
6681712d DS |
747 | |
748 | if (tb[NDA_VNI]) { | |
749 | if (nla_len(tb[NDA_VNI]) != sizeof(u32)) | |
750 | return -EINVAL; | |
54bfd872 | 751 | *vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI])); |
f0b074be MR |
752 | } else { |
753 | *vni = vxlan->default_dst.remote_vni; | |
754 | } | |
6681712d DS |
755 | |
756 | if (tb[NDA_IFINDEX]) { | |
5abb0029 | 757 | struct net_device *tdev; |
6681712d DS |
758 | |
759 | if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32)) | |
760 | return -EINVAL; | |
f0b074be | 761 | *ifindex = nla_get_u32(tb[NDA_IFINDEX]); |
73763949 | 762 | tdev = __dev_get_by_index(net, *ifindex); |
5abb0029 | 763 | if (!tdev) |
6681712d | 764 | return -EADDRNOTAVAIL; |
f0b074be MR |
765 | } else { |
766 | *ifindex = 0; | |
767 | } | |
768 | ||
769 | return 0; | |
770 | } | |
771 | ||
772 | /* Add static entry (via netlink) */ | |
773 | static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[], | |
774 | struct net_device *dev, | |
f6f6424b | 775 | const unsigned char *addr, u16 vid, u16 flags) |
f0b074be MR |
776 | { |
777 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
778 | /* struct net *net = dev_net(vxlan->dev); */ | |
e4c7ed41 | 779 | union vxlan_addr ip; |
f0b074be | 780 | __be16 port; |
54bfd872 JB |
781 | __be32 vni; |
782 | u32 ifindex; | |
f0b074be MR |
783 | int err; |
784 | ||
785 | if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) { | |
786 | pr_info("RTM_NEWNEIGH with invalid state %#x\n", | |
787 | ndm->ndm_state); | |
788 | return -EINVAL; | |
789 | } | |
790 | ||
791 | if (tb[NDA_DST] == NULL) | |
792 | return -EINVAL; | |
793 | ||
794 | err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex); | |
795 | if (err) | |
796 | return err; | |
6681712d | 797 | |
5933a7bb MR |
798 | if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family) |
799 | return -EAFNOSUPPORT; | |
800 | ||
d342894c | 801 | spin_lock_bh(&vxlan->hash_lock); |
e4c7ed41 | 802 | err = vxlan_fdb_create(vxlan, addr, &ip, ndm->ndm_state, flags, |
73cf3317 | 803 | port, vni, ifindex, ndm->ndm_flags); |
d342894c | 804 | spin_unlock_bh(&vxlan->hash_lock); |
805 | ||
806 | return err; | |
807 | } | |
808 | ||
809 | /* Delete entry (via netlink) */ | |
1690be63 VY |
810 | static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[], |
811 | struct net_device *dev, | |
f6f6424b | 812 | const unsigned char *addr, u16 vid) |
d342894c | 813 | { |
814 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
815 | struct vxlan_fdb *f; | |
bc7892ba | 816 | struct vxlan_rdst *rd = NULL; |
e4c7ed41 | 817 | union vxlan_addr ip; |
bc7892ba | 818 | __be16 port; |
54bfd872 JB |
819 | __be32 vni; |
820 | u32 ifindex; | |
bc7892ba MR |
821 | int err; |
822 | ||
823 | err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex); | |
824 | if (err) | |
825 | return err; | |
826 | ||
827 | err = -ENOENT; | |
d342894c | 828 | |
829 | spin_lock_bh(&vxlan->hash_lock); | |
830 | f = vxlan_find_mac(vxlan, addr); | |
bc7892ba MR |
831 | if (!f) |
832 | goto out; | |
833 | ||
e4c7ed41 CW |
834 | if (!vxlan_addr_any(&ip)) { |
835 | rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex); | |
bc7892ba MR |
836 | if (!rd) |
837 | goto out; | |
838 | } | |
839 | ||
840 | err = 0; | |
841 | ||
842 | /* remove a destination if it's not the only one on the list, | |
843 | * otherwise destroy the fdb entry | |
844 | */ | |
845 | if (rd && !list_is_singular(&f->remotes)) { | |
846 | list_del_rcu(&rd->list); | |
9e4b93f9 | 847 | vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH); |
dcdc7a59 | 848 | kfree_rcu(rd, rcu); |
bc7892ba | 849 | goto out; |
d342894c | 850 | } |
bc7892ba MR |
851 | |
852 | vxlan_fdb_destroy(vxlan, f); | |
853 | ||
854 | out: | |
d342894c | 855 | spin_unlock_bh(&vxlan->hash_lock); |
856 | ||
857 | return err; | |
858 | } | |
859 | ||
860 | /* Dump forwarding table */ | |
861 | static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb, | |
5d5eacb3 | 862 | struct net_device *dev, |
d297653d | 863 | struct net_device *filter_dev, int *idx) |
d342894c | 864 | { |
865 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
866 | unsigned int h; | |
d297653d | 867 | int err = 0; |
d342894c | 868 | |
869 | for (h = 0; h < FDB_HASH_SIZE; ++h) { | |
870 | struct vxlan_fdb *f; | |
d342894c | 871 | |
b67bfe0d | 872 | hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) { |
6681712d | 873 | struct vxlan_rdst *rd; |
6681712d | 874 | |
3e61aa8f | 875 | list_for_each_entry_rcu(rd, &f->remotes, list) { |
d297653d | 876 | if (*idx < cb->args[2]) |
07a51cd3 AW |
877 | goto skip; |
878 | ||
6681712d DS |
879 | err = vxlan_fdb_info(skb, vxlan, f, |
880 | NETLINK_CB(cb->skb).portid, | |
881 | cb->nlh->nlmsg_seq, | |
882 | RTM_NEWNEIGH, | |
883 | NLM_F_MULTI, rd); | |
d297653d | 884 | if (err < 0) |
3e61aa8f | 885 | goto out; |
3e61aa8f | 886 | skip: |
d297653d | 887 | *idx += 1; |
07a51cd3 | 888 | } |
d342894c | 889 | } |
890 | } | |
3e61aa8f | 891 | out: |
d297653d | 892 | return err; |
d342894c | 893 | } |
894 | ||
895 | /* Watch incoming packets to learn mapping between Ethernet address | |
896 | * and Tunnel endpoint. | |
c4b49512 | 897 | * Return true if packet is bogus and should be dropped. |
d342894c | 898 | */ |
26a41ae6 | 899 | static bool vxlan_snoop(struct net_device *dev, |
e4c7ed41 | 900 | union vxlan_addr *src_ip, const u8 *src_mac) |
d342894c | 901 | { |
902 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
903 | struct vxlan_fdb *f; | |
d342894c | 904 | |
905 | f = vxlan_find_mac(vxlan, src_mac); | |
906 | if (likely(f)) { | |
5ca5461c | 907 | struct vxlan_rdst *rdst = first_remote_rcu(f); |
3e61aa8f | 908 | |
e4c7ed41 | 909 | if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip))) |
26a41ae6 | 910 | return false; |
911 | ||
912 | /* Don't migrate static entries, drop packets */ | |
eb064c3b | 913 | if (f->state & NUD_NOARP) |
26a41ae6 | 914 | return true; |
d342894c | 915 | |
916 | if (net_ratelimit()) | |
917 | netdev_info(dev, | |
e4c7ed41 | 918 | "%pM migrated from %pIS to %pIS\n", |
a4870f79 | 919 | src_mac, &rdst->remote_ip.sa, &src_ip->sa); |
d342894c | 920 | |
e4c7ed41 | 921 | rdst->remote_ip = *src_ip; |
d342894c | 922 | f->updated = jiffies; |
9e4b93f9 | 923 | vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH); |
d342894c | 924 | } else { |
925 | /* learned new entry */ | |
926 | spin_lock(&vxlan->hash_lock); | |
3bf74b1a | 927 | |
928 | /* close off race between vxlan_flush and incoming packets */ | |
929 | if (netif_running(dev)) | |
930 | vxlan_fdb_create(vxlan, src_mac, src_ip, | |
931 | NUD_REACHABLE, | |
932 | NLM_F_EXCL|NLM_F_CREATE, | |
0dfbdf41 | 933 | vxlan->cfg.dst_port, |
3bf74b1a | 934 | vxlan->default_dst.remote_vni, |
935 | 0, NTF_SELF); | |
d342894c | 936 | spin_unlock(&vxlan->hash_lock); |
937 | } | |
26a41ae6 | 938 | |
939 | return false; | |
d342894c | 940 | } |
941 | ||
d342894c | 942 | /* See if multicast group is already in use by other ID */ |
95ab0991 | 943 | static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev) |
d342894c | 944 | { |
553675fb | 945 | struct vxlan_dev *vxlan; |
b1be00a6 | 946 | unsigned short family = dev->default_dst.remote_ip.sa.sa_family; |
d342894c | 947 | |
95ab0991 G |
948 | /* The vxlan_sock is only used by dev, leaving group has |
949 | * no effect on other vxlan devices. | |
950 | */ | |
b1be00a6 JB |
951 | if (family == AF_INET && dev->vn4_sock && |
952 | atomic_read(&dev->vn4_sock->refcnt) == 1) | |
95ab0991 | 953 | return false; |
b1be00a6 JB |
954 | #if IS_ENABLED(CONFIG_IPV6) |
955 | if (family == AF_INET6 && dev->vn6_sock && | |
956 | atomic_read(&dev->vn6_sock->refcnt) == 1) | |
957 | return false; | |
958 | #endif | |
95ab0991 | 959 | |
553675fb | 960 | list_for_each_entry(vxlan, &vn->vxlan_list, next) { |
95ab0991 | 961 | if (!netif_running(vxlan->dev) || vxlan == dev) |
553675fb | 962 | continue; |
d342894c | 963 | |
b1be00a6 | 964 | if (family == AF_INET && vxlan->vn4_sock != dev->vn4_sock) |
95ab0991 | 965 | continue; |
b1be00a6 JB |
966 | #if IS_ENABLED(CONFIG_IPV6) |
967 | if (family == AF_INET6 && vxlan->vn6_sock != dev->vn6_sock) | |
968 | continue; | |
969 | #endif | |
95ab0991 G |
970 | |
971 | if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip, | |
972 | &dev->default_dst.remote_ip)) | |
973 | continue; | |
974 | ||
975 | if (vxlan->default_dst.remote_ifindex != | |
976 | dev->default_dst.remote_ifindex) | |
977 | continue; | |
978 | ||
979 | return true; | |
553675fb | 980 | } |
d342894c | 981 | |
982 | return false; | |
983 | } | |
984 | ||
544a773a | 985 | static bool __vxlan_sock_release_prep(struct vxlan_sock *vs) |
7c47cedf | 986 | { |
b1be00a6 | 987 | struct vxlan_net *vn; |
012a5729 | 988 | |
b1be00a6 | 989 | if (!vs) |
544a773a | 990 | return false; |
7c47cedf | 991 | if (!atomic_dec_and_test(&vs->refcnt)) |
544a773a | 992 | return false; |
d342894c | 993 | |
b1be00a6 | 994 | vn = net_generic(sock_net(vs->sock->sk), vxlan_net_id); |
1c51a915 | 995 | spin_lock(&vn->sock_lock); |
7c47cedf | 996 | hlist_del_rcu(&vs->hlist); |
e7b3db5e | 997 | udp_tunnel_notify_del_rx_port(vs->sock, |
b9adcd69 AD |
998 | (vs->flags & VXLAN_F_GPE) ? |
999 | UDP_TUNNEL_TYPE_VXLAN_GPE : | |
e7b3db5e | 1000 | UDP_TUNNEL_TYPE_VXLAN); |
1c51a915 SH |
1001 | spin_unlock(&vn->sock_lock); |
1002 | ||
544a773a | 1003 | return true; |
d342894c | 1004 | } |
1005 | ||
b1be00a6 JB |
1006 | static void vxlan_sock_release(struct vxlan_dev *vxlan) |
1007 | { | |
544a773a | 1008 | bool ipv4 = __vxlan_sock_release_prep(vxlan->vn4_sock); |
b1be00a6 | 1009 | #if IS_ENABLED(CONFIG_IPV6) |
544a773a HFS |
1010 | bool ipv6 = __vxlan_sock_release_prep(vxlan->vn6_sock); |
1011 | #endif | |
1012 | ||
1013 | synchronize_net(); | |
1014 | ||
1015 | if (ipv4) { | |
1016 | udp_tunnel_sock_release(vxlan->vn4_sock->sock); | |
1017 | kfree(vxlan->vn4_sock); | |
1018 | } | |
1019 | ||
1020 | #if IS_ENABLED(CONFIG_IPV6) | |
1021 | if (ipv6) { | |
1022 | udp_tunnel_sock_release(vxlan->vn6_sock->sock); | |
1023 | kfree(vxlan->vn6_sock); | |
1024 | } | |
b1be00a6 JB |
1025 | #endif |
1026 | } | |
1027 | ||
56ef9c90 | 1028 | /* Update multicast group membership when first VNI on |
c4b49512 | 1029 | * multicast address is brought up |
7c47cedf | 1030 | */ |
56ef9c90 | 1031 | static int vxlan_igmp_join(struct vxlan_dev *vxlan) |
d342894c | 1032 | { |
b1be00a6 | 1033 | struct sock *sk; |
e4c7ed41 CW |
1034 | union vxlan_addr *ip = &vxlan->default_dst.remote_ip; |
1035 | int ifindex = vxlan->default_dst.remote_ifindex; | |
149d7549 | 1036 | int ret = -EINVAL; |
d342894c | 1037 | |
e4c7ed41 CW |
1038 | if (ip->sa.sa_family == AF_INET) { |
1039 | struct ip_mreqn mreq = { | |
1040 | .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr, | |
1041 | .imr_ifindex = ifindex, | |
1042 | }; | |
1043 | ||
b1be00a6 JB |
1044 | sk = vxlan->vn4_sock->sock->sk; |
1045 | lock_sock(sk); | |
56ef9c90 | 1046 | ret = ip_mc_join_group(sk, &mreq); |
b1be00a6 | 1047 | release_sock(sk); |
e4c7ed41 CW |
1048 | #if IS_ENABLED(CONFIG_IPV6) |
1049 | } else { | |
b1be00a6 JB |
1050 | sk = vxlan->vn6_sock->sock->sk; |
1051 | lock_sock(sk); | |
56ef9c90 MRL |
1052 | ret = ipv6_stub->ipv6_sock_mc_join(sk, ifindex, |
1053 | &ip->sin6.sin6_addr); | |
b1be00a6 | 1054 | release_sock(sk); |
e4c7ed41 CW |
1055 | #endif |
1056 | } | |
3fc2de2f | 1057 | |
56ef9c90 | 1058 | return ret; |
3fc2de2f | 1059 | } |
1060 | ||
1061 | /* Inverse of vxlan_igmp_join when last VNI is brought down */ | |
56ef9c90 | 1062 | static int vxlan_igmp_leave(struct vxlan_dev *vxlan) |
3fc2de2f | 1063 | { |
b1be00a6 | 1064 | struct sock *sk; |
e4c7ed41 CW |
1065 | union vxlan_addr *ip = &vxlan->default_dst.remote_ip; |
1066 | int ifindex = vxlan->default_dst.remote_ifindex; | |
149d7549 | 1067 | int ret = -EINVAL; |
3fc2de2f | 1068 | |
e4c7ed41 CW |
1069 | if (ip->sa.sa_family == AF_INET) { |
1070 | struct ip_mreqn mreq = { | |
1071 | .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr, | |
1072 | .imr_ifindex = ifindex, | |
1073 | }; | |
1074 | ||
b1be00a6 JB |
1075 | sk = vxlan->vn4_sock->sock->sk; |
1076 | lock_sock(sk); | |
56ef9c90 | 1077 | ret = ip_mc_leave_group(sk, &mreq); |
b1be00a6 | 1078 | release_sock(sk); |
e4c7ed41 CW |
1079 | #if IS_ENABLED(CONFIG_IPV6) |
1080 | } else { | |
b1be00a6 JB |
1081 | sk = vxlan->vn6_sock->sock->sk; |
1082 | lock_sock(sk); | |
56ef9c90 MRL |
1083 | ret = ipv6_stub->ipv6_sock_mc_drop(sk, ifindex, |
1084 | &ip->sin6.sin6_addr); | |
b1be00a6 | 1085 | release_sock(sk); |
e4c7ed41 CW |
1086 | #endif |
1087 | } | |
d342894c | 1088 | |
56ef9c90 | 1089 | return ret; |
d342894c | 1090 | } |
1091 | ||
f14ecebb JB |
1092 | static bool vxlan_remcsum(struct vxlanhdr *unparsed, |
1093 | struct sk_buff *skb, u32 vxflags) | |
dfd8645e | 1094 | { |
7d34fa75 | 1095 | size_t start, offset; |
dfd8645e | 1096 | |
f14ecebb JB |
1097 | if (!(unparsed->vx_flags & VXLAN_HF_RCO) || skb->remcsum_offload) |
1098 | goto out; | |
b7fe10e5 | 1099 | |
f14ecebb JB |
1100 | start = vxlan_rco_start(unparsed->vx_vni); |
1101 | offset = start + vxlan_rco_offset(unparsed->vx_vni); | |
dfd8645e | 1102 | |
7d34fa75 | 1103 | if (!pskb_may_pull(skb, offset + sizeof(u16))) |
be5cfeab | 1104 | return false; |
dfd8645e | 1105 | |
be5cfeab JB |
1106 | skb_remcsum_process(skb, (void *)(vxlan_hdr(skb) + 1), start, offset, |
1107 | !!(vxflags & VXLAN_F_REMCSUM_NOPARTIAL)); | |
f14ecebb JB |
1108 | out: |
1109 | unparsed->vx_flags &= ~VXLAN_HF_RCO; | |
1110 | unparsed->vx_vni &= VXLAN_VNI_MASK; | |
be5cfeab | 1111 | return true; |
dfd8645e TH |
1112 | } |
1113 | ||
f14ecebb | 1114 | static void vxlan_parse_gbp_hdr(struct vxlanhdr *unparsed, |
64f87d36 | 1115 | struct sk_buff *skb, u32 vxflags, |
10a5af23 | 1116 | struct vxlan_metadata *md) |
3288af08 | 1117 | { |
f14ecebb | 1118 | struct vxlanhdr_gbp *gbp = (struct vxlanhdr_gbp *)unparsed; |
10a5af23 | 1119 | struct metadata_dst *tun_dst; |
f14ecebb JB |
1120 | |
1121 | if (!(unparsed->vx_flags & VXLAN_HF_GBP)) | |
1122 | goto out; | |
3288af08 | 1123 | |
3288af08 JB |
1124 | md->gbp = ntohs(gbp->policy_id); |
1125 | ||
10a5af23 | 1126 | tun_dst = (struct metadata_dst *)skb_dst(skb); |
810813c4 | 1127 | if (tun_dst) { |
3288af08 | 1128 | tun_dst->u.tun_info.key.tun_flags |= TUNNEL_VXLAN_OPT; |
810813c4 DM |
1129 | tun_dst->u.tun_info.options_len = sizeof(*md); |
1130 | } | |
3288af08 JB |
1131 | if (gbp->dont_learn) |
1132 | md->gbp |= VXLAN_GBP_DONT_LEARN; | |
1133 | ||
1134 | if (gbp->policy_applied) | |
1135 | md->gbp |= VXLAN_GBP_POLICY_APPLIED; | |
f14ecebb | 1136 | |
64f87d36 JB |
1137 | /* In flow-based mode, GBP is carried in dst_metadata */ |
1138 | if (!(vxflags & VXLAN_F_COLLECT_METADATA)) | |
1139 | skb->mark = md->gbp; | |
f14ecebb JB |
1140 | out: |
1141 | unparsed->vx_flags &= ~VXLAN_GBP_USED_BITS; | |
3288af08 JB |
1142 | } |
1143 | ||
e1e5314d | 1144 | static bool vxlan_parse_gpe_hdr(struct vxlanhdr *unparsed, |
61618eea | 1145 | __be16 *protocol, |
e1e5314d JB |
1146 | struct sk_buff *skb, u32 vxflags) |
1147 | { | |
1148 | struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)unparsed; | |
1149 | ||
1150 | /* Need to have Next Protocol set for interfaces in GPE mode. */ | |
1151 | if (!gpe->np_applied) | |
1152 | return false; | |
1153 | /* "The initial version is 0. If a receiver does not support the | |
1154 | * version indicated it MUST drop the packet. | |
1155 | */ | |
1156 | if (gpe->version != 0) | |
1157 | return false; | |
1158 | /* "When the O bit is set to 1, the packet is an OAM packet and OAM | |
1159 | * processing MUST occur." However, we don't implement OAM | |
1160 | * processing, thus drop the packet. | |
1161 | */ | |
1162 | if (gpe->oam_flag) | |
1163 | return false; | |
1164 | ||
1165 | switch (gpe->next_protocol) { | |
1166 | case VXLAN_GPE_NP_IPV4: | |
1167 | *protocol = htons(ETH_P_IP); | |
1168 | break; | |
1169 | case VXLAN_GPE_NP_IPV6: | |
1170 | *protocol = htons(ETH_P_IPV6); | |
1171 | break; | |
1172 | case VXLAN_GPE_NP_ETHERNET: | |
1173 | *protocol = htons(ETH_P_TEB); | |
1174 | break; | |
1175 | default: | |
1176 | return false; | |
1177 | } | |
1178 | ||
1179 | unparsed->vx_flags &= ~VXLAN_GPE_USED_BITS; | |
1180 | return true; | |
1181 | } | |
1182 | ||
1ab016e2 JB |
1183 | static bool vxlan_set_mac(struct vxlan_dev *vxlan, |
1184 | struct vxlan_sock *vs, | |
1185 | struct sk_buff *skb) | |
614732ea | 1186 | { |
614732ea | 1187 | union vxlan_addr saddr; |
614732ea | 1188 | |
614732ea | 1189 | skb_reset_mac_header(skb); |
614732ea TG |
1190 | skb->protocol = eth_type_trans(skb, vxlan->dev); |
1191 | skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); | |
1192 | ||
1193 | /* Ignore packet loops (and multicast echo) */ | |
1194 | if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr)) | |
1ab016e2 | 1195 | return false; |
614732ea | 1196 | |
760c6805 | 1197 | /* Get address from the outer IP header */ |
ce212d0f | 1198 | if (vxlan_get_sk_family(vs) == AF_INET) { |
1ab016e2 | 1199 | saddr.sin.sin_addr.s_addr = ip_hdr(skb)->saddr; |
614732ea TG |
1200 | saddr.sa.sa_family = AF_INET; |
1201 | #if IS_ENABLED(CONFIG_IPV6) | |
1202 | } else { | |
1ab016e2 | 1203 | saddr.sin6.sin6_addr = ipv6_hdr(skb)->saddr; |
614732ea TG |
1204 | saddr.sa.sa_family = AF_INET6; |
1205 | #endif | |
1206 | } | |
1207 | ||
1ab016e2 JB |
1208 | if ((vxlan->flags & VXLAN_F_LEARN) && |
1209 | vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source)) | |
1210 | return false; | |
1211 | ||
1212 | return true; | |
1213 | } | |
1214 | ||
760c6805 JB |
1215 | static bool vxlan_ecn_decapsulate(struct vxlan_sock *vs, void *oiph, |
1216 | struct sk_buff *skb) | |
1217 | { | |
1218 | int err = 0; | |
1219 | ||
1220 | if (vxlan_get_sk_family(vs) == AF_INET) | |
1221 | err = IP_ECN_decapsulate(oiph, skb); | |
1222 | #if IS_ENABLED(CONFIG_IPV6) | |
1223 | else | |
1224 | err = IP6_ECN_decapsulate(oiph, skb); | |
1225 | #endif | |
1226 | ||
1227 | if (unlikely(err) && log_ecn_error) { | |
1228 | if (vxlan_get_sk_family(vs) == AF_INET) | |
1229 | net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n", | |
1230 | &((struct iphdr *)oiph)->saddr, | |
1231 | ((struct iphdr *)oiph)->tos); | |
1232 | else | |
1233 | net_info_ratelimited("non-ECT from %pI6\n", | |
1234 | &((struct ipv6hdr *)oiph)->saddr); | |
1235 | } | |
1236 | return err <= 1; | |
1237 | } | |
1238 | ||
d342894c | 1239 | /* Callback from net/ipv4/udp.c to receive packets */ |
f2d1968e | 1240 | static int vxlan_rcv(struct sock *sk, struct sk_buff *skb) |
d342894c | 1241 | { |
f2d1968e | 1242 | struct pcpu_sw_netstats *stats; |
c9e78efb | 1243 | struct vxlan_dev *vxlan; |
5cfccc5a | 1244 | struct vxlan_sock *vs; |
f14ecebb | 1245 | struct vxlanhdr unparsed; |
ee122c79 TG |
1246 | struct vxlan_metadata _md; |
1247 | struct vxlan_metadata *md = &_md; | |
61618eea | 1248 | __be16 protocol = htons(ETH_P_TEB); |
e1e5314d | 1249 | bool raw_proto = false; |
f2d1968e | 1250 | void *oiph; |
d342894c | 1251 | |
e1e5314d | 1252 | /* Need UDP and VXLAN header to be present */ |
7ce04758 | 1253 | if (!pskb_may_pull(skb, VXLAN_HLEN)) |
e5aed006 | 1254 | goto drop; |
d342894c | 1255 | |
f14ecebb | 1256 | unparsed = *vxlan_hdr(skb); |
288b01c8 JB |
1257 | /* VNI flag always required to be set */ |
1258 | if (!(unparsed.vx_flags & VXLAN_HF_VNI)) { | |
1259 | netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n", | |
1260 | ntohl(vxlan_hdr(skb)->vx_flags), | |
1261 | ntohl(vxlan_hdr(skb)->vx_vni)); | |
1262 | /* Return non vxlan pkt */ | |
e5aed006 | 1263 | goto drop; |
d342894c | 1264 | } |
288b01c8 JB |
1265 | unparsed.vx_flags &= ~VXLAN_HF_VNI; |
1266 | unparsed.vx_vni &= ~VXLAN_VNI_MASK; | |
d342894c | 1267 | |
559835ea | 1268 | vs = rcu_dereference_sk_user_data(sk); |
5cfccc5a | 1269 | if (!vs) |
d342894c | 1270 | goto drop; |
d342894c | 1271 | |
c9e78efb JB |
1272 | vxlan = vxlan_vs_find_vni(vs, vxlan_vni(vxlan_hdr(skb)->vx_vni)); |
1273 | if (!vxlan) | |
1274 | goto drop; | |
1275 | ||
e1e5314d JB |
1276 | /* For backwards compatibility, only allow reserved fields to be |
1277 | * used by VXLAN extensions if explicitly requested. | |
1278 | */ | |
1279 | if (vs->flags & VXLAN_F_GPE) { | |
1280 | if (!vxlan_parse_gpe_hdr(&unparsed, &protocol, skb, vs->flags)) | |
1281 | goto drop; | |
1282 | raw_proto = true; | |
1283 | } | |
1284 | ||
1285 | if (__iptunnel_pull_header(skb, VXLAN_HLEN, protocol, raw_proto, | |
1286 | !net_eq(vxlan->net, dev_net(vxlan->dev)))) | |
1287 | goto drop; | |
c9e78efb | 1288 | |
ee122c79 | 1289 | if (vxlan_collect_metadata(vs)) { |
07dabf20 | 1290 | __be32 vni = vxlan_vni(vxlan_hdr(skb)->vx_vni); |
10a5af23 | 1291 | struct metadata_dst *tun_dst; |
07dabf20 | 1292 | |
c29a70d2 | 1293 | tun_dst = udp_tun_rx_dst(skb, vxlan_get_sk_family(vs), TUNNEL_KEY, |
07dabf20 | 1294 | vxlan_vni_to_tun_id(vni), sizeof(*md)); |
c29a70d2 | 1295 | |
ee122c79 TG |
1296 | if (!tun_dst) |
1297 | goto drop; | |
1298 | ||
0f1b7354 | 1299 | md = ip_tunnel_info_opts(&tun_dst->u.tun_info); |
10a5af23 JB |
1300 | |
1301 | skb_dst_set(skb, (struct dst_entry *)tun_dst); | |
ee122c79 TG |
1302 | } else { |
1303 | memset(md, 0, sizeof(*md)); | |
1304 | } | |
1305 | ||
f14ecebb JB |
1306 | if (vs->flags & VXLAN_F_REMCSUM_RX) |
1307 | if (!vxlan_remcsum(&unparsed, skb, vs->flags)) | |
1308 | goto drop; | |
1309 | if (vs->flags & VXLAN_F_GBP) | |
10a5af23 | 1310 | vxlan_parse_gbp_hdr(&unparsed, skb, vs->flags, md); |
e1e5314d JB |
1311 | /* Note that GBP and GPE can never be active together. This is |
1312 | * ensured in vxlan_dev_configure. | |
1313 | */ | |
3511494c | 1314 | |
f14ecebb | 1315 | if (unparsed.vx_flags || unparsed.vx_vni) { |
3bf39475 TH |
1316 | /* If there are any unprocessed flags remaining treat |
1317 | * this as a malformed packet. This behavior diverges from | |
1318 | * VXLAN RFC (RFC7348) which stipulates that bits in reserved | |
1319 | * in reserved fields are to be ignored. The approach here | |
c4b49512 | 1320 | * maintains compatibility with previous stack code, and also |
3bf39475 TH |
1321 | * is more robust and provides a little more security in |
1322 | * adding extensions to VXLAN. | |
1323 | */ | |
288b01c8 | 1324 | goto drop; |
3bf39475 TH |
1325 | } |
1326 | ||
e1e5314d JB |
1327 | if (!raw_proto) { |
1328 | if (!vxlan_set_mac(vxlan, vs, skb)) | |
1329 | goto drop; | |
1330 | } else { | |
8be0cfa4 | 1331 | skb_reset_mac_header(skb); |
e1e5314d JB |
1332 | skb->dev = vxlan->dev; |
1333 | skb->pkt_type = PACKET_HOST; | |
1334 | } | |
f2d1968e | 1335 | |
f2d1968e JB |
1336 | oiph = skb_network_header(skb); |
1337 | skb_reset_network_header(skb); | |
1338 | ||
1339 | if (!vxlan_ecn_decapsulate(vs, oiph, skb)) { | |
1340 | ++vxlan->dev->stats.rx_frame_errors; | |
1341 | ++vxlan->dev->stats.rx_errors; | |
1342 | goto drop; | |
1343 | } | |
1344 | ||
1345 | stats = this_cpu_ptr(vxlan->dev->tstats); | |
1346 | u64_stats_update_begin(&stats->syncp); | |
1347 | stats->rx_packets++; | |
1348 | stats->rx_bytes += skb->len; | |
1349 | u64_stats_update_end(&stats->syncp); | |
1350 | ||
1351 | gro_cells_receive(&vxlan->gro_cells, skb); | |
5cfccc5a PS |
1352 | return 0; |
1353 | ||
1354 | drop: | |
288b01c8 JB |
1355 | /* Consume bad packet */ |
1356 | kfree_skb(skb); | |
1357 | return 0; | |
5cfccc5a PS |
1358 | } |
1359 | ||
e4f67add DS |
1360 | static int arp_reduce(struct net_device *dev, struct sk_buff *skb) |
1361 | { | |
1362 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
1363 | struct arphdr *parp; | |
1364 | u8 *arpptr, *sha; | |
1365 | __be32 sip, tip; | |
1366 | struct neighbour *n; | |
1367 | ||
1368 | if (dev->flags & IFF_NOARP) | |
1369 | goto out; | |
1370 | ||
1371 | if (!pskb_may_pull(skb, arp_hdr_len(dev))) { | |
1372 | dev->stats.tx_dropped++; | |
1373 | goto out; | |
1374 | } | |
1375 | parp = arp_hdr(skb); | |
1376 | ||
1377 | if ((parp->ar_hrd != htons(ARPHRD_ETHER) && | |
1378 | parp->ar_hrd != htons(ARPHRD_IEEE802)) || | |
1379 | parp->ar_pro != htons(ETH_P_IP) || | |
1380 | parp->ar_op != htons(ARPOP_REQUEST) || | |
1381 | parp->ar_hln != dev->addr_len || | |
1382 | parp->ar_pln != 4) | |
1383 | goto out; | |
1384 | arpptr = (u8 *)parp + sizeof(struct arphdr); | |
1385 | sha = arpptr; | |
1386 | arpptr += dev->addr_len; /* sha */ | |
1387 | memcpy(&sip, arpptr, sizeof(sip)); | |
1388 | arpptr += sizeof(sip); | |
1389 | arpptr += dev->addr_len; /* tha */ | |
1390 | memcpy(&tip, arpptr, sizeof(tip)); | |
1391 | ||
1392 | if (ipv4_is_loopback(tip) || | |
1393 | ipv4_is_multicast(tip)) | |
1394 | goto out; | |
1395 | ||
1396 | n = neigh_lookup(&arp_tbl, &tip, dev); | |
1397 | ||
1398 | if (n) { | |
e4f67add DS |
1399 | struct vxlan_fdb *f; |
1400 | struct sk_buff *reply; | |
1401 | ||
1402 | if (!(n->nud_state & NUD_CONNECTED)) { | |
1403 | neigh_release(n); | |
1404 | goto out; | |
1405 | } | |
1406 | ||
1407 | f = vxlan_find_mac(vxlan, n->ha); | |
e4c7ed41 | 1408 | if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) { |
e4f67add DS |
1409 | /* bridge-local neighbor */ |
1410 | neigh_release(n); | |
1411 | goto out; | |
1412 | } | |
1413 | ||
1414 | reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha, | |
1415 | n->ha, sha); | |
1416 | ||
1417 | neigh_release(n); | |
1418 | ||
7346135d DS |
1419 | if (reply == NULL) |
1420 | goto out; | |
1421 | ||
e4f67add DS |
1422 | skb_reset_mac_header(reply); |
1423 | __skb_pull(reply, skb_network_offset(reply)); | |
1424 | reply->ip_summed = CHECKSUM_UNNECESSARY; | |
1425 | reply->pkt_type = PACKET_HOST; | |
1426 | ||
1427 | if (netif_rx_ni(reply) == NET_RX_DROP) | |
1428 | dev->stats.rx_dropped++; | |
e4c7ed41 CW |
1429 | } else if (vxlan->flags & VXLAN_F_L3MISS) { |
1430 | union vxlan_addr ipa = { | |
1431 | .sin.sin_addr.s_addr = tip, | |
a45e92a5 | 1432 | .sin.sin_family = AF_INET, |
e4c7ed41 CW |
1433 | }; |
1434 | ||
1435 | vxlan_ip_miss(dev, &ipa); | |
1436 | } | |
e4f67add DS |
1437 | out: |
1438 | consume_skb(skb); | |
1439 | return NETDEV_TX_OK; | |
1440 | } | |
1441 | ||
f564f45c | 1442 | #if IS_ENABLED(CONFIG_IPV6) |
4b29dba9 DS |
1443 | static struct sk_buff *vxlan_na_create(struct sk_buff *request, |
1444 | struct neighbour *n, bool isrouter) | |
1445 | { | |
1446 | struct net_device *dev = request->dev; | |
1447 | struct sk_buff *reply; | |
1448 | struct nd_msg *ns, *na; | |
1449 | struct ipv6hdr *pip6; | |
1450 | u8 *daddr; | |
1451 | int na_olen = 8; /* opt hdr + ETH_ALEN for target */ | |
1452 | int ns_olen; | |
1453 | int i, len; | |
1454 | ||
1455 | if (dev == NULL) | |
1456 | return NULL; | |
1457 | ||
1458 | len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) + | |
1459 | sizeof(*na) + na_olen + dev->needed_tailroom; | |
1460 | reply = alloc_skb(len, GFP_ATOMIC); | |
1461 | if (reply == NULL) | |
1462 | return NULL; | |
1463 | ||
1464 | reply->protocol = htons(ETH_P_IPV6); | |
1465 | reply->dev = dev; | |
1466 | skb_reserve(reply, LL_RESERVED_SPACE(request->dev)); | |
1467 | skb_push(reply, sizeof(struct ethhdr)); | |
6297b91c | 1468 | skb_reset_mac_header(reply); |
4b29dba9 DS |
1469 | |
1470 | ns = (struct nd_msg *)skb_transport_header(request); | |
1471 | ||
1472 | daddr = eth_hdr(request)->h_source; | |
1473 | ns_olen = request->len - skb_transport_offset(request) - sizeof(*ns); | |
1474 | for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) { | |
1475 | if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) { | |
1476 | daddr = ns->opt + i + sizeof(struct nd_opt_hdr); | |
1477 | break; | |
1478 | } | |
1479 | } | |
1480 | ||
1481 | /* Ethernet header */ | |
1482 | ether_addr_copy(eth_hdr(reply)->h_dest, daddr); | |
1483 | ether_addr_copy(eth_hdr(reply)->h_source, n->ha); | |
1484 | eth_hdr(reply)->h_proto = htons(ETH_P_IPV6); | |
1485 | reply->protocol = htons(ETH_P_IPV6); | |
1486 | ||
1487 | skb_pull(reply, sizeof(struct ethhdr)); | |
6297b91c | 1488 | skb_reset_network_header(reply); |
4b29dba9 DS |
1489 | skb_put(reply, sizeof(struct ipv6hdr)); |
1490 | ||
1491 | /* IPv6 header */ | |
1492 | ||
1493 | pip6 = ipv6_hdr(reply); | |
1494 | memset(pip6, 0, sizeof(struct ipv6hdr)); | |
1495 | pip6->version = 6; | |
1496 | pip6->priority = ipv6_hdr(request)->priority; | |
1497 | pip6->nexthdr = IPPROTO_ICMPV6; | |
1498 | pip6->hop_limit = 255; | |
1499 | pip6->daddr = ipv6_hdr(request)->saddr; | |
1500 | pip6->saddr = *(struct in6_addr *)n->primary_key; | |
1501 | ||
1502 | skb_pull(reply, sizeof(struct ipv6hdr)); | |
6297b91c | 1503 | skb_reset_transport_header(reply); |
4b29dba9 DS |
1504 | |
1505 | na = (struct nd_msg *)skb_put(reply, sizeof(*na) + na_olen); | |
1506 | ||
1507 | /* Neighbor Advertisement */ | |
1508 | memset(na, 0, sizeof(*na)+na_olen); | |
1509 | na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT; | |
1510 | na->icmph.icmp6_router = isrouter; | |
1511 | na->icmph.icmp6_override = 1; | |
1512 | na->icmph.icmp6_solicited = 1; | |
1513 | na->target = ns->target; | |
1514 | ether_addr_copy(&na->opt[2], n->ha); | |
1515 | na->opt[0] = ND_OPT_TARGET_LL_ADDR; | |
1516 | na->opt[1] = na_olen >> 3; | |
1517 | ||
1518 | na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr, | |
1519 | &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6, | |
1520 | csum_partial(na, sizeof(*na)+na_olen, 0)); | |
1521 | ||
1522 | pip6->payload_len = htons(sizeof(*na)+na_olen); | |
1523 | ||
1524 | skb_push(reply, sizeof(struct ipv6hdr)); | |
1525 | ||
1526 | reply->ip_summed = CHECKSUM_UNNECESSARY; | |
1527 | ||
1528 | return reply; | |
1529 | } | |
1530 | ||
f564f45c CW |
1531 | static int neigh_reduce(struct net_device *dev, struct sk_buff *skb) |
1532 | { | |
1533 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
4b29dba9 | 1534 | struct nd_msg *msg; |
f564f45c CW |
1535 | const struct ipv6hdr *iphdr; |
1536 | const struct in6_addr *saddr, *daddr; | |
4b29dba9 DS |
1537 | struct neighbour *n; |
1538 | struct inet6_dev *in6_dev; | |
f564f45c CW |
1539 | |
1540 | in6_dev = __in6_dev_get(dev); | |
1541 | if (!in6_dev) | |
1542 | goto out; | |
1543 | ||
f564f45c CW |
1544 | iphdr = ipv6_hdr(skb); |
1545 | saddr = &iphdr->saddr; | |
1546 | daddr = &iphdr->daddr; | |
1547 | ||
f564f45c CW |
1548 | msg = (struct nd_msg *)skb_transport_header(skb); |
1549 | if (msg->icmph.icmp6_code != 0 || | |
1550 | msg->icmph.icmp6_type != NDISC_NEIGHBOUR_SOLICITATION) | |
1551 | goto out; | |
1552 | ||
4b29dba9 DS |
1553 | if (ipv6_addr_loopback(daddr) || |
1554 | ipv6_addr_is_multicast(&msg->target)) | |
1555 | goto out; | |
1556 | ||
1557 | n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev); | |
f564f45c CW |
1558 | |
1559 | if (n) { | |
1560 | struct vxlan_fdb *f; | |
4b29dba9 | 1561 | struct sk_buff *reply; |
f564f45c CW |
1562 | |
1563 | if (!(n->nud_state & NUD_CONNECTED)) { | |
1564 | neigh_release(n); | |
1565 | goto out; | |
1566 | } | |
1567 | ||
1568 | f = vxlan_find_mac(vxlan, n->ha); | |
1569 | if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) { | |
1570 | /* bridge-local neighbor */ | |
1571 | neigh_release(n); | |
1572 | goto out; | |
1573 | } | |
1574 | ||
4b29dba9 DS |
1575 | reply = vxlan_na_create(skb, n, |
1576 | !!(f ? f->flags & NTF_ROUTER : 0)); | |
1577 | ||
f564f45c | 1578 | neigh_release(n); |
4b29dba9 DS |
1579 | |
1580 | if (reply == NULL) | |
1581 | goto out; | |
1582 | ||
1583 | if (netif_rx_ni(reply) == NET_RX_DROP) | |
1584 | dev->stats.rx_dropped++; | |
1585 | ||
f564f45c | 1586 | } else if (vxlan->flags & VXLAN_F_L3MISS) { |
4b29dba9 DS |
1587 | union vxlan_addr ipa = { |
1588 | .sin6.sin6_addr = msg->target, | |
a45e92a5 | 1589 | .sin6.sin6_family = AF_INET6, |
4b29dba9 DS |
1590 | }; |
1591 | ||
f564f45c CW |
1592 | vxlan_ip_miss(dev, &ipa); |
1593 | } | |
1594 | ||
1595 | out: | |
1596 | consume_skb(skb); | |
1597 | return NETDEV_TX_OK; | |
1598 | } | |
1599 | #endif | |
1600 | ||
e4f67add DS |
1601 | static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb) |
1602 | { | |
1603 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
1604 | struct neighbour *n; | |
e4f67add DS |
1605 | |
1606 | if (is_multicast_ether_addr(eth_hdr(skb)->h_dest)) | |
1607 | return false; | |
1608 | ||
1609 | n = NULL; | |
1610 | switch (ntohs(eth_hdr(skb)->h_proto)) { | |
1611 | case ETH_P_IP: | |
e15a00aa CW |
1612 | { |
1613 | struct iphdr *pip; | |
1614 | ||
e4f67add DS |
1615 | if (!pskb_may_pull(skb, sizeof(struct iphdr))) |
1616 | return false; | |
1617 | pip = ip_hdr(skb); | |
1618 | n = neigh_lookup(&arp_tbl, &pip->daddr, dev); | |
e4c7ed41 CW |
1619 | if (!n && (vxlan->flags & VXLAN_F_L3MISS)) { |
1620 | union vxlan_addr ipa = { | |
1621 | .sin.sin_addr.s_addr = pip->daddr, | |
a45e92a5 | 1622 | .sin.sin_family = AF_INET, |
e4c7ed41 CW |
1623 | }; |
1624 | ||
1625 | vxlan_ip_miss(dev, &ipa); | |
1626 | return false; | |
1627 | } | |
1628 | ||
e4f67add | 1629 | break; |
e15a00aa CW |
1630 | } |
1631 | #if IS_ENABLED(CONFIG_IPV6) | |
1632 | case ETH_P_IPV6: | |
1633 | { | |
1634 | struct ipv6hdr *pip6; | |
1635 | ||
1636 | if (!pskb_may_pull(skb, sizeof(struct ipv6hdr))) | |
1637 | return false; | |
1638 | pip6 = ipv6_hdr(skb); | |
1639 | n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev); | |
1640 | if (!n && (vxlan->flags & VXLAN_F_L3MISS)) { | |
1641 | union vxlan_addr ipa = { | |
1642 | .sin6.sin6_addr = pip6->daddr, | |
a45e92a5 | 1643 | .sin6.sin6_family = AF_INET6, |
e15a00aa CW |
1644 | }; |
1645 | ||
1646 | vxlan_ip_miss(dev, &ipa); | |
1647 | return false; | |
1648 | } | |
1649 | ||
1650 | break; | |
1651 | } | |
1652 | #endif | |
e4f67add DS |
1653 | default: |
1654 | return false; | |
1655 | } | |
1656 | ||
1657 | if (n) { | |
1658 | bool diff; | |
1659 | ||
7367d0b5 | 1660 | diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha); |
e4f67add DS |
1661 | if (diff) { |
1662 | memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest, | |
1663 | dev->addr_len); | |
1664 | memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len); | |
1665 | } | |
1666 | neigh_release(n); | |
1667 | return diff; | |
e4c7ed41 CW |
1668 | } |
1669 | ||
e4f67add DS |
1670 | return false; |
1671 | } | |
1672 | ||
af33c1ad | 1673 | static void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, u32 vxflags, |
3511494c TG |
1674 | struct vxlan_metadata *md) |
1675 | { | |
1676 | struct vxlanhdr_gbp *gbp; | |
1677 | ||
db79a621 TG |
1678 | if (!md->gbp) |
1679 | return; | |
1680 | ||
3511494c | 1681 | gbp = (struct vxlanhdr_gbp *)vxh; |
54bfd872 | 1682 | vxh->vx_flags |= VXLAN_HF_GBP; |
3511494c TG |
1683 | |
1684 | if (md->gbp & VXLAN_GBP_DONT_LEARN) | |
1685 | gbp->dont_learn = 1; | |
1686 | ||
1687 | if (md->gbp & VXLAN_GBP_POLICY_APPLIED) | |
1688 | gbp->policy_applied = 1; | |
1689 | ||
1690 | gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK); | |
1691 | } | |
1692 | ||
e1e5314d JB |
1693 | static int vxlan_build_gpe_hdr(struct vxlanhdr *vxh, u32 vxflags, |
1694 | __be16 protocol) | |
1695 | { | |
1696 | struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)vxh; | |
1697 | ||
1698 | gpe->np_applied = 1; | |
1699 | ||
1700 | switch (protocol) { | |
1701 | case htons(ETH_P_IP): | |
1702 | gpe->next_protocol = VXLAN_GPE_NP_IPV4; | |
1703 | return 0; | |
1704 | case htons(ETH_P_IPV6): | |
1705 | gpe->next_protocol = VXLAN_GPE_NP_IPV6; | |
1706 | return 0; | |
1707 | case htons(ETH_P_TEB): | |
1708 | gpe->next_protocol = VXLAN_GPE_NP_ETHERNET; | |
1709 | return 0; | |
1710 | } | |
1711 | return -EPFNOSUPPORT; | |
1712 | } | |
1713 | ||
f491e56d JB |
1714 | static int vxlan_build_skb(struct sk_buff *skb, struct dst_entry *dst, |
1715 | int iphdr_len, __be32 vni, | |
1716 | struct vxlan_metadata *md, u32 vxflags, | |
b4ed5cad | 1717 | bool udp_sum) |
e4c7ed41 | 1718 | { |
e4c7ed41 | 1719 | struct vxlanhdr *vxh; |
e4c7ed41 CW |
1720 | int min_headroom; |
1721 | int err; | |
dfd8645e | 1722 | int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL; |
e1e5314d | 1723 | __be16 inner_protocol = htons(ETH_P_TEB); |
dfd8645e | 1724 | |
af33c1ad | 1725 | if ((vxflags & VXLAN_F_REMCSUM_TX) && |
dfd8645e TH |
1726 | skb->ip_summed == CHECKSUM_PARTIAL) { |
1727 | int csum_start = skb_checksum_start_offset(skb); | |
1728 | ||
1729 | if (csum_start <= VXLAN_MAX_REMCSUM_START && | |
1730 | !(csum_start & VXLAN_RCO_SHIFT_MASK) && | |
1731 | (skb->csum_offset == offsetof(struct udphdr, check) || | |
b5708501 | 1732 | skb->csum_offset == offsetof(struct tcphdr, check))) |
dfd8645e | 1733 | type |= SKB_GSO_TUNNEL_REMCSUM; |
dfd8645e | 1734 | } |
e4c7ed41 | 1735 | |
e4c7ed41 | 1736 | min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len |
f491e56d | 1737 | + VXLAN_HLEN + iphdr_len |
df8a39de | 1738 | + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0); |
649c5b8b PS |
1739 | |
1740 | /* Need space for new headers (invalidates iph ptr) */ | |
1741 | err = skb_cow_head(skb, min_headroom); | |
e1e5314d JB |
1742 | if (unlikely(err)) |
1743 | goto out_free; | |
649c5b8b | 1744 | |
5968250c JP |
1745 | skb = vlan_hwaccel_push_inside(skb); |
1746 | if (WARN_ON(!skb)) | |
1747 | return -ENOMEM; | |
1eaa8178 | 1748 | |
aed069df AD |
1749 | err = iptunnel_handle_offloads(skb, type); |
1750 | if (err) | |
1751 | goto out_free; | |
b736a623 | 1752 | |
49560532 | 1753 | vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh)); |
54bfd872 JB |
1754 | vxh->vx_flags = VXLAN_HF_VNI; |
1755 | vxh->vx_vni = vxlan_vni_field(vni); | |
49560532 | 1756 | |
dfd8645e | 1757 | if (type & SKB_GSO_TUNNEL_REMCSUM) { |
54bfd872 | 1758 | unsigned int start; |
dfd8645e | 1759 | |
54bfd872 JB |
1760 | start = skb_checksum_start_offset(skb) - sizeof(struct vxlanhdr); |
1761 | vxh->vx_vni |= vxlan_compute_rco(start, skb->csum_offset); | |
1762 | vxh->vx_flags |= VXLAN_HF_RCO; | |
dfd8645e TH |
1763 | |
1764 | if (!skb_is_gso(skb)) { | |
1765 | skb->ip_summed = CHECKSUM_NONE; | |
1766 | skb->encapsulation = 0; | |
1767 | } | |
1768 | } | |
1769 | ||
af33c1ad TH |
1770 | if (vxflags & VXLAN_F_GBP) |
1771 | vxlan_build_gbp_hdr(vxh, vxflags, md); | |
e1e5314d JB |
1772 | if (vxflags & VXLAN_F_GPE) { |
1773 | err = vxlan_build_gpe_hdr(vxh, vxflags, skb->protocol); | |
1774 | if (err < 0) | |
1775 | goto out_free; | |
1776 | inner_protocol = skb->protocol; | |
1777 | } | |
3511494c | 1778 | |
e1e5314d | 1779 | skb_set_inner_protocol(skb, inner_protocol); |
039f5062 | 1780 | return 0; |
e1e5314d JB |
1781 | |
1782 | out_free: | |
1783 | kfree_skb(skb); | |
1784 | return err; | |
49560532 | 1785 | } |
49560532 | 1786 | |
1a8496ba JB |
1787 | static struct rtable *vxlan_get_route(struct vxlan_dev *vxlan, |
1788 | struct sk_buff *skb, int oif, u8 tos, | |
0c1d70af PA |
1789 | __be32 daddr, __be32 *saddr, |
1790 | struct dst_cache *dst_cache, | |
db3c6139 | 1791 | const struct ip_tunnel_info *info) |
1a8496ba | 1792 | { |
db3c6139 | 1793 | bool use_cache = ip_tunnel_dst_cache_usable(skb, info); |
1a8496ba JB |
1794 | struct rtable *rt = NULL; |
1795 | struct flowi4 fl4; | |
1796 | ||
db3c6139 DB |
1797 | if (tos && !info) |
1798 | use_cache = false; | |
1799 | if (use_cache) { | |
0c1d70af PA |
1800 | rt = dst_cache_get_ip4(dst_cache, saddr); |
1801 | if (rt) | |
1802 | return rt; | |
1803 | } | |
1804 | ||
1a8496ba JB |
1805 | memset(&fl4, 0, sizeof(fl4)); |
1806 | fl4.flowi4_oif = oif; | |
1807 | fl4.flowi4_tos = RT_TOS(tos); | |
1808 | fl4.flowi4_mark = skb->mark; | |
1809 | fl4.flowi4_proto = IPPROTO_UDP; | |
1810 | fl4.daddr = daddr; | |
272d96a5 | 1811 | fl4.saddr = *saddr; |
1a8496ba JB |
1812 | |
1813 | rt = ip_route_output_key(vxlan->net, &fl4); | |
0c1d70af | 1814 | if (!IS_ERR(rt)) { |
1a8496ba | 1815 | *saddr = fl4.saddr; |
0c1d70af PA |
1816 | if (use_cache) |
1817 | dst_cache_set_ip4(dst_cache, &rt->dst, fl4.saddr); | |
1818 | } | |
1a8496ba JB |
1819 | return rt; |
1820 | } | |
1821 | ||
e5d4b29f JB |
1822 | #if IS_ENABLED(CONFIG_IPV6) |
1823 | static struct dst_entry *vxlan6_get_route(struct vxlan_dev *vxlan, | |
1400615d | 1824 | struct sk_buff *skb, int oif, u8 tos, |
e7f70af1 | 1825 | __be32 label, |
e5d4b29f | 1826 | const struct in6_addr *daddr, |
0c1d70af | 1827 | struct in6_addr *saddr, |
db3c6139 DB |
1828 | struct dst_cache *dst_cache, |
1829 | const struct ip_tunnel_info *info) | |
e5d4b29f | 1830 | { |
db3c6139 | 1831 | bool use_cache = ip_tunnel_dst_cache_usable(skb, info); |
e5d4b29f JB |
1832 | struct dst_entry *ndst; |
1833 | struct flowi6 fl6; | |
1834 | int err; | |
1835 | ||
1400615d DB |
1836 | if (tos && !info) |
1837 | use_cache = false; | |
db3c6139 | 1838 | if (use_cache) { |
0c1d70af PA |
1839 | ndst = dst_cache_get_ip6(dst_cache, saddr); |
1840 | if (ndst) | |
1841 | return ndst; | |
1842 | } | |
1843 | ||
e5d4b29f JB |
1844 | memset(&fl6, 0, sizeof(fl6)); |
1845 | fl6.flowi6_oif = oif; | |
1846 | fl6.daddr = *daddr; | |
272d96a5 | 1847 | fl6.saddr = *saddr; |
eaa93bf4 | 1848 | fl6.flowlabel = ip6_make_flowinfo(RT_TOS(tos), label); |
e5d4b29f JB |
1849 | fl6.flowi6_mark = skb->mark; |
1850 | fl6.flowi6_proto = IPPROTO_UDP; | |
1851 | ||
1852 | err = ipv6_stub->ipv6_dst_lookup(vxlan->net, | |
1853 | vxlan->vn6_sock->sock->sk, | |
1854 | &ndst, &fl6); | |
1855 | if (err < 0) | |
1856 | return ERR_PTR(err); | |
1857 | ||
1858 | *saddr = fl6.saddr; | |
db3c6139 | 1859 | if (use_cache) |
0c1d70af | 1860 | dst_cache_set_ip6(dst_cache, ndst, saddr); |
e5d4b29f JB |
1861 | return ndst; |
1862 | } | |
1863 | #endif | |
1864 | ||
9dcc71e1 SS |
1865 | /* Bypass encapsulation if the destination is local */ |
1866 | static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan, | |
1867 | struct vxlan_dev *dst_vxlan) | |
1868 | { | |
8f84985f | 1869 | struct pcpu_sw_netstats *tx_stats, *rx_stats; |
e4c7ed41 CW |
1870 | union vxlan_addr loopback; |
1871 | union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip; | |
ce6502a8 LR |
1872 | struct net_device *dev = skb->dev; |
1873 | int len = skb->len; | |
9dcc71e1 | 1874 | |
8f84985f LR |
1875 | tx_stats = this_cpu_ptr(src_vxlan->dev->tstats); |
1876 | rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats); | |
9dcc71e1 SS |
1877 | skb->pkt_type = PACKET_HOST; |
1878 | skb->encapsulation = 0; | |
1879 | skb->dev = dst_vxlan->dev; | |
1880 | __skb_pull(skb, skb_network_offset(skb)); | |
1881 | ||
e4c7ed41 CW |
1882 | if (remote_ip->sa.sa_family == AF_INET) { |
1883 | loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK); | |
1884 | loopback.sa.sa_family = AF_INET; | |
1885 | #if IS_ENABLED(CONFIG_IPV6) | |
1886 | } else { | |
1887 | loopback.sin6.sin6_addr = in6addr_loopback; | |
1888 | loopback.sa.sa_family = AF_INET6; | |
1889 | #endif | |
1890 | } | |
1891 | ||
9dcc71e1 | 1892 | if (dst_vxlan->flags & VXLAN_F_LEARN) |
e4c7ed41 | 1893 | vxlan_snoop(skb->dev, &loopback, eth_hdr(skb)->h_source); |
9dcc71e1 SS |
1894 | |
1895 | u64_stats_update_begin(&tx_stats->syncp); | |
1896 | tx_stats->tx_packets++; | |
ce6502a8 | 1897 | tx_stats->tx_bytes += len; |
9dcc71e1 SS |
1898 | u64_stats_update_end(&tx_stats->syncp); |
1899 | ||
1900 | if (netif_rx(skb) == NET_RX_SUCCESS) { | |
1901 | u64_stats_update_begin(&rx_stats->syncp); | |
1902 | rx_stats->rx_packets++; | |
ce6502a8 | 1903 | rx_stats->rx_bytes += len; |
9dcc71e1 SS |
1904 | u64_stats_update_end(&rx_stats->syncp); |
1905 | } else { | |
ce6502a8 | 1906 | dev->stats.rx_dropped++; |
9dcc71e1 SS |
1907 | } |
1908 | } | |
1909 | ||
4ad16930 SH |
1910 | static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev, |
1911 | struct vxlan_rdst *rdst, bool did_rsc) | |
d342894c | 1912 | { |
d71785ff | 1913 | struct dst_cache *dst_cache; |
3093fbe7 | 1914 | struct ip_tunnel_info *info; |
d342894c | 1915 | struct vxlan_dev *vxlan = netdev_priv(dev); |
b1be00a6 | 1916 | struct sock *sk; |
e4c7ed41 | 1917 | struct rtable *rt = NULL; |
d342894c | 1918 | const struct iphdr *old_iph; |
e4c7ed41 | 1919 | union vxlan_addr *dst; |
272d96a5 | 1920 | union vxlan_addr remote_ip, local_ip; |
1921 | union vxlan_addr *src; | |
ee122c79 TG |
1922 | struct vxlan_metadata _md; |
1923 | struct vxlan_metadata *md = &_md; | |
e4c7ed41 | 1924 | __be16 src_port = 0, dst_port; |
e7f70af1 | 1925 | __be32 vni, label; |
d342894c | 1926 | __be16 df = 0; |
1927 | __u8 tos, ttl; | |
0e6fbc5b | 1928 | int err; |
ee122c79 | 1929 | u32 flags = vxlan->flags; |
b4ed5cad | 1930 | bool udp_sum = false; |
f491e56d | 1931 | bool xnet = !net_eq(vxlan->net, dev_net(vxlan->dev)); |
d342894c | 1932 | |
61adedf3 | 1933 | info = skb_tunnel_info(skb); |
3093fbe7 | 1934 | |
ee122c79 | 1935 | if (rdst) { |
0dfbdf41 | 1936 | dst_port = rdst->remote_port ? rdst->remote_port : vxlan->cfg.dst_port; |
ee122c79 TG |
1937 | vni = rdst->remote_vni; |
1938 | dst = &rdst->remote_ip; | |
272d96a5 | 1939 | src = &vxlan->cfg.saddr; |
d71785ff | 1940 | dst_cache = &rdst->dst_cache; |
ee122c79 TG |
1941 | } else { |
1942 | if (!info) { | |
1943 | WARN_ONCE(1, "%s: Missing encapsulation instructions\n", | |
1944 | dev->name); | |
1945 | goto drop; | |
1946 | } | |
0dfbdf41 | 1947 | dst_port = info->key.tp_dst ? : vxlan->cfg.dst_port; |
54bfd872 | 1948 | vni = vxlan_tun_id_to_vni(info->key.tun_id); |
b1be00a6 | 1949 | remote_ip.sa.sa_family = ip_tunnel_info_af(info); |
272d96a5 | 1950 | if (remote_ip.sa.sa_family == AF_INET) { |
a725e514 | 1951 | remote_ip.sin.sin_addr.s_addr = info->key.u.ipv4.dst; |
272d96a5 | 1952 | local_ip.sin.sin_addr.s_addr = info->key.u.ipv4.src; |
1953 | } else { | |
a725e514 | 1954 | remote_ip.sin6.sin6_addr = info->key.u.ipv6.dst; |
272d96a5 | 1955 | local_ip.sin6.sin6_addr = info->key.u.ipv6.src; |
1956 | } | |
ee122c79 | 1957 | dst = &remote_ip; |
272d96a5 | 1958 | src = &local_ip; |
d71785ff | 1959 | dst_cache = &info->dst_cache; |
ee122c79 | 1960 | } |
e4f67add | 1961 | |
e4c7ed41 | 1962 | if (vxlan_addr_any(dst)) { |
e4f67add | 1963 | if (did_rsc) { |
e4f67add | 1964 | /* short-circuited back to local bridge */ |
9dcc71e1 | 1965 | vxlan_encap_bypass(skb, vxlan, vxlan); |
4ad16930 | 1966 | return; |
e4f67add | 1967 | } |
ef59febe | 1968 | goto drop; |
e4f67add | 1969 | } |
ef59febe | 1970 | |
d342894c | 1971 | old_iph = ip_hdr(skb); |
1972 | ||
0dfbdf41 | 1973 | ttl = vxlan->cfg.ttl; |
e4c7ed41 | 1974 | if (!ttl && vxlan_addr_multicast(dst)) |
d342894c | 1975 | ttl = 1; |
1976 | ||
0dfbdf41 | 1977 | tos = vxlan->cfg.tos; |
d342894c | 1978 | if (tos == 1) |
206aaafc | 1979 | tos = ip_tunnel_get_dsfield(old_iph, skb); |
d342894c | 1980 | |
e7f70af1 | 1981 | label = vxlan->cfg.label; |
0dfbdf41 TG |
1982 | src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min, |
1983 | vxlan->cfg.port_max, true); | |
d342894c | 1984 | |
a725e514 | 1985 | if (info) { |
a725e514 JB |
1986 | ttl = info->key.ttl; |
1987 | tos = info->key.tos; | |
e7f70af1 | 1988 | label = info->key.label; |
b4ed5cad | 1989 | udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM); |
a725e514 JB |
1990 | |
1991 | if (info->options_len) | |
4c222798 | 1992 | md = ip_tunnel_info_opts(info); |
a725e514 JB |
1993 | } else { |
1994 | md->gbp = skb->mark; | |
1995 | } | |
1996 | ||
e4c7ed41 | 1997 | if (dst->sa.sa_family == AF_INET) { |
b1be00a6 JB |
1998 | if (!vxlan->vn4_sock) |
1999 | goto drop; | |
2000 | sk = vxlan->vn4_sock->sock->sk; | |
2001 | ||
1a8496ba JB |
2002 | rt = vxlan_get_route(vxlan, skb, |
2003 | rdst ? rdst->remote_ifindex : 0, tos, | |
272d96a5 | 2004 | dst->sin.sin_addr.s_addr, |
2005 | &src->sin.sin_addr.s_addr, | |
d71785ff | 2006 | dst_cache, info); |
e4c7ed41 CW |
2007 | if (IS_ERR(rt)) { |
2008 | netdev_dbg(dev, "no route to %pI4\n", | |
2009 | &dst->sin.sin_addr.s_addr); | |
2010 | dev->stats.tx_carrier_errors++; | |
2011 | goto tx_error; | |
2012 | } | |
d342894c | 2013 | |
e4c7ed41 CW |
2014 | if (rt->dst.dev == dev) { |
2015 | netdev_dbg(dev, "circular route to %pI4\n", | |
2016 | &dst->sin.sin_addr.s_addr); | |
2017 | dev->stats.collisions++; | |
fffc15a5 | 2018 | goto rt_tx_error; |
e4c7ed41 CW |
2019 | } |
2020 | ||
2021 | /* Bypass encapsulation if the destination is local */ | |
bbec7802 | 2022 | if (!info && rt->rt_flags & RTCF_LOCAL && |
e4c7ed41 CW |
2023 | !(rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) { |
2024 | struct vxlan_dev *dst_vxlan; | |
2025 | ||
2026 | ip_rt_put(rt); | |
19ca9fc1 | 2027 | dst_vxlan = vxlan_find_vni(vxlan->net, vni, |
ac5132d1 TG |
2028 | dst->sa.sa_family, dst_port, |
2029 | vxlan->flags); | |
e4c7ed41 CW |
2030 | if (!dst_vxlan) |
2031 | goto tx_error; | |
2032 | vxlan_encap_bypass(skb, vxlan, dst_vxlan); | |
2033 | return; | |
2034 | } | |
2035 | ||
6ceb31ca AD |
2036 | if (!info) |
2037 | udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM_TX); | |
2038 | else if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT) | |
2039 | df = htons(IP_DF); | |
2040 | ||
e4c7ed41 CW |
2041 | tos = ip_tunnel_ecn_encap(tos, old_iph, skb); |
2042 | ttl = ttl ? : ip4_dst_hoplimit(&rt->dst); | |
f491e56d | 2043 | err = vxlan_build_skb(skb, &rt->dst, sizeof(struct iphdr), |
54bfd872 | 2044 | vni, md, flags, udp_sum); |
f491e56d JB |
2045 | if (err < 0) |
2046 | goto xmit_tx_error; | |
2047 | ||
272d96a5 | 2048 | udp_tunnel_xmit_skb(rt, sk, skb, src->sin.sin_addr.s_addr, |
f491e56d JB |
2049 | dst->sin.sin_addr.s_addr, tos, ttl, df, |
2050 | src_port, dst_port, xnet, !udp_sum); | |
e4c7ed41 CW |
2051 | #if IS_ENABLED(CONFIG_IPV6) |
2052 | } else { | |
e4c7ed41 | 2053 | struct dst_entry *ndst; |
6f264e47 | 2054 | u32 rt6i_flags; |
e4c7ed41 | 2055 | |
b1be00a6 JB |
2056 | if (!vxlan->vn6_sock) |
2057 | goto drop; | |
2058 | sk = vxlan->vn6_sock->sock->sk; | |
2059 | ||
e5d4b29f | 2060 | ndst = vxlan6_get_route(vxlan, skb, |
1400615d | 2061 | rdst ? rdst->remote_ifindex : 0, tos, |
272d96a5 | 2062 | label, &dst->sin6.sin6_addr, |
2063 | &src->sin6.sin6_addr, | |
db3c6139 | 2064 | dst_cache, info); |
e5d4b29f | 2065 | if (IS_ERR(ndst)) { |
e4c7ed41 CW |
2066 | netdev_dbg(dev, "no route to %pI6\n", |
2067 | &dst->sin6.sin6_addr); | |
2068 | dev->stats.tx_carrier_errors++; | |
9dcc71e1 | 2069 | goto tx_error; |
e4c7ed41 | 2070 | } |
d342894c | 2071 | |
e4c7ed41 CW |
2072 | if (ndst->dev == dev) { |
2073 | netdev_dbg(dev, "circular route to %pI6\n", | |
2074 | &dst->sin6.sin6_addr); | |
2075 | dst_release(ndst); | |
2076 | dev->stats.collisions++; | |
2077 | goto tx_error; | |
2078 | } | |
0e6fbc5b | 2079 | |
e4c7ed41 | 2080 | /* Bypass encapsulation if the destination is local */ |
6f264e47 | 2081 | rt6i_flags = ((struct rt6_info *)ndst)->rt6i_flags; |
bbec7802 | 2082 | if (!info && rt6i_flags & RTF_LOCAL && |
6f264e47 | 2083 | !(rt6i_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) { |
e4c7ed41 CW |
2084 | struct vxlan_dev *dst_vxlan; |
2085 | ||
2086 | dst_release(ndst); | |
19ca9fc1 | 2087 | dst_vxlan = vxlan_find_vni(vxlan->net, vni, |
ac5132d1 TG |
2088 | dst->sa.sa_family, dst_port, |
2089 | vxlan->flags); | |
e4c7ed41 CW |
2090 | if (!dst_vxlan) |
2091 | goto tx_error; | |
2092 | vxlan_encap_bypass(skb, vxlan, dst_vxlan); | |
2093 | return; | |
2094 | } | |
49560532 | 2095 | |
b4ed5cad JB |
2096 | if (!info) |
2097 | udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM6_TX); | |
35e2d115 | 2098 | |
1400615d | 2099 | tos = ip_tunnel_ecn_encap(tos, old_iph, skb); |
e4c7ed41 | 2100 | ttl = ttl ? : ip6_dst_hoplimit(ndst); |
f491e56d JB |
2101 | skb_scrub_packet(skb, xnet); |
2102 | err = vxlan_build_skb(skb, ndst, sizeof(struct ipv6hdr), | |
54bfd872 | 2103 | vni, md, flags, udp_sum); |
f491e56d JB |
2104 | if (err < 0) { |
2105 | dst_release(ndst); | |
5e1e61a3 | 2106 | dev->stats.tx_errors++; |
f491e56d JB |
2107 | return; |
2108 | } | |
2109 | udp_tunnel6_xmit_skb(ndst, sk, skb, dev, | |
272d96a5 | 2110 | &src->sin6.sin6_addr, |
2111 | &dst->sin6.sin6_addr, tos, ttl, | |
e7f70af1 | 2112 | label, src_port, dst_port, !udp_sum); |
e4c7ed41 CW |
2113 | #endif |
2114 | } | |
0e6fbc5b | 2115 | |
4ad16930 | 2116 | return; |
d342894c | 2117 | |
2118 | drop: | |
2119 | dev->stats.tx_dropped++; | |
2120 | goto tx_free; | |
2121 | ||
f491e56d JB |
2122 | xmit_tx_error: |
2123 | /* skb is already freed. */ | |
2124 | skb = NULL; | |
49560532 PS |
2125 | rt_tx_error: |
2126 | ip_rt_put(rt); | |
d342894c | 2127 | tx_error: |
2128 | dev->stats.tx_errors++; | |
2129 | tx_free: | |
2130 | dev_kfree_skb(skb); | |
d342894c | 2131 | } |
2132 | ||
6681712d DS |
2133 | /* Transmit local packets over Vxlan |
2134 | * | |
2135 | * Outer IP header inherits ECN and DF from inner header. | |
2136 | * Outer UDP destination is the VXLAN assigned port. | |
2137 | * source port is based on hash of flow | |
2138 | */ | |
2139 | static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev) | |
2140 | { | |
2141 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3093fbe7 | 2142 | const struct ip_tunnel_info *info; |
6681712d DS |
2143 | struct ethhdr *eth; |
2144 | bool did_rsc = false; | |
8f646c92 | 2145 | struct vxlan_rdst *rdst, *fdst = NULL; |
6681712d | 2146 | struct vxlan_fdb *f; |
6681712d | 2147 | |
61adedf3 | 2148 | info = skb_tunnel_info(skb); |
3093fbe7 | 2149 | |
6681712d | 2150 | skb_reset_mac_header(skb); |
6681712d | 2151 | |
47e5d1b0 JB |
2152 | if (vxlan->flags & VXLAN_F_COLLECT_METADATA) { |
2153 | if (info && info->mode & IP_TUNNEL_INFO_TX) | |
2154 | vxlan_xmit_one(skb, dev, NULL, false); | |
2155 | else | |
2156 | kfree_skb(skb); | |
2157 | return NETDEV_TX_OK; | |
2158 | } | |
2159 | ||
2160 | if (vxlan->flags & VXLAN_F_PROXY) { | |
2161 | eth = eth_hdr(skb); | |
f564f45c CW |
2162 | if (ntohs(eth->h_proto) == ETH_P_ARP) |
2163 | return arp_reduce(dev, skb); | |
2164 | #if IS_ENABLED(CONFIG_IPV6) | |
2165 | else if (ntohs(eth->h_proto) == ETH_P_IPV6 && | |
91269e39 LR |
2166 | pskb_may_pull(skb, sizeof(struct ipv6hdr) |
2167 | + sizeof(struct nd_msg)) && | |
f564f45c CW |
2168 | ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) { |
2169 | struct nd_msg *msg; | |
2170 | ||
2171 | msg = (struct nd_msg *)skb_transport_header(skb); | |
2172 | if (msg->icmph.icmp6_code == 0 && | |
2173 | msg->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION) | |
2174 | return neigh_reduce(dev, skb); | |
2175 | } | |
2176 | #endif | |
2177 | } | |
6681712d | 2178 | |
47e5d1b0 | 2179 | eth = eth_hdr(skb); |
6681712d | 2180 | f = vxlan_find_mac(vxlan, eth->h_dest); |
ae884082 DS |
2181 | did_rsc = false; |
2182 | ||
2183 | if (f && (f->flags & NTF_ROUTER) && (vxlan->flags & VXLAN_F_RSC) && | |
e15a00aa CW |
2184 | (ntohs(eth->h_proto) == ETH_P_IP || |
2185 | ntohs(eth->h_proto) == ETH_P_IPV6)) { | |
ae884082 DS |
2186 | did_rsc = route_shortcircuit(dev, skb); |
2187 | if (did_rsc) | |
2188 | f = vxlan_find_mac(vxlan, eth->h_dest); | |
2189 | } | |
2190 | ||
6681712d | 2191 | if (f == NULL) { |
afbd8bae MR |
2192 | f = vxlan_find_mac(vxlan, all_zeros_mac); |
2193 | if (f == NULL) { | |
2194 | if ((vxlan->flags & VXLAN_F_L2MISS) && | |
2195 | !is_multicast_ether_addr(eth->h_dest)) | |
2196 | vxlan_fdb_miss(vxlan, eth->h_dest); | |
2197 | ||
2198 | dev->stats.tx_dropped++; | |
8f646c92 | 2199 | kfree_skb(skb); |
afbd8bae MR |
2200 | return NETDEV_TX_OK; |
2201 | } | |
2202 | } | |
6681712d | 2203 | |
afbd8bae MR |
2204 | list_for_each_entry_rcu(rdst, &f->remotes, list) { |
2205 | struct sk_buff *skb1; | |
6681712d | 2206 | |
8f646c92 ED |
2207 | if (!fdst) { |
2208 | fdst = rdst; | |
2209 | continue; | |
2210 | } | |
afbd8bae MR |
2211 | skb1 = skb_clone(skb, GFP_ATOMIC); |
2212 | if (skb1) | |
2213 | vxlan_xmit_one(skb1, dev, rdst, did_rsc); | |
6681712d DS |
2214 | } |
2215 | ||
8f646c92 ED |
2216 | if (fdst) |
2217 | vxlan_xmit_one(skb, dev, fdst, did_rsc); | |
2218 | else | |
2219 | kfree_skb(skb); | |
4ad16930 | 2220 | return NETDEV_TX_OK; |
6681712d DS |
2221 | } |
2222 | ||
d342894c | 2223 | /* Walk the forwarding table and purge stale entries */ |
2224 | static void vxlan_cleanup(unsigned long arg) | |
2225 | { | |
2226 | struct vxlan_dev *vxlan = (struct vxlan_dev *) arg; | |
2227 | unsigned long next_timer = jiffies + FDB_AGE_INTERVAL; | |
2228 | unsigned int h; | |
2229 | ||
2230 | if (!netif_running(vxlan->dev)) | |
2231 | return; | |
2232 | ||
d342894c | 2233 | for (h = 0; h < FDB_HASH_SIZE; ++h) { |
2234 | struct hlist_node *p, *n; | |
14e1d0fa SD |
2235 | |
2236 | spin_lock_bh(&vxlan->hash_lock); | |
d342894c | 2237 | hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) { |
2238 | struct vxlan_fdb *f | |
2239 | = container_of(p, struct vxlan_fdb, hlist); | |
2240 | unsigned long timeout; | |
2241 | ||
3c172868 | 2242 | if (f->state & NUD_PERMANENT) |
d342894c | 2243 | continue; |
2244 | ||
0dfbdf41 | 2245 | timeout = f->used + vxlan->cfg.age_interval * HZ; |
d342894c | 2246 | if (time_before_eq(timeout, jiffies)) { |
2247 | netdev_dbg(vxlan->dev, | |
2248 | "garbage collect %pM\n", | |
2249 | f->eth_addr); | |
2250 | f->state = NUD_STALE; | |
2251 | vxlan_fdb_destroy(vxlan, f); | |
2252 | } else if (time_before(timeout, next_timer)) | |
2253 | next_timer = timeout; | |
2254 | } | |
14e1d0fa | 2255 | spin_unlock_bh(&vxlan->hash_lock); |
d342894c | 2256 | } |
d342894c | 2257 | |
2258 | mod_timer(&vxlan->age_timer, next_timer); | |
2259 | } | |
2260 | ||
9c2e24e1 PS |
2261 | static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan) |
2262 | { | |
56ef9c90 | 2263 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
54bfd872 | 2264 | __be32 vni = vxlan->default_dst.remote_vni; |
9c2e24e1 | 2265 | |
56ef9c90 | 2266 | spin_lock(&vn->sock_lock); |
9c2e24e1 | 2267 | hlist_add_head_rcu(&vxlan->hlist, vni_head(vs, vni)); |
56ef9c90 | 2268 | spin_unlock(&vn->sock_lock); |
9c2e24e1 PS |
2269 | } |
2270 | ||
d342894c | 2271 | /* Setup stats when device is created */ |
2272 | static int vxlan_init(struct net_device *dev) | |
2273 | { | |
1c213bd2 | 2274 | dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); |
e8171045 | 2275 | if (!dev->tstats) |
d342894c | 2276 | return -ENOMEM; |
2277 | ||
2278 | return 0; | |
2279 | } | |
2280 | ||
ba609e9b | 2281 | static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan) |
afbd8bae MR |
2282 | { |
2283 | struct vxlan_fdb *f; | |
2284 | ||
2285 | spin_lock_bh(&vxlan->hash_lock); | |
2286 | f = __vxlan_find_mac(vxlan, all_zeros_mac); | |
2287 | if (f) | |
2288 | vxlan_fdb_destroy(vxlan, f); | |
2289 | spin_unlock_bh(&vxlan->hash_lock); | |
2290 | } | |
2291 | ||
ebf4063e SH |
2292 | static void vxlan_uninit(struct net_device *dev) |
2293 | { | |
2294 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
ebf4063e | 2295 | |
ba609e9b | 2296 | vxlan_fdb_delete_default(vxlan); |
afbd8bae | 2297 | |
ebf4063e SH |
2298 | free_percpu(dev->tstats); |
2299 | } | |
2300 | ||
d342894c | 2301 | /* Start ageing timer and join group when device is brought up */ |
2302 | static int vxlan_open(struct net_device *dev) | |
2303 | { | |
2304 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
205f356d | 2305 | int ret; |
56ef9c90 | 2306 | |
205f356d JB |
2307 | ret = vxlan_sock_add(vxlan); |
2308 | if (ret < 0) | |
2309 | return ret; | |
d342894c | 2310 | |
79d4a94f | 2311 | if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) { |
56ef9c90 | 2312 | ret = vxlan_igmp_join(vxlan); |
bef0057b MRL |
2313 | if (ret == -EADDRINUSE) |
2314 | ret = 0; | |
56ef9c90 | 2315 | if (ret) { |
205f356d | 2316 | vxlan_sock_release(vxlan); |
56ef9c90 MRL |
2317 | return ret; |
2318 | } | |
d342894c | 2319 | } |
2320 | ||
0dfbdf41 | 2321 | if (vxlan->cfg.age_interval) |
d342894c | 2322 | mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL); |
2323 | ||
56ef9c90 | 2324 | return ret; |
d342894c | 2325 | } |
2326 | ||
2327 | /* Purge the forwarding table */ | |
2328 | static void vxlan_flush(struct vxlan_dev *vxlan) | |
2329 | { | |
31fec5aa | 2330 | unsigned int h; |
d342894c | 2331 | |
2332 | spin_lock_bh(&vxlan->hash_lock); | |
2333 | for (h = 0; h < FDB_HASH_SIZE; ++h) { | |
2334 | struct hlist_node *p, *n; | |
2335 | hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) { | |
2336 | struct vxlan_fdb *f | |
2337 | = container_of(p, struct vxlan_fdb, hlist); | |
afbd8bae MR |
2338 | /* the all_zeros_mac entry is deleted at vxlan_uninit */ |
2339 | if (!is_zero_ether_addr(f->eth_addr)) | |
2340 | vxlan_fdb_destroy(vxlan, f); | |
d342894c | 2341 | } |
2342 | } | |
2343 | spin_unlock_bh(&vxlan->hash_lock); | |
2344 | } | |
2345 | ||
2346 | /* Cleanup timer and forwarding table on shutdown */ | |
2347 | static int vxlan_stop(struct net_device *dev) | |
2348 | { | |
2349 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
f01ec1c0 | 2350 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
56ef9c90 | 2351 | int ret = 0; |
d342894c | 2352 | |
24c0e683 | 2353 | if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip) && |
f13b1689 | 2354 | !vxlan_group_used(vn, vxlan)) |
56ef9c90 | 2355 | ret = vxlan_igmp_leave(vxlan); |
d342894c | 2356 | |
2357 | del_timer_sync(&vxlan->age_timer); | |
2358 | ||
2359 | vxlan_flush(vxlan); | |
205f356d | 2360 | vxlan_sock_release(vxlan); |
d342894c | 2361 | |
56ef9c90 | 2362 | return ret; |
d342894c | 2363 | } |
2364 | ||
d342894c | 2365 | /* Stub, nothing needs to be done. */ |
2366 | static void vxlan_set_multicast_list(struct net_device *dev) | |
2367 | { | |
2368 | } | |
2369 | ||
72564b59 DW |
2370 | static int __vxlan_change_mtu(struct net_device *dev, |
2371 | struct net_device *lowerdev, | |
2372 | struct vxlan_rdst *dst, int new_mtu, bool strict) | |
345010b5 | 2373 | { |
72564b59 | 2374 | int max_mtu = IP_MAX_MTU; |
345010b5 | 2375 | |
72564b59 DW |
2376 | if (lowerdev) |
2377 | max_mtu = lowerdev->mtu; | |
345010b5 DB |
2378 | |
2379 | if (dst->remote_ip.sa.sa_family == AF_INET6) | |
72564b59 | 2380 | max_mtu -= VXLAN6_HEADROOM; |
345010b5 | 2381 | else |
72564b59 | 2382 | max_mtu -= VXLAN_HEADROOM; |
345010b5 | 2383 | |
72564b59 | 2384 | if (new_mtu < 68) |
345010b5 DB |
2385 | return -EINVAL; |
2386 | ||
72564b59 DW |
2387 | if (new_mtu > max_mtu) { |
2388 | if (strict) | |
2389 | return -EINVAL; | |
2390 | ||
2391 | new_mtu = max_mtu; | |
2392 | } | |
2393 | ||
345010b5 DB |
2394 | dev->mtu = new_mtu; |
2395 | return 0; | |
2396 | } | |
2397 | ||
72564b59 DW |
2398 | static int vxlan_change_mtu(struct net_device *dev, int new_mtu) |
2399 | { | |
2400 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
2401 | struct vxlan_rdst *dst = &vxlan->default_dst; | |
2402 | struct net_device *lowerdev = __dev_get_by_index(vxlan->net, | |
2403 | dst->remote_ifindex); | |
2404 | return __vxlan_change_mtu(dev, lowerdev, dst, new_mtu, true); | |
2405 | } | |
2406 | ||
fc4099f1 PS |
2407 | static int vxlan_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) |
2408 | { | |
2409 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
2410 | struct ip_tunnel_info *info = skb_tunnel_info(skb); | |
2411 | __be16 sport, dport; | |
2412 | ||
2413 | sport = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min, | |
2414 | vxlan->cfg.port_max, true); | |
2415 | dport = info->key.tp_dst ? : vxlan->cfg.dst_port; | |
2416 | ||
239e944f | 2417 | if (ip_tunnel_info_af(info) == AF_INET) { |
1a8496ba JB |
2418 | struct rtable *rt; |
2419 | ||
239e944f JB |
2420 | if (!vxlan->vn4_sock) |
2421 | return -EINVAL; | |
1a8496ba JB |
2422 | rt = vxlan_get_route(vxlan, skb, 0, info->key.tos, |
2423 | info->key.u.ipv4.dst, | |
0c1d70af | 2424 | &info->key.u.ipv4.src, NULL, info); |
1a8496ba JB |
2425 | if (IS_ERR(rt)) |
2426 | return PTR_ERR(rt); | |
2427 | ip_rt_put(rt); | |
239e944f JB |
2428 | } else { |
2429 | #if IS_ENABLED(CONFIG_IPV6) | |
2430 | struct dst_entry *ndst; | |
2431 | ||
2432 | if (!vxlan->vn6_sock) | |
2433 | return -EINVAL; | |
1400615d | 2434 | ndst = vxlan6_get_route(vxlan, skb, 0, info->key.tos, |
e7f70af1 | 2435 | info->key.label, &info->key.u.ipv6.dst, |
db3c6139 | 2436 | &info->key.u.ipv6.src, NULL, info); |
239e944f JB |
2437 | if (IS_ERR(ndst)) |
2438 | return PTR_ERR(ndst); | |
2439 | dst_release(ndst); | |
239e944f JB |
2440 | #else /* !CONFIG_IPV6 */ |
2441 | return -EPFNOSUPPORT; | |
2442 | #endif | |
2443 | } | |
1a8496ba JB |
2444 | info->key.tp_src = sport; |
2445 | info->key.tp_dst = dport; | |
239e944f | 2446 | return 0; |
fc4099f1 PS |
2447 | } |
2448 | ||
0c867c9b | 2449 | static const struct net_device_ops vxlan_netdev_ether_ops = { |
d342894c | 2450 | .ndo_init = vxlan_init, |
ebf4063e | 2451 | .ndo_uninit = vxlan_uninit, |
d342894c | 2452 | .ndo_open = vxlan_open, |
2453 | .ndo_stop = vxlan_stop, | |
2454 | .ndo_start_xmit = vxlan_xmit, | |
e8171045 | 2455 | .ndo_get_stats64 = ip_tunnel_get_stats64, |
d342894c | 2456 | .ndo_set_rx_mode = vxlan_set_multicast_list, |
345010b5 | 2457 | .ndo_change_mtu = vxlan_change_mtu, |
d342894c | 2458 | .ndo_validate_addr = eth_validate_addr, |
2459 | .ndo_set_mac_address = eth_mac_addr, | |
2460 | .ndo_fdb_add = vxlan_fdb_add, | |
2461 | .ndo_fdb_del = vxlan_fdb_delete, | |
2462 | .ndo_fdb_dump = vxlan_fdb_dump, | |
fc4099f1 | 2463 | .ndo_fill_metadata_dst = vxlan_fill_metadata_dst, |
d342894c | 2464 | }; |
2465 | ||
e1e5314d JB |
2466 | static const struct net_device_ops vxlan_netdev_raw_ops = { |
2467 | .ndo_init = vxlan_init, | |
2468 | .ndo_uninit = vxlan_uninit, | |
2469 | .ndo_open = vxlan_open, | |
2470 | .ndo_stop = vxlan_stop, | |
2471 | .ndo_start_xmit = vxlan_xmit, | |
2472 | .ndo_get_stats64 = ip_tunnel_get_stats64, | |
2473 | .ndo_change_mtu = vxlan_change_mtu, | |
2474 | .ndo_fill_metadata_dst = vxlan_fill_metadata_dst, | |
2475 | }; | |
2476 | ||
d342894c | 2477 | /* Info for udev, that this is a virtual tunnel endpoint */ |
2478 | static struct device_type vxlan_type = { | |
2479 | .name = "vxlan", | |
2480 | }; | |
2481 | ||
e5de25dc | 2482 | /* Calls the ndo_udp_tunnel_add of the caller in order to |
35e42379 | 2483 | * supply the listening VXLAN udp ports. Callers are expected |
e5de25dc | 2484 | * to implement the ndo_udp_tunnel_add. |
53cf5275 | 2485 | */ |
b7aade15 | 2486 | static void vxlan_push_rx_ports(struct net_device *dev) |
53cf5275 JG |
2487 | { |
2488 | struct vxlan_sock *vs; | |
2489 | struct net *net = dev_net(dev); | |
2490 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
35e42379 | 2491 | unsigned int i; |
53cf5275 JG |
2492 | |
2493 | spin_lock(&vn->sock_lock); | |
2494 | for (i = 0; i < PORT_HASH_SIZE; ++i) { | |
e7b3db5e AD |
2495 | hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) |
2496 | udp_tunnel_push_rx_port(dev, vs->sock, | |
b9adcd69 AD |
2497 | (vs->flags & VXLAN_F_GPE) ? |
2498 | UDP_TUNNEL_TYPE_VXLAN_GPE : | |
e7b3db5e | 2499 | UDP_TUNNEL_TYPE_VXLAN); |
53cf5275 JG |
2500 | } |
2501 | spin_unlock(&vn->sock_lock); | |
2502 | } | |
53cf5275 | 2503 | |
d342894c | 2504 | /* Initialize the device structure. */ |
2505 | static void vxlan_setup(struct net_device *dev) | |
2506 | { | |
2507 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
31fec5aa | 2508 | unsigned int h; |
d342894c | 2509 | |
65226ef8 JB |
2510 | eth_hw_addr_random(dev); |
2511 | ether_setup(dev); | |
2512 | ||
ebf4063e | 2513 | dev->destructor = free_netdev; |
d342894c | 2514 | SET_NETDEV_DEVTYPE(dev, &vxlan_type); |
2515 | ||
d342894c | 2516 | dev->features |= NETIF_F_LLTX; |
d6727fe3 | 2517 | dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM; |
0afb1666 | 2518 | dev->features |= NETIF_F_RXCSUM; |
05c0db08 | 2519 | dev->features |= NETIF_F_GSO_SOFTWARE; |
0afb1666 | 2520 | |
1eaa8178 PS |
2521 | dev->vlan_features = dev->features; |
2522 | dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX; | |
0afb1666 | 2523 | dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM; |
05c0db08 | 2524 | dev->hw_features |= NETIF_F_GSO_SOFTWARE; |
1eaa8178 | 2525 | dev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX; |
02875878 | 2526 | netif_keep_dst(dev); |
0c867c9b | 2527 | dev->priv_flags |= IFF_NO_QUEUE; |
d342894c | 2528 | |
553675fb | 2529 | INIT_LIST_HEAD(&vxlan->next); |
d342894c | 2530 | spin_lock_init(&vxlan->hash_lock); |
2531 | ||
2532 | init_timer_deferrable(&vxlan->age_timer); | |
2533 | vxlan->age_timer.function = vxlan_cleanup; | |
2534 | vxlan->age_timer.data = (unsigned long) vxlan; | |
2535 | ||
0dfbdf41 | 2536 | vxlan->cfg.dst_port = htons(vxlan_port); |
05f47d69 | 2537 | |
d342894c | 2538 | vxlan->dev = dev; |
2539 | ||
58ce31cc TH |
2540 | gro_cells_init(&vxlan->gro_cells, dev); |
2541 | ||
d342894c | 2542 | for (h = 0; h < FDB_HASH_SIZE; ++h) |
2543 | INIT_HLIST_HEAD(&vxlan->fdb_head[h]); | |
2544 | } | |
2545 | ||
0c867c9b JB |
2546 | static void vxlan_ether_setup(struct net_device *dev) |
2547 | { | |
0c867c9b JB |
2548 | dev->priv_flags &= ~IFF_TX_SKB_SHARING; |
2549 | dev->priv_flags |= IFF_LIVE_ADDR_CHANGE; | |
2550 | dev->netdev_ops = &vxlan_netdev_ether_ops; | |
2551 | } | |
2552 | ||
e1e5314d JB |
2553 | static void vxlan_raw_setup(struct net_device *dev) |
2554 | { | |
65226ef8 | 2555 | dev->header_ops = NULL; |
e1e5314d JB |
2556 | dev->type = ARPHRD_NONE; |
2557 | dev->hard_header_len = 0; | |
2558 | dev->addr_len = 0; | |
e1e5314d JB |
2559 | dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST; |
2560 | dev->netdev_ops = &vxlan_netdev_raw_ops; | |
2561 | } | |
2562 | ||
d342894c | 2563 | static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = { |
2564 | [IFLA_VXLAN_ID] = { .type = NLA_U32 }, | |
5d174dd8 | 2565 | [IFLA_VXLAN_GROUP] = { .len = FIELD_SIZEOF(struct iphdr, daddr) }, |
e4c7ed41 | 2566 | [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) }, |
d342894c | 2567 | [IFLA_VXLAN_LINK] = { .type = NLA_U32 }, |
2568 | [IFLA_VXLAN_LOCAL] = { .len = FIELD_SIZEOF(struct iphdr, saddr) }, | |
e4c7ed41 | 2569 | [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) }, |
d342894c | 2570 | [IFLA_VXLAN_TOS] = { .type = NLA_U8 }, |
2571 | [IFLA_VXLAN_TTL] = { .type = NLA_U8 }, | |
e7f70af1 | 2572 | [IFLA_VXLAN_LABEL] = { .type = NLA_U32 }, |
d342894c | 2573 | [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 }, |
2574 | [IFLA_VXLAN_AGEING] = { .type = NLA_U32 }, | |
2575 | [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 }, | |
05f47d69 | 2576 | [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) }, |
e4f67add DS |
2577 | [IFLA_VXLAN_PROXY] = { .type = NLA_U8 }, |
2578 | [IFLA_VXLAN_RSC] = { .type = NLA_U8 }, | |
2579 | [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 }, | |
2580 | [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 }, | |
f8a9b1bc | 2581 | [IFLA_VXLAN_COLLECT_METADATA] = { .type = NLA_U8 }, |
823aa873 | 2582 | [IFLA_VXLAN_PORT] = { .type = NLA_U16 }, |
5c91ae08 TH |
2583 | [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 }, |
2584 | [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 }, | |
2585 | [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 }, | |
dfd8645e TH |
2586 | [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 }, |
2587 | [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 }, | |
3511494c | 2588 | [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, }, |
e1e5314d | 2589 | [IFLA_VXLAN_GPE] = { .type = NLA_FLAG, }, |
0ace2ca8 | 2590 | [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG }, |
d342894c | 2591 | }; |
2592 | ||
2593 | static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[]) | |
2594 | { | |
2595 | if (tb[IFLA_ADDRESS]) { | |
2596 | if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) { | |
2597 | pr_debug("invalid link address (not ethernet)\n"); | |
2598 | return -EINVAL; | |
2599 | } | |
2600 | ||
2601 | if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) { | |
2602 | pr_debug("invalid all zero ethernet address\n"); | |
2603 | return -EADDRNOTAVAIL; | |
2604 | } | |
2605 | } | |
2606 | ||
2607 | if (!data) | |
2608 | return -EINVAL; | |
2609 | ||
2610 | if (data[IFLA_VXLAN_ID]) { | |
2611 | __u32 id = nla_get_u32(data[IFLA_VXLAN_ID]); | |
2612 | if (id >= VXLAN_VID_MASK) | |
2613 | return -ERANGE; | |
2614 | } | |
2615 | ||
05f47d69 | 2616 | if (data[IFLA_VXLAN_PORT_RANGE]) { |
2617 | const struct ifla_vxlan_port_range *p | |
2618 | = nla_data(data[IFLA_VXLAN_PORT_RANGE]); | |
2619 | ||
2620 | if (ntohs(p->high) < ntohs(p->low)) { | |
2621 | pr_debug("port range %u .. %u not valid\n", | |
2622 | ntohs(p->low), ntohs(p->high)); | |
2623 | return -EINVAL; | |
2624 | } | |
2625 | } | |
2626 | ||
d342894c | 2627 | return 0; |
2628 | } | |
2629 | ||
1b13c97f YB |
2630 | static void vxlan_get_drvinfo(struct net_device *netdev, |
2631 | struct ethtool_drvinfo *drvinfo) | |
2632 | { | |
2633 | strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version)); | |
2634 | strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver)); | |
2635 | } | |
2636 | ||
2637 | static const struct ethtool_ops vxlan_ethtool_ops = { | |
2638 | .get_drvinfo = vxlan_get_drvinfo, | |
2639 | .get_link = ethtool_op_get_link, | |
2640 | }; | |
2641 | ||
3ee64f39 TH |
2642 | static struct socket *vxlan_create_sock(struct net *net, bool ipv6, |
2643 | __be16 port, u32 flags) | |
553675fb | 2644 | { |
e4c7ed41 | 2645 | struct socket *sock; |
3ee64f39 TH |
2646 | struct udp_port_cfg udp_conf; |
2647 | int err; | |
e4c7ed41 | 2648 | |
3ee64f39 | 2649 | memset(&udp_conf, 0, sizeof(udp_conf)); |
e4c7ed41 | 2650 | |
3ee64f39 TH |
2651 | if (ipv6) { |
2652 | udp_conf.family = AF_INET6; | |
3ee64f39 | 2653 | udp_conf.use_udp6_rx_checksums = |
3dc2b6a8 | 2654 | !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX); |
a43a9ef6 | 2655 | udp_conf.ipv6_v6only = 1; |
3ee64f39 TH |
2656 | } else { |
2657 | udp_conf.family = AF_INET; | |
e4c7ed41 | 2658 | } |
e4c7ed41 | 2659 | |
3ee64f39 | 2660 | udp_conf.local_udp_port = port; |
553675fb | 2661 | |
3ee64f39 TH |
2662 | /* Open UDP socket */ |
2663 | err = udp_sock_create(net, &udp_conf, &sock); | |
2664 | if (err < 0) | |
2665 | return ERR_PTR(err); | |
e4c7ed41 | 2666 | |
39deb2c7 | 2667 | return sock; |
e4c7ed41 CW |
2668 | } |
2669 | ||
2670 | /* Create new listen socket if needed */ | |
b1be00a6 JB |
2671 | static struct vxlan_sock *vxlan_socket_create(struct net *net, bool ipv6, |
2672 | __be16 port, u32 flags) | |
e4c7ed41 CW |
2673 | { |
2674 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
2675 | struct vxlan_sock *vs; | |
2676 | struct socket *sock; | |
e4c7ed41 | 2677 | unsigned int h; |
acbf74a7 | 2678 | struct udp_tunnel_sock_cfg tunnel_cfg; |
e4c7ed41 | 2679 | |
dc01e7d3 | 2680 | vs = kzalloc(sizeof(*vs), GFP_KERNEL); |
e4c7ed41 CW |
2681 | if (!vs) |
2682 | return ERR_PTR(-ENOMEM); | |
2683 | ||
2684 | for (h = 0; h < VNI_HASH_SIZE; ++h) | |
2685 | INIT_HLIST_HEAD(&vs->vni_list[h]); | |
2686 | ||
3ee64f39 | 2687 | sock = vxlan_create_sock(net, ipv6, port, flags); |
39deb2c7 | 2688 | if (IS_ERR(sock)) { |
56ef9c90 MRL |
2689 | pr_info("Cannot bind port %d, err=%ld\n", ntohs(port), |
2690 | PTR_ERR(sock)); | |
553675fb | 2691 | kfree(vs); |
e50fddc8 | 2692 | return ERR_CAST(sock); |
553675fb | 2693 | } |
e4c7ed41 CW |
2694 | |
2695 | vs->sock = sock; | |
9c2e24e1 | 2696 | atomic_set(&vs->refcnt, 1); |
af33c1ad | 2697 | vs->flags = (flags & VXLAN_F_RCV_FLAGS); |
553675fb | 2698 | |
9c2e24e1 PS |
2699 | spin_lock(&vn->sock_lock); |
2700 | hlist_add_head_rcu(&vs->hlist, vs_head(net, port)); | |
e7b3db5e | 2701 | udp_tunnel_notify_add_rx_port(sock, |
b9adcd69 AD |
2702 | (vs->flags & VXLAN_F_GPE) ? |
2703 | UDP_TUNNEL_TYPE_VXLAN_GPE : | |
e7b3db5e | 2704 | UDP_TUNNEL_TYPE_VXLAN); |
9c2e24e1 | 2705 | spin_unlock(&vn->sock_lock); |
553675fb | 2706 | |
2707 | /* Mark socket as an encapsulation socket. */ | |
5602c48c | 2708 | memset(&tunnel_cfg, 0, sizeof(tunnel_cfg)); |
acbf74a7 AZ |
2709 | tunnel_cfg.sk_user_data = vs; |
2710 | tunnel_cfg.encap_type = 1; | |
f2d1968e | 2711 | tunnel_cfg.encap_rcv = vxlan_rcv; |
acbf74a7 | 2712 | tunnel_cfg.encap_destroy = NULL; |
5602c48c TH |
2713 | tunnel_cfg.gro_receive = vxlan_gro_receive; |
2714 | tunnel_cfg.gro_complete = vxlan_gro_complete; | |
acbf74a7 AZ |
2715 | |
2716 | setup_udp_tunnel_sock(net, sock, &tunnel_cfg); | |
e4c7ed41 | 2717 | |
9c2e24e1 PS |
2718 | return vs; |
2719 | } | |
2720 | ||
b1be00a6 | 2721 | static int __vxlan_sock_add(struct vxlan_dev *vxlan, bool ipv6) |
9c2e24e1 | 2722 | { |
205f356d JB |
2723 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
2724 | struct vxlan_sock *vs = NULL; | |
9c2e24e1 | 2725 | |
205f356d | 2726 | if (!vxlan->cfg.no_share) { |
56ef9c90 | 2727 | spin_lock(&vn->sock_lock); |
205f356d JB |
2728 | vs = vxlan_find_sock(vxlan->net, ipv6 ? AF_INET6 : AF_INET, |
2729 | vxlan->cfg.dst_port, vxlan->flags); | |
2730 | if (vs && !atomic_add_unless(&vs->refcnt, 1, 0)) { | |
56ef9c90 | 2731 | spin_unlock(&vn->sock_lock); |
205f356d | 2732 | return -EBUSY; |
56ef9c90 MRL |
2733 | } |
2734 | spin_unlock(&vn->sock_lock); | |
2735 | } | |
205f356d | 2736 | if (!vs) |
b1be00a6 JB |
2737 | vs = vxlan_socket_create(vxlan->net, ipv6, |
2738 | vxlan->cfg.dst_port, vxlan->flags); | |
205f356d JB |
2739 | if (IS_ERR(vs)) |
2740 | return PTR_ERR(vs); | |
b1be00a6 JB |
2741 | #if IS_ENABLED(CONFIG_IPV6) |
2742 | if (ipv6) | |
2743 | vxlan->vn6_sock = vs; | |
2744 | else | |
2745 | #endif | |
2746 | vxlan->vn4_sock = vs; | |
205f356d JB |
2747 | vxlan_vs_add_dev(vs, vxlan); |
2748 | return 0; | |
553675fb | 2749 | } |
2750 | ||
b1be00a6 JB |
2751 | static int vxlan_sock_add(struct vxlan_dev *vxlan) |
2752 | { | |
2753 | bool ipv6 = vxlan->flags & VXLAN_F_IPV6; | |
2754 | bool metadata = vxlan->flags & VXLAN_F_COLLECT_METADATA; | |
2755 | int ret = 0; | |
2756 | ||
2757 | vxlan->vn4_sock = NULL; | |
2758 | #if IS_ENABLED(CONFIG_IPV6) | |
2759 | vxlan->vn6_sock = NULL; | |
2760 | if (ipv6 || metadata) | |
2761 | ret = __vxlan_sock_add(vxlan, true); | |
2762 | #endif | |
2763 | if (!ret && (!ipv6 || metadata)) | |
2764 | ret = __vxlan_sock_add(vxlan, false); | |
2765 | if (ret < 0) | |
2766 | vxlan_sock_release(vxlan); | |
2767 | return ret; | |
2768 | } | |
2769 | ||
0dfbdf41 TG |
2770 | static int vxlan_dev_configure(struct net *src_net, struct net_device *dev, |
2771 | struct vxlan_config *conf) | |
d342894c | 2772 | { |
33564bbb | 2773 | struct vxlan_net *vn = net_generic(src_net, vxlan_net_id); |
07b9b37c | 2774 | struct vxlan_dev *vxlan = netdev_priv(dev), *tmp; |
c7995c43 | 2775 | struct vxlan_rdst *dst = &vxlan->default_dst; |
b1be00a6 | 2776 | unsigned short needed_headroom = ETH_HLEN; |
d342894c | 2777 | int err; |
e4c7ed41 | 2778 | bool use_ipv6 = false; |
0dfbdf41 | 2779 | __be16 default_port = vxlan->cfg.dst_port; |
7e059158 | 2780 | struct net_device *lowerdev = NULL; |
d342894c | 2781 | |
e1e5314d JB |
2782 | if (conf->flags & VXLAN_F_GPE) { |
2783 | if (conf->flags & ~VXLAN_F_ALLOWED_GPE) | |
2784 | return -EINVAL; | |
2785 | /* For now, allow GPE only together with COLLECT_METADATA. | |
2786 | * This can be relaxed later; in such case, the other side | |
2787 | * of the PtP link will have to be provided. | |
2788 | */ | |
2789 | if (!(conf->flags & VXLAN_F_COLLECT_METADATA)) | |
2790 | return -EINVAL; | |
2791 | ||
2792 | vxlan_raw_setup(dev); | |
2793 | } else { | |
2794 | vxlan_ether_setup(dev); | |
2795 | } | |
0c867c9b | 2796 | |
33564bbb | 2797 | vxlan->net = src_net; |
f01ec1c0 | 2798 | |
0dfbdf41 | 2799 | dst->remote_vni = conf->vni; |
d342894c | 2800 | |
0dfbdf41 | 2801 | memcpy(&dst->remote_ip, &conf->remote_ip, sizeof(conf->remote_ip)); |
d342894c | 2802 | |
0dfbdf41 TG |
2803 | /* Unless IPv6 is explicitly requested, assume IPv4 */ |
2804 | if (!dst->remote_ip.sa.sa_family) | |
2805 | dst->remote_ip.sa.sa_family = AF_INET; | |
e4c7ed41 | 2806 | |
0dfbdf41 | 2807 | if (dst->remote_ip.sa.sa_family == AF_INET6 || |
057ba29b JB |
2808 | vxlan->cfg.saddr.sa.sa_family == AF_INET6) { |
2809 | if (!IS_ENABLED(CONFIG_IPV6)) | |
2810 | return -EPFNOSUPPORT; | |
e4c7ed41 | 2811 | use_ipv6 = true; |
b1be00a6 | 2812 | vxlan->flags |= VXLAN_F_IPV6; |
057ba29b | 2813 | } |
d342894c | 2814 | |
e7f70af1 DB |
2815 | if (conf->label && !use_ipv6) { |
2816 | pr_info("label only supported in use with IPv6\n"); | |
2817 | return -EINVAL; | |
2818 | } | |
2819 | ||
0dfbdf41 | 2820 | if (conf->remote_ifindex) { |
7e059158 | 2821 | lowerdev = __dev_get_by_index(src_net, conf->remote_ifindex); |
0dfbdf41 | 2822 | dst->remote_ifindex = conf->remote_ifindex; |
34e02aa1 | 2823 | |
2824 | if (!lowerdev) { | |
c7995c43 | 2825 | pr_info("ifindex %d does not exist\n", dst->remote_ifindex); |
34e02aa1 | 2826 | return -ENODEV; |
2827 | } | |
d342894c | 2828 | |
e4c7ed41 CW |
2829 | #if IS_ENABLED(CONFIG_IPV6) |
2830 | if (use_ipv6) { | |
2831 | struct inet6_dev *idev = __in6_dev_get(lowerdev); | |
2832 | if (idev && idev->cnf.disable_ipv6) { | |
2833 | pr_info("IPv6 is disabled via sysctl\n"); | |
2834 | return -EPERM; | |
2835 | } | |
e4c7ed41 CW |
2836 | } |
2837 | #endif | |
2838 | ||
0dfbdf41 | 2839 | if (!conf->mtu) |
e4c7ed41 | 2840 | dev->mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM); |
1ba56fb4 | 2841 | |
b1be00a6 | 2842 | needed_headroom = lowerdev->hard_header_len; |
9dc2ad10 | 2843 | } |
d342894c | 2844 | |
7e059158 DW |
2845 | if (conf->mtu) { |
2846 | err = __vxlan_change_mtu(dev, lowerdev, dst, conf->mtu, false); | |
2847 | if (err) | |
2848 | return err; | |
2849 | } | |
2850 | ||
b1be00a6 JB |
2851 | if (use_ipv6 || conf->flags & VXLAN_F_COLLECT_METADATA) |
2852 | needed_headroom += VXLAN6_HEADROOM; | |
2853 | else | |
2854 | needed_headroom += VXLAN_HEADROOM; | |
2855 | dev->needed_headroom = needed_headroom; | |
2856 | ||
0dfbdf41 | 2857 | memcpy(&vxlan->cfg, conf, sizeof(*conf)); |
e1e5314d JB |
2858 | if (!vxlan->cfg.dst_port) { |
2859 | if (conf->flags & VXLAN_F_GPE) | |
2860 | vxlan->cfg.dst_port = 4790; /* IANA assigned VXLAN-GPE port */ | |
2861 | else | |
2862 | vxlan->cfg.dst_port = default_port; | |
2863 | } | |
0dfbdf41 TG |
2864 | vxlan->flags |= conf->flags; |
2865 | ||
2866 | if (!vxlan->cfg.age_interval) | |
2867 | vxlan->cfg.age_interval = FDB_AGE_DEFAULT; | |
2868 | ||
07b9b37c ND |
2869 | list_for_each_entry(tmp, &vn->vxlan_list, next) { |
2870 | if (tmp->cfg.vni == conf->vni && | |
2871 | (tmp->default_dst.remote_ip.sa.sa_family == AF_INET6 || | |
2872 | tmp->cfg.saddr.sa.sa_family == AF_INET6) == use_ipv6 && | |
2873 | tmp->cfg.dst_port == vxlan->cfg.dst_port && | |
2874 | (tmp->flags & VXLAN_F_RCV_FLAGS) == | |
2875 | (vxlan->flags & VXLAN_F_RCV_FLAGS)) | |
0dfbdf41 | 2876 | return -EEXIST; |
07b9b37c | 2877 | } |
0dfbdf41 TG |
2878 | |
2879 | dev->ethtool_ops = &vxlan_ethtool_ops; | |
2880 | ||
2881 | /* create an fdb entry for a valid default destination */ | |
2882 | if (!vxlan_addr_any(&vxlan->default_dst.remote_ip)) { | |
2883 | err = vxlan_fdb_create(vxlan, all_zeros_mac, | |
2884 | &vxlan->default_dst.remote_ip, | |
2885 | NUD_REACHABLE|NUD_PERMANENT, | |
2886 | NLM_F_EXCL|NLM_F_CREATE, | |
2887 | vxlan->cfg.dst_port, | |
2888 | vxlan->default_dst.remote_vni, | |
2889 | vxlan->default_dst.remote_ifindex, | |
2890 | NTF_SELF); | |
2891 | if (err) | |
2892 | return err; | |
2893 | } | |
2894 | ||
2895 | err = register_netdevice(dev); | |
2896 | if (err) { | |
2897 | vxlan_fdb_delete_default(vxlan); | |
2898 | return err; | |
2899 | } | |
2900 | ||
2901 | list_add(&vxlan->next, &vn->vxlan_list); | |
2902 | ||
2903 | return 0; | |
2904 | } | |
2905 | ||
0dfbdf41 TG |
2906 | static int vxlan_newlink(struct net *src_net, struct net_device *dev, |
2907 | struct nlattr *tb[], struct nlattr *data[]) | |
2908 | { | |
2909 | struct vxlan_config conf; | |
2910 | int err; | |
2911 | ||
0dfbdf41 | 2912 | memset(&conf, 0, sizeof(conf)); |
e277de5f JG |
2913 | |
2914 | if (data[IFLA_VXLAN_ID]) | |
54bfd872 | 2915 | conf.vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID])); |
0dfbdf41 TG |
2916 | |
2917 | if (data[IFLA_VXLAN_GROUP]) { | |
2918 | conf.remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]); | |
2919 | } else if (data[IFLA_VXLAN_GROUP6]) { | |
2920 | if (!IS_ENABLED(CONFIG_IPV6)) | |
2921 | return -EPFNOSUPPORT; | |
2922 | ||
2923 | conf.remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]); | |
2924 | conf.remote_ip.sa.sa_family = AF_INET6; | |
2925 | } | |
2926 | ||
2927 | if (data[IFLA_VXLAN_LOCAL]) { | |
2928 | conf.saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]); | |
2929 | conf.saddr.sa.sa_family = AF_INET; | |
2930 | } else if (data[IFLA_VXLAN_LOCAL6]) { | |
2931 | if (!IS_ENABLED(CONFIG_IPV6)) | |
2932 | return -EPFNOSUPPORT; | |
2933 | ||
2934 | /* TODO: respect scope id */ | |
2935 | conf.saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]); | |
2936 | conf.saddr.sa.sa_family = AF_INET6; | |
2937 | } | |
2938 | ||
2939 | if (data[IFLA_VXLAN_LINK]) | |
2940 | conf.remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]); | |
2941 | ||
d342894c | 2942 | if (data[IFLA_VXLAN_TOS]) |
0dfbdf41 | 2943 | conf.tos = nla_get_u8(data[IFLA_VXLAN_TOS]); |
d342894c | 2944 | |
afb97186 | 2945 | if (data[IFLA_VXLAN_TTL]) |
0dfbdf41 | 2946 | conf.ttl = nla_get_u8(data[IFLA_VXLAN_TTL]); |
afb97186 | 2947 | |
e7f70af1 DB |
2948 | if (data[IFLA_VXLAN_LABEL]) |
2949 | conf.label = nla_get_be32(data[IFLA_VXLAN_LABEL]) & | |
2950 | IPV6_FLOWLABEL_MASK; | |
2951 | ||
d342894c | 2952 | if (!data[IFLA_VXLAN_LEARNING] || nla_get_u8(data[IFLA_VXLAN_LEARNING])) |
0dfbdf41 | 2953 | conf.flags |= VXLAN_F_LEARN; |
d342894c | 2954 | |
2955 | if (data[IFLA_VXLAN_AGEING]) | |
0dfbdf41 | 2956 | conf.age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]); |
d342894c | 2957 | |
e4f67add | 2958 | if (data[IFLA_VXLAN_PROXY] && nla_get_u8(data[IFLA_VXLAN_PROXY])) |
0dfbdf41 | 2959 | conf.flags |= VXLAN_F_PROXY; |
e4f67add DS |
2960 | |
2961 | if (data[IFLA_VXLAN_RSC] && nla_get_u8(data[IFLA_VXLAN_RSC])) | |
0dfbdf41 | 2962 | conf.flags |= VXLAN_F_RSC; |
e4f67add DS |
2963 | |
2964 | if (data[IFLA_VXLAN_L2MISS] && nla_get_u8(data[IFLA_VXLAN_L2MISS])) | |
0dfbdf41 | 2965 | conf.flags |= VXLAN_F_L2MISS; |
e4f67add DS |
2966 | |
2967 | if (data[IFLA_VXLAN_L3MISS] && nla_get_u8(data[IFLA_VXLAN_L3MISS])) | |
0dfbdf41 | 2968 | conf.flags |= VXLAN_F_L3MISS; |
e4f67add | 2969 | |
d342894c | 2970 | if (data[IFLA_VXLAN_LIMIT]) |
0dfbdf41 | 2971 | conf.addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]); |
d342894c | 2972 | |
f8a9b1bc AS |
2973 | if (data[IFLA_VXLAN_COLLECT_METADATA] && |
2974 | nla_get_u8(data[IFLA_VXLAN_COLLECT_METADATA])) | |
2975 | conf.flags |= VXLAN_F_COLLECT_METADATA; | |
2976 | ||
05f47d69 | 2977 | if (data[IFLA_VXLAN_PORT_RANGE]) { |
2978 | const struct ifla_vxlan_port_range *p | |
2979 | = nla_data(data[IFLA_VXLAN_PORT_RANGE]); | |
0dfbdf41 TG |
2980 | conf.port_min = ntohs(p->low); |
2981 | conf.port_max = ntohs(p->high); | |
05f47d69 | 2982 | } |
2983 | ||
823aa873 | 2984 | if (data[IFLA_VXLAN_PORT]) |
0dfbdf41 | 2985 | conf.dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]); |
823aa873 | 2986 | |
6ceb31ca AD |
2987 | if (data[IFLA_VXLAN_UDP_CSUM] && |
2988 | !nla_get_u8(data[IFLA_VXLAN_UDP_CSUM])) | |
2989 | conf.flags |= VXLAN_F_UDP_ZERO_CSUM_TX; | |
359a0ea9 TH |
2990 | |
2991 | if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX] && | |
2992 | nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX])) | |
0dfbdf41 | 2993 | conf.flags |= VXLAN_F_UDP_ZERO_CSUM6_TX; |
359a0ea9 TH |
2994 | |
2995 | if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX] && | |
2996 | nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX])) | |
0dfbdf41 | 2997 | conf.flags |= VXLAN_F_UDP_ZERO_CSUM6_RX; |
359a0ea9 | 2998 | |
dfd8645e TH |
2999 | if (data[IFLA_VXLAN_REMCSUM_TX] && |
3000 | nla_get_u8(data[IFLA_VXLAN_REMCSUM_TX])) | |
0dfbdf41 | 3001 | conf.flags |= VXLAN_F_REMCSUM_TX; |
dfd8645e TH |
3002 | |
3003 | if (data[IFLA_VXLAN_REMCSUM_RX] && | |
3004 | nla_get_u8(data[IFLA_VXLAN_REMCSUM_RX])) | |
0dfbdf41 | 3005 | conf.flags |= VXLAN_F_REMCSUM_RX; |
dfd8645e | 3006 | |
3511494c | 3007 | if (data[IFLA_VXLAN_GBP]) |
0dfbdf41 | 3008 | conf.flags |= VXLAN_F_GBP; |
3511494c | 3009 | |
e1e5314d JB |
3010 | if (data[IFLA_VXLAN_GPE]) |
3011 | conf.flags |= VXLAN_F_GPE; | |
3012 | ||
0ace2ca8 | 3013 | if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL]) |
0dfbdf41 | 3014 | conf.flags |= VXLAN_F_REMCSUM_NOPARTIAL; |
0ace2ca8 | 3015 | |
ce577668 CH |
3016 | if (tb[IFLA_MTU]) |
3017 | conf.mtu = nla_get_u32(tb[IFLA_MTU]); | |
3018 | ||
0dfbdf41 TG |
3019 | err = vxlan_dev_configure(src_net, dev, &conf); |
3020 | switch (err) { | |
3021 | case -ENODEV: | |
3022 | pr_info("ifindex %d does not exist\n", conf.remote_ifindex); | |
3023 | break; | |
1b13c97f | 3024 | |
0dfbdf41 TG |
3025 | case -EPERM: |
3026 | pr_info("IPv6 is disabled via sysctl\n"); | |
3027 | break; | |
d342894c | 3028 | |
0dfbdf41 | 3029 | case -EEXIST: |
54bfd872 | 3030 | pr_info("duplicate VNI %u\n", be32_to_cpu(conf.vni)); |
0dfbdf41 | 3031 | break; |
e1e5314d JB |
3032 | |
3033 | case -EINVAL: | |
3034 | pr_info("unsupported combination of extensions\n"); | |
3035 | break; | |
afbd8bae MR |
3036 | } |
3037 | ||
0dfbdf41 | 3038 | return err; |
d342894c | 3039 | } |
3040 | ||
3041 | static void vxlan_dellink(struct net_device *dev, struct list_head *head) | |
3042 | { | |
3043 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
f01ec1c0 | 3044 | struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id); |
d342894c | 3045 | |
fe5c3561 | 3046 | spin_lock(&vn->sock_lock); |
9c2e24e1 PS |
3047 | if (!hlist_unhashed(&vxlan->hlist)) |
3048 | hlist_del_rcu(&vxlan->hlist); | |
fe5c3561 | 3049 | spin_unlock(&vn->sock_lock); |
3050 | ||
58ce31cc | 3051 | gro_cells_destroy(&vxlan->gro_cells); |
553675fb | 3052 | list_del(&vxlan->next); |
d342894c | 3053 | unregister_netdevice_queue(dev, head); |
3054 | } | |
3055 | ||
3056 | static size_t vxlan_get_size(const struct net_device *dev) | |
3057 | { | |
3058 | ||
3059 | return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */ | |
e4c7ed41 | 3060 | nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */ |
d342894c | 3061 | nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */ |
e4c7ed41 | 3062 | nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */ |
d342894c | 3063 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */ |
3064 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */ | |
e7f70af1 | 3065 | nla_total_size(sizeof(__be32)) + /* IFLA_VXLAN_LABEL */ |
d342894c | 3066 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */ |
e4f67add DS |
3067 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */ |
3068 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */ | |
3069 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */ | |
3070 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */ | |
da8b43c0 | 3071 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_COLLECT_METADATA */ |
d342894c | 3072 | nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */ |
3073 | nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */ | |
05f47d69 | 3074 | nla_total_size(sizeof(struct ifla_vxlan_port_range)) + |
359a0ea9 TH |
3075 | nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */ |
3076 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */ | |
3077 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */ | |
3078 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */ | |
dfd8645e TH |
3079 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */ |
3080 | nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */ | |
d342894c | 3081 | 0; |
3082 | } | |
3083 | ||
3084 | static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev) | |
3085 | { | |
3086 | const struct vxlan_dev *vxlan = netdev_priv(dev); | |
c7995c43 | 3087 | const struct vxlan_rdst *dst = &vxlan->default_dst; |
05f47d69 | 3088 | struct ifla_vxlan_port_range ports = { |
0dfbdf41 TG |
3089 | .low = htons(vxlan->cfg.port_min), |
3090 | .high = htons(vxlan->cfg.port_max), | |
05f47d69 | 3091 | }; |
d342894c | 3092 | |
54bfd872 | 3093 | if (nla_put_u32(skb, IFLA_VXLAN_ID, be32_to_cpu(dst->remote_vni))) |
d342894c | 3094 | goto nla_put_failure; |
3095 | ||
e4c7ed41 CW |
3096 | if (!vxlan_addr_any(&dst->remote_ip)) { |
3097 | if (dst->remote_ip.sa.sa_family == AF_INET) { | |
930345ea JB |
3098 | if (nla_put_in_addr(skb, IFLA_VXLAN_GROUP, |
3099 | dst->remote_ip.sin.sin_addr.s_addr)) | |
e4c7ed41 CW |
3100 | goto nla_put_failure; |
3101 | #if IS_ENABLED(CONFIG_IPV6) | |
3102 | } else { | |
930345ea JB |
3103 | if (nla_put_in6_addr(skb, IFLA_VXLAN_GROUP6, |
3104 | &dst->remote_ip.sin6.sin6_addr)) | |
e4c7ed41 CW |
3105 | goto nla_put_failure; |
3106 | #endif | |
3107 | } | |
3108 | } | |
d342894c | 3109 | |
c7995c43 | 3110 | if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex)) |
d342894c | 3111 | goto nla_put_failure; |
3112 | ||
0dfbdf41 TG |
3113 | if (!vxlan_addr_any(&vxlan->cfg.saddr)) { |
3114 | if (vxlan->cfg.saddr.sa.sa_family == AF_INET) { | |
930345ea | 3115 | if (nla_put_in_addr(skb, IFLA_VXLAN_LOCAL, |
0dfbdf41 | 3116 | vxlan->cfg.saddr.sin.sin_addr.s_addr)) |
e4c7ed41 CW |
3117 | goto nla_put_failure; |
3118 | #if IS_ENABLED(CONFIG_IPV6) | |
3119 | } else { | |
930345ea | 3120 | if (nla_put_in6_addr(skb, IFLA_VXLAN_LOCAL6, |
0dfbdf41 | 3121 | &vxlan->cfg.saddr.sin6.sin6_addr)) |
e4c7ed41 CW |
3122 | goto nla_put_failure; |
3123 | #endif | |
3124 | } | |
3125 | } | |
d342894c | 3126 | |
0dfbdf41 TG |
3127 | if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->cfg.ttl) || |
3128 | nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->cfg.tos) || | |
e7f70af1 | 3129 | nla_put_be32(skb, IFLA_VXLAN_LABEL, vxlan->cfg.label) || |
e4f67add DS |
3130 | nla_put_u8(skb, IFLA_VXLAN_LEARNING, |
3131 | !!(vxlan->flags & VXLAN_F_LEARN)) || | |
3132 | nla_put_u8(skb, IFLA_VXLAN_PROXY, | |
3133 | !!(vxlan->flags & VXLAN_F_PROXY)) || | |
3134 | nla_put_u8(skb, IFLA_VXLAN_RSC, !!(vxlan->flags & VXLAN_F_RSC)) || | |
3135 | nla_put_u8(skb, IFLA_VXLAN_L2MISS, | |
3136 | !!(vxlan->flags & VXLAN_F_L2MISS)) || | |
3137 | nla_put_u8(skb, IFLA_VXLAN_L3MISS, | |
3138 | !!(vxlan->flags & VXLAN_F_L3MISS)) || | |
da8b43c0 AS |
3139 | nla_put_u8(skb, IFLA_VXLAN_COLLECT_METADATA, |
3140 | !!(vxlan->flags & VXLAN_F_COLLECT_METADATA)) || | |
0dfbdf41 TG |
3141 | nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->cfg.age_interval) || |
3142 | nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->cfg.addrmax) || | |
3143 | nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->cfg.dst_port) || | |
359a0ea9 | 3144 | nla_put_u8(skb, IFLA_VXLAN_UDP_CSUM, |
6ceb31ca | 3145 | !(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM_TX)) || |
359a0ea9 TH |
3146 | nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, |
3147 | !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_TX)) || | |
3148 | nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, | |
dfd8645e TH |
3149 | !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_RX)) || |
3150 | nla_put_u8(skb, IFLA_VXLAN_REMCSUM_TX, | |
3151 | !!(vxlan->flags & VXLAN_F_REMCSUM_TX)) || | |
3152 | nla_put_u8(skb, IFLA_VXLAN_REMCSUM_RX, | |
3153 | !!(vxlan->flags & VXLAN_F_REMCSUM_RX))) | |
d342894c | 3154 | goto nla_put_failure; |
3155 | ||
05f47d69 | 3156 | if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports)) |
3157 | goto nla_put_failure; | |
3158 | ||
3511494c TG |
3159 | if (vxlan->flags & VXLAN_F_GBP && |
3160 | nla_put_flag(skb, IFLA_VXLAN_GBP)) | |
3161 | goto nla_put_failure; | |
3162 | ||
e1e5314d JB |
3163 | if (vxlan->flags & VXLAN_F_GPE && |
3164 | nla_put_flag(skb, IFLA_VXLAN_GPE)) | |
3165 | goto nla_put_failure; | |
3166 | ||
0ace2ca8 TH |
3167 | if (vxlan->flags & VXLAN_F_REMCSUM_NOPARTIAL && |
3168 | nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL)) | |
3169 | goto nla_put_failure; | |
3170 | ||
d342894c | 3171 | return 0; |
3172 | ||
3173 | nla_put_failure: | |
3174 | return -EMSGSIZE; | |
3175 | } | |
3176 | ||
1728d4fa ND |
3177 | static struct net *vxlan_get_link_net(const struct net_device *dev) |
3178 | { | |
3179 | struct vxlan_dev *vxlan = netdev_priv(dev); | |
3180 | ||
3181 | return vxlan->net; | |
3182 | } | |
3183 | ||
d342894c | 3184 | static struct rtnl_link_ops vxlan_link_ops __read_mostly = { |
3185 | .kind = "vxlan", | |
3186 | .maxtype = IFLA_VXLAN_MAX, | |
3187 | .policy = vxlan_policy, | |
3188 | .priv_size = sizeof(struct vxlan_dev), | |
3189 | .setup = vxlan_setup, | |
3190 | .validate = vxlan_validate, | |
3191 | .newlink = vxlan_newlink, | |
3192 | .dellink = vxlan_dellink, | |
3193 | .get_size = vxlan_get_size, | |
3194 | .fill_info = vxlan_fill_info, | |
1728d4fa | 3195 | .get_link_net = vxlan_get_link_net, |
d342894c | 3196 | }; |
3197 | ||
cf5da330 ND |
3198 | struct net_device *vxlan_dev_create(struct net *net, const char *name, |
3199 | u8 name_assign_type, | |
3200 | struct vxlan_config *conf) | |
3201 | { | |
3202 | struct nlattr *tb[IFLA_MAX + 1]; | |
3203 | struct net_device *dev; | |
3204 | int err; | |
3205 | ||
3206 | memset(&tb, 0, sizeof(tb)); | |
3207 | ||
3208 | dev = rtnl_create_link(net, name, name_assign_type, | |
3209 | &vxlan_link_ops, tb); | |
3210 | if (IS_ERR(dev)) | |
3211 | return dev; | |
3212 | ||
3213 | err = vxlan_dev_configure(net, dev, conf); | |
3214 | if (err < 0) { | |
3215 | free_netdev(dev); | |
3216 | return ERR_PTR(err); | |
3217 | } | |
3218 | ||
3219 | err = rtnl_configure_link(dev, NULL); | |
3220 | if (err < 0) { | |
3221 | LIST_HEAD(list_kill); | |
3222 | ||
3223 | vxlan_dellink(dev, &list_kill); | |
3224 | unregister_netdevice_many(&list_kill); | |
3225 | return ERR_PTR(err); | |
3226 | } | |
3227 | ||
3228 | return dev; | |
3229 | } | |
3230 | EXPORT_SYMBOL_GPL(vxlan_dev_create); | |
3231 | ||
acaf4e70 DB |
3232 | static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn, |
3233 | struct net_device *dev) | |
3234 | { | |
3235 | struct vxlan_dev *vxlan, *next; | |
3236 | LIST_HEAD(list_kill); | |
3237 | ||
3238 | list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) { | |
3239 | struct vxlan_rdst *dst = &vxlan->default_dst; | |
3240 | ||
3241 | /* In case we created vxlan device with carrier | |
3242 | * and we loose the carrier due to module unload | |
3243 | * we also need to remove vxlan device. In other | |
3244 | * cases, it's not necessary and remote_ifindex | |
3245 | * is 0 here, so no matches. | |
3246 | */ | |
3247 | if (dst->remote_ifindex == dev->ifindex) | |
3248 | vxlan_dellink(vxlan->dev, &list_kill); | |
3249 | } | |
3250 | ||
3251 | unregister_netdevice_many(&list_kill); | |
3252 | } | |
3253 | ||
b7aade15 HFS |
3254 | static int vxlan_netdevice_event(struct notifier_block *unused, |
3255 | unsigned long event, void *ptr) | |
acaf4e70 DB |
3256 | { |
3257 | struct net_device *dev = netdev_notifier_info_to_dev(ptr); | |
783c1463 | 3258 | struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id); |
acaf4e70 | 3259 | |
783c1463 | 3260 | if (event == NETDEV_UNREGISTER) |
acaf4e70 | 3261 | vxlan_handle_lowerdev_unregister(vn, dev); |
7c46a640 | 3262 | else if (event == NETDEV_UDP_TUNNEL_PUSH_INFO) |
b7aade15 | 3263 | vxlan_push_rx_ports(dev); |
acaf4e70 DB |
3264 | |
3265 | return NOTIFY_DONE; | |
3266 | } | |
3267 | ||
3268 | static struct notifier_block vxlan_notifier_block __read_mostly = { | |
b7aade15 | 3269 | .notifier_call = vxlan_netdevice_event, |
acaf4e70 DB |
3270 | }; |
3271 | ||
d342894c | 3272 | static __net_init int vxlan_init_net(struct net *net) |
3273 | { | |
3274 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
31fec5aa | 3275 | unsigned int h; |
d342894c | 3276 | |
553675fb | 3277 | INIT_LIST_HEAD(&vn->vxlan_list); |
1c51a915 | 3278 | spin_lock_init(&vn->sock_lock); |
d342894c | 3279 | |
553675fb | 3280 | for (h = 0; h < PORT_HASH_SIZE; ++h) |
3281 | INIT_HLIST_HEAD(&vn->sock_list[h]); | |
d342894c | 3282 | |
3283 | return 0; | |
3284 | } | |
3285 | ||
f01ec1c0 ND |
3286 | static void __net_exit vxlan_exit_net(struct net *net) |
3287 | { | |
3288 | struct vxlan_net *vn = net_generic(net, vxlan_net_id); | |
3289 | struct vxlan_dev *vxlan, *next; | |
3290 | struct net_device *dev, *aux; | |
3291 | LIST_HEAD(list); | |
3292 | ||
3293 | rtnl_lock(); | |
3294 | for_each_netdev_safe(net, dev, aux) | |
3295 | if (dev->rtnl_link_ops == &vxlan_link_ops) | |
3296 | unregister_netdevice_queue(dev, &list); | |
3297 | ||
3298 | list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) { | |
3299 | /* If vxlan->dev is in the same netns, it has already been added | |
3300 | * to the list by the previous loop. | |
3301 | */ | |
58ce31cc TH |
3302 | if (!net_eq(dev_net(vxlan->dev), net)) { |
3303 | gro_cells_destroy(&vxlan->gro_cells); | |
13c3ed6a | 3304 | unregister_netdevice_queue(vxlan->dev, &list); |
58ce31cc | 3305 | } |
f01ec1c0 ND |
3306 | } |
3307 | ||
3308 | unregister_netdevice_many(&list); | |
3309 | rtnl_unlock(); | |
3310 | } | |
3311 | ||
d342894c | 3312 | static struct pernet_operations vxlan_net_ops = { |
3313 | .init = vxlan_init_net, | |
f01ec1c0 | 3314 | .exit = vxlan_exit_net, |
d342894c | 3315 | .id = &vxlan_net_id, |
3316 | .size = sizeof(struct vxlan_net), | |
3317 | }; | |
3318 | ||
3319 | static int __init vxlan_init_module(void) | |
3320 | { | |
3321 | int rc; | |
3322 | ||
3323 | get_random_bytes(&vxlan_salt, sizeof(vxlan_salt)); | |
3324 | ||
783c1463 | 3325 | rc = register_pernet_subsys(&vxlan_net_ops); |
d342894c | 3326 | if (rc) |
3327 | goto out1; | |
3328 | ||
acaf4e70 | 3329 | rc = register_netdevice_notifier(&vxlan_notifier_block); |
d342894c | 3330 | if (rc) |
3331 | goto out2; | |
3332 | ||
acaf4e70 DB |
3333 | rc = rtnl_link_register(&vxlan_link_ops); |
3334 | if (rc) | |
3335 | goto out3; | |
d342894c | 3336 | |
acaf4e70 DB |
3337 | return 0; |
3338 | out3: | |
3339 | unregister_netdevice_notifier(&vxlan_notifier_block); | |
d342894c | 3340 | out2: |
783c1463 | 3341 | unregister_pernet_subsys(&vxlan_net_ops); |
d342894c | 3342 | out1: |
3343 | return rc; | |
3344 | } | |
7332a13b | 3345 | late_initcall(vxlan_init_module); |
d342894c | 3346 | |
3347 | static void __exit vxlan_cleanup_module(void) | |
3348 | { | |
b7153984 | 3349 | rtnl_link_unregister(&vxlan_link_ops); |
acaf4e70 | 3350 | unregister_netdevice_notifier(&vxlan_notifier_block); |
783c1463 DB |
3351 | unregister_pernet_subsys(&vxlan_net_ops); |
3352 | /* rcu_barrier() is called by netns */ | |
d342894c | 3353 | } |
3354 | module_exit(vxlan_cleanup_module); | |
3355 | ||
3356 | MODULE_LICENSE("GPL"); | |
3357 | MODULE_VERSION(VXLAN_VERSION); | |
3b8df3c6 | 3358 | MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>"); |
ead5139a | 3359 | MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic"); |
d342894c | 3360 | MODULE_ALIAS_RTNL_LINK("vxlan"); |