projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge tag 'vfs-6.7.misc' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs
[linux-block.git] / drivers / net / vxlan / vxlan_core.c
CommitLineData
d2912cb1 1// SPDX-License-Identifier: GPL-2.0-only
d342894c 2/*
eb5ce439 3 * VXLAN: Virtual eXtensible Local Area Network
d342894c 4 *
3b8df3c6 5 * Copyright (c) 2012-2013 Vyatta Inc.
d342894c 6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/kernel.h>
d342894c 11#include <linux/module.h>
12#include <linux/errno.h>
13#include <linux/slab.h>
d342894c 14#include <linux/udp.h>
15#include <linux/igmp.h>
d342894c 16#include <linux/if_ether.h>
1b13c97f 17#include <linux/ethtool.h>
e4f67add
DS		 18#include <net/arp.h>
19#include <net/ndisc.h>
4721031c 20#include <net/gro.h>
3616d08b 21#include <net/ipv6_stubs.h>
d342894c 22#include <net/ip.h>
23#include <net/icmp.h>
d342894c 24#include <net/rtnetlink.h>
d342894c 25#include <net/inet_ecn.h>
26#include <net/net_namespace.h>
27#include <net/netns/generic.h>
fa20e0e3 28#include <net/tun_proto.h>
012a5729 29#include <net/vxlan.h>
1274e1cc 30#include <net/nexthop.h>
40d29af0 31
e4c7ed41 32#if IS_ENABLED(CONFIG_IPV6)
e4c7ed41 33#include <net/ip6_tunnel.h>
660d98ca 34#include <net/ip6_checksum.h>
e4c7ed41 35#endif
d342894c 36
76fc217d
RP		 37#include "vxlan_private.h"
38
d342894c 39#define VXLAN_VERSION "0.1"
40
d342894c 41#define FDB_AGE_DEFAULT 300 /* 5 min */
42#define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */
43
23c578bf 44/* UDP port for VXLAN traffic.
45 * The IANA assigned port is 4789, but the Linux default is 8472
234f5b73 46 * for compatibility with early adopters.
23c578bf 47 */
9daaa397
SH		 48static unsigned short vxlan_port __read_mostly = 8472;
49module_param_named(udp_port, vxlan_port, ushort, 0444);
d342894c 50MODULE_PARM_DESC(udp_port, "Destination UDP port");
51
52static bool log_ecn_error = true;
53module_param(log_ecn_error, bool, 0644);
54MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
55
76fc217d 56unsigned int vxlan_net_id;
553675fb 57
76fc217d
RP		 58const u8 all_zeros_mac[ETH_ALEN + 2];
59static struct rtnl_link_ops vxlan_link_ops;
afbd8bae 60
205f356d 61static int vxlan_sock_add(struct vxlan_dev *vxlan);
614732ea 62
a53cb29b
MB		 63static void vxlan_vs_del_dev(struct vxlan_dev *vxlan);
64
d342894c 65/* salt for hash table */
66static u32 vxlan_salt __read_mostly;
67
ee122c79
TG		 68static inline bool vxlan_collect_metadata(struct vxlan_sock *vs)
69{
e7030878
TG		 70 return vs->flags & VXLAN_F_COLLECT_METADATA ||
71 ip_tunnel_collect_metadata();
ee122c79
TG		 72}
73
78ec710e
FF		 74/* Find VXLAN socket based on network namespace, address family, UDP port,
75 * enabled unshareable flags and socket device binding (see l3mdev with
76 * non-default VRF).
ac5132d1
TG		 77 */
78static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family,
aab8cc36 79 __be16 port, u32 flags, int ifindex)
553675fb 80{
81 struct vxlan_sock *vs;
af33c1ad
TH		 82
83 flags &= VXLAN_F_RCV_FLAGS;
553675fb 84
85 hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) {
19ca9fc1 86 if (inet_sk(vs->sock->sk)->inet_sport == port &&
705cc62f 87 vxlan_get_sk_family(vs) == family &&
aab8cc36
AB		 88 vs->flags == flags &&
89 vs->sock->sk->sk_bound_dev_if == ifindex)
553675fb 90 return vs;
91 }
92 return NULL;
d342894c 93}
94
4095e0e1
NA		 95static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs,
96 int ifindex, __be32 vni,
97 struct vxlan_vni_node **vninode)
d342894c 98{
4095e0e1 99 struct vxlan_vni_node *vnode;
69e76661 100 struct vxlan_dev_node *node;
d342894c 101
b9167b2e 102 /* For flow based devices, map all packets to VNI 0 */
f9c4bb0b
RP		 103 if (vs->flags & VXLAN_F_COLLECT_METADATA &&
104 !(vs->flags & VXLAN_F_VNIFILTER))
b9167b2e
JB		 105 vni = 0;
106
69e76661 107 hlist_for_each_entry_rcu(node, vni_head(vs, vni), hlist) {
f9c4bb0b 108 if (!node->vxlan)
49f810f0 109 continue;
4095e0e1 110 vnode = NULL;
f9c4bb0b 111 if (node->vxlan->cfg.flags & VXLAN_F_VNIFILTER) {
4095e0e1
NA		 112 vnode = vxlan_vnifilter_lookup(node->vxlan, vni);
113 if (!vnode)
f9c4bb0b
RP		 114 continue;
115 } else if (node->vxlan->default_dst.remote_vni != vni) {
116 continue;
117 }
49f810f0
MS		 118
119 if (IS_ENABLED(CONFIG_IPV6)) {
69e76661 120 const struct vxlan_config *cfg = &node->vxlan->cfg;
49f810f0
MS		 121
122 if ((cfg->flags & VXLAN_F_IPV6_LINKLOCAL) &&
123 cfg->remote_ifindex != ifindex)
124 continue;
125 }
126
4095e0e1
NA		 127 if (vninode)
128 *vninode = vnode;
69e76661 129 return node->vxlan;
d342894c 130 }
131
132 return NULL;
133}
134
5cfccc5a 135/* Look up VNI in a per net namespace table */
49f810f0
MS		 136static struct vxlan_dev *vxlan_find_vni(struct net *net, int ifindex,
137 __be32 vni, sa_family_t family,
138 __be16 port, u32 flags)
5cfccc5a
PS		 139{
140 struct vxlan_sock *vs;
141
aab8cc36 142 vs = vxlan_find_sock(net, family, port, flags, ifindex);
5cfccc5a
PS		 143 if (!vs)
144 return NULL;
145
4095e0e1 146 return vxlan_vs_find_vni(vs, ifindex, vni, NULL);
5cfccc5a
PS		 147}
148
d342894c 149/* Fill in neighbour message in skbuff. */
150static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan,
234f5b73
SH		 151 const struct vxlan_fdb *fdb,
152 u32 portid, u32 seq, int type, unsigned int flags,
153 const struct vxlan_rdst *rdst)
d342894c 154{
155 unsigned long now = jiffies;
156 struct nda_cacheinfo ci;
1274e1cc 157 bool send_ip, send_eth;
d342894c 158 struct nlmsghdr *nlh;
1274e1cc 159 struct nexthop *nh;
d342894c 160 struct ndmsg *ndm;
06ec313e
IS		 161 int nh_family;
162 u32 nh_id;
d342894c 163
164 nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags);
165 if (nlh == NULL)
166 return -EMSGSIZE;
167
168 ndm = nlmsg_data(nlh);
169 memset(ndm, 0, sizeof(*ndm));
e4f67add
DS		 170
171 send_eth = send_ip = true;
172
06ec313e
IS		 173 rcu_read_lock();
174 nh = rcu_dereference(fdb->nh);
175 if (nh) {
176 nh_family = nexthop_get_family(nh);
177 nh_id = nh->id;
178 }
179 rcu_read_unlock();
180
e4f67add 181 if (type == RTM_GETNEIGH) {
1274e1cc
RP		 182 if (rdst) {
183 send_ip = !vxlan_addr_any(&rdst->remote_ip);
184 ndm->ndm_family = send_ip ? rdst->remote_ip.sa.sa_family : AF_INET;
185 } else if (nh) {
06ec313e 186 ndm->ndm_family = nh_family;
1274e1cc 187 }
e4f67add
DS		 188 send_eth = !is_zero_ether_addr(fdb->eth_addr);
189 } else
190 ndm->ndm_family = AF_BRIDGE;
d342894c 191 ndm->ndm_state = fdb->state;
192 ndm->ndm_ifindex = vxlan->dev->ifindex;
ae884082 193 ndm->ndm_flags = fdb->flags;
1274e1cc 194 if (rdst && rdst->offloaded)
0efe1173 195 ndm->ndm_flags |= NTF_OFFLOADED;
545469f7 196 ndm->ndm_type = RTN_UNICAST;
d342894c 197
193523bf 198 if (!net_eq(dev_net(vxlan->dev), vxlan->net) &&
4967082b 199 nla_put_s32(skb, NDA_LINK_NETNSID,
38f507f1 200 peernet2id(dev_net(vxlan->dev), vxlan->net)))
193523bf
ND		 201 goto nla_put_failure;
202
e4f67add 203 if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr))
d342894c 204 goto nla_put_failure;
1274e1cc 205 if (nh) {
06ec313e 206 if (nla_put_u32(skb, NDA_NH_ID, nh_id))
1274e1cc
RP		 207 goto nla_put_failure;
208 } else if (rdst) {
209 if (send_ip && vxlan_nla_put_addr(skb, NDA_DST,
210 &rdst->remote_ip))
211 goto nla_put_failure;
212
213 if (rdst->remote_port &&
214 rdst->remote_port != vxlan->cfg.dst_port &&
215 nla_put_be16(skb, NDA_PORT, rdst->remote_port))
216 goto nla_put_failure;
217 if (rdst->remote_vni != vxlan->default_dst.remote_vni &&
218 nla_put_u32(skb, NDA_VNI, be32_to_cpu(rdst->remote_vni)))
219 goto nla_put_failure;
220 if (rdst->remote_ifindex &&
221 nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex))
222 goto nla_put_failure;
223 }
d342894c 224
dc5321d7 225 if ((vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) && fdb->vni &&
3ad7a4b1
RP		 226 nla_put_u32(skb, NDA_SRC_VNI,
227 be32_to_cpu(fdb->vni)))
228 goto nla_put_failure;
d342894c 229
230 ci.ndm_used = jiffies_to_clock_t(now - fdb->used);
231 ci.ndm_confirmed = 0;
232 ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated);
233 ci.ndm_refcnt = 0;
234
235 if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci))
236 goto nla_put_failure;
237
053c095a
JB		 238 nlmsg_end(skb, nlh);
239 return 0;
d342894c 240
241nla_put_failure:
242 nlmsg_cancel(skb, nlh);
243 return -EMSGSIZE;
244}
245
246static inline size_t vxlan_nlmsg_size(void)
247{
248 return NLMSG_ALIGN(sizeof(struct ndmsg))
249 + nla_total_size(ETH_ALEN) /* NDA_LLADDR */
e4c7ed41 250 + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */
73cf3317 251 + nla_total_size(sizeof(__be16)) /* NDA_PORT */
6681712d
DS		 252 + nla_total_size(sizeof(__be32)) /* NDA_VNI */
253 + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */
4967082b 254 + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */
d342894c 255 + nla_total_size(sizeof(struct nda_cacheinfo));
256}
257
9a997353
PM		 258static void __vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
259 struct vxlan_rdst *rd, int type)
d342894c 260{
261 struct net *net = dev_net(vxlan->dev);
262 struct sk_buff *skb;
263 int err = -ENOBUFS;
264
265 skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC);
266 if (skb == NULL)
267 goto errout;
268
9e4b93f9 269 err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd);
d342894c 270 if (err < 0) {
271 /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */
272 WARN_ON(err == -EMSGSIZE);
273 kfree_skb(skb);
274 goto errout;
275 }
276
277 rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC);
278 return;
279errout:
280 if (err < 0)
281 rtnl_set_sk_err(net, RTNLGRP_NEIGH, err);
282}
283
ff23b91c
PM		 284static void vxlan_fdb_switchdev_notifier_info(const struct vxlan_dev *vxlan,
285 const struct vxlan_fdb *fdb,
286 const struct vxlan_rdst *rd,
4c59b7d1 287 struct netlink_ext_ack *extack,
ff23b91c
PM		 288 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
289{
290 fdb_info->info.dev = vxlan->dev;
4c59b7d1 291 fdb_info->info.extack = extack;
ff23b91c
PM		 292 fdb_info->remote_ip = rd->remote_ip;
293 fdb_info->remote_port = rd->remote_port;
294 fdb_info->remote_vni = rd->remote_vni;
295 fdb_info->remote_ifindex = rd->remote_ifindex;
296 memcpy(fdb_info->eth_addr, fdb->eth_addr, ETH_ALEN);
297 fdb_info->vni = fdb->vni;
298 fdb_info->offloaded = rd->offloaded;
299 fdb_info->added_by_user = fdb->flags & NTF_VXLAN_ADDED_BY_USER;
300}
301
61f46fe8
PM		 302static int vxlan_fdb_switchdev_call_notifiers(struct vxlan_dev *vxlan,
303 struct vxlan_fdb *fdb,
304 struct vxlan_rdst *rd,
4c59b7d1
PM		 305 bool adding,
306 struct netlink_ext_ack *extack)
9a997353
PM		 307{
308 struct switchdev_notifier_vxlan_fdb_info info;
309 enum switchdev_notifier_type notifier_type;
61f46fe8 310 int ret;
9a997353
PM		 311
312 if (WARN_ON(!rd))
61f46fe8 313 return 0;
9a997353
PM		 314
315 notifier_type = adding ? SWITCHDEV_VXLAN_FDB_ADD_TO_DEVICE
316 : SWITCHDEV_VXLAN_FDB_DEL_TO_DEVICE;
4c59b7d1 317 vxlan_fdb_switchdev_notifier_info(vxlan, fdb, rd, NULL, &info);
61f46fe8 318 ret = call_switchdev_notifiers(notifier_type, vxlan->dev,
6685987c 319 &info.info, extack);
61f46fe8 320 return notifier_to_errno(ret);
9a997353
PM		 321}
322
61f46fe8 323static int vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
4c59b7d1
PM		 324 struct vxlan_rdst *rd, int type, bool swdev_notify,
325 struct netlink_ext_ack *extack)
9a997353 326{
61f46fe8
PM		 327 int err;
328
1274e1cc 329 if (swdev_notify && rd) {
0e6160f3
PM		 330 switch (type) {
331 case RTM_NEWNEIGH:
61f46fe8 332 err = vxlan_fdb_switchdev_call_notifiers(vxlan, fdb, rd,
4c59b7d1 333 true, extack);
61f46fe8
PM		 334 if (err)
335 return err;
0e6160f3
PM		 336 break;
337 case RTM_DELNEIGH:
338 vxlan_fdb_switchdev_call_notifiers(vxlan, fdb, rd,
4c59b7d1 339 false, extack);
0e6160f3
PM		 340 break;
341 }
9a997353
PM		 342 }
343
344 __vxlan_fdb_notify(vxlan, fdb, rd, type);
61f46fe8 345 return 0;
9a997353
PM		 346}
347
e4c7ed41 348static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa)
e4f67add
DS		 349{
350 struct vxlan_dev *vxlan = netdev_priv(dev);
bb3fd687
SH		 351 struct vxlan_fdb f = {
352 .state = NUD_STALE,
353 };
354 struct vxlan_rdst remote = {
e4c7ed41 355 .remote_ip = *ipa, /* goes to NDA_DST */
54bfd872 356 .remote_vni = cpu_to_be32(VXLAN_N_VID),
bb3fd687 357 };
3e61aa8f 358
4c59b7d1 359 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH, true, NULL);
e4f67add
DS		 360}
361
362static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN])
363{
bb3fd687
SH		 364 struct vxlan_fdb f = {
365 .state = NUD_STALE,
366 };
9e4b93f9 367 struct vxlan_rdst remote = { };
e4f67add 368
e4f67add
DS		 369 memcpy(f.eth_addr, eth_addr, ETH_ALEN);
370
4c59b7d1 371 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH, true, NULL);
e4f67add
DS		 372}
373
d342894c 374/* Hash Ethernet address */
375static u32 eth_hash(const unsigned char *addr)
376{
377 u64 value = get_unaligned((u64 *)addr);
378
379 /* only want 6 bytes */
380#ifdef __BIG_ENDIAN
d342894c 381 value >>= 16;
321fb991 382#else
383 value <<= 16;
d342894c 384#endif
385 return hash_64(value, FDB_HASH_BITS);
386}
387
c63053e0 388u32 eth_vni_hash(const unsigned char *addr, __be32 vni)
3ad7a4b1
RP		 389{
390 /* use 1 byte of OUI and 3 bytes of NIC */
391 u32 key = get_unaligned((u32 *)(addr + 2));
392
393 return jhash_2words(key, vni, vxlan_salt) & (FDB_HASH_SIZE - 1);
394}
395
c63053e0 396u32 fdb_head_index(struct vxlan_dev *vxlan, const u8 *mac, __be32 vni)
fe1e0713
L		 397{
398 if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA)
399 return eth_vni_hash(mac, vni);
400 else
401 return eth_hash(mac);
402}
403
d342894c 404/* Hash chain to use given mac address */
405static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan,
3ad7a4b1 406 const u8 *mac, __be32 vni)
d342894c 407{
fe1e0713 408 return &vxlan->fdb_head[fdb_head_index(vxlan, mac, vni)];
d342894c 409}
410
411/* Look up Ethernet address in forwarding table */
014be2c8 412static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan,
3ad7a4b1 413 const u8 *mac, __be32 vni)
d342894c 414{
3ad7a4b1 415 struct hlist_head *head = vxlan_fdb_head(vxlan, mac, vni);
d342894c 416 struct vxlan_fdb *f;
d342894c 417
b67bfe0d 418 hlist_for_each_entry_rcu(f, head, hlist) {
3ad7a4b1 419 if (ether_addr_equal(mac, f->eth_addr)) {
dc5321d7 420 if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) {
3ad7a4b1
RP		 421 if (vni == f->vni)
422 return f;
423 } else {
424 return f;
425 }
426 }
d342894c 427 }
428
429 return NULL;
430}
431
014be2c8 432static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan,
3ad7a4b1 433 const u8 *mac, __be32 vni)
014be2c8
SS		 434{
435 struct vxlan_fdb *f;
436
3ad7a4b1 437 f = __vxlan_find_mac(vxlan, mac, vni);
016f3d18 438 if (f && f->used != jiffies)
014be2c8
SS		 439 f->used = jiffies;
440
441 return f;
442}
443
a5e7c10a
MR		 444/* caller should hold vxlan->hash_lock */
445static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f,
e4c7ed41 446 union vxlan_addr *ip, __be16 port,
54bfd872 447 __be32 vni, __u32 ifindex)
6681712d 448{
3e61aa8f 449 struct vxlan_rdst *rd;
6681712d 450
3e61aa8f 451 list_for_each_entry(rd, &f->remotes, list) {
e4c7ed41 452 if (vxlan_addr_equal(&rd->remote_ip, ip) &&
6681712d
DS		 453 rd->remote_port == port &&
454 rd->remote_vni == vni &&
455 rd->remote_ifindex == ifindex)
a5e7c10a 456 return rd;
6681712d 457 }
3e61aa8f 458
a5e7c10a
MR		 459 return NULL;
460}
461
1941f1d6
PM		 462int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
463 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
464{
465 struct vxlan_dev *vxlan = netdev_priv(dev);
466 u8 eth_addr[ETH_ALEN + 2] = { 0 };
467 struct vxlan_rdst *rdst;
468 struct vxlan_fdb *f;
469 int rc = 0;
470
471 if (is_multicast_ether_addr(mac) ||
472 is_zero_ether_addr(mac))
473 return -EINVAL;
474
475 ether_addr_copy(eth_addr, mac);
476
477 rcu_read_lock();
478
479 f = __vxlan_find_mac(vxlan, eth_addr, vni);
480 if (!f) {
481 rc = -ENOENT;
482 goto out;
483 }
484
485 rdst = first_remote_rcu(f);
4c59b7d1 486 vxlan_fdb_switchdev_notifier_info(vxlan, f, rdst, NULL, fdb_info);
1941f1d6
PM		 487
488out:
489 rcu_read_unlock();
490 return rc;
491}
492EXPORT_SYMBOL_GPL(vxlan_fdb_find_uc);
493
4f89f5b5
PM		 494static int vxlan_fdb_notify_one(struct notifier_block *nb,
495 const struct vxlan_dev *vxlan,
496 const struct vxlan_fdb *f,
4c59b7d1
PM		 497 const struct vxlan_rdst *rdst,
498 struct netlink_ext_ack *extack)
4f89f5b5
PM		 499{
500 struct switchdev_notifier_vxlan_fdb_info fdb_info;
501 int rc;
502
4c59b7d1 503 vxlan_fdb_switchdev_notifier_info(vxlan, f, rdst, extack, &fdb_info);
4f89f5b5
PM		 504 rc = nb->notifier_call(nb, SWITCHDEV_VXLAN_FDB_ADD_TO_DEVICE,
505 &fdb_info);
506 return notifier_to_errno(rc);
507}
508
509int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
4c59b7d1
PM		 510 struct notifier_block *nb,
511 struct netlink_ext_ack *extack)
4f89f5b5
PM		 512{
513 struct vxlan_dev *vxlan;
514 struct vxlan_rdst *rdst;
515 struct vxlan_fdb *f;
516 unsigned int h;
517 int rc = 0;
518
519 if (!netif_is_vxlan(dev))
520 return -EINVAL;
521 vxlan = netdev_priv(dev);
522
4f89f5b5 523 for (h = 0; h < FDB_HASH_SIZE; ++h) {
fe1e0713 524 spin_lock_bh(&vxlan->hash_lock[h]);
4f89f5b5
PM		 525 hlist_for_each_entry(f, &vxlan->fdb_head[h], hlist) {
526 if (f->vni == vni) {
527 list_for_each_entry(rdst, &f->remotes, list) {
528 rc = vxlan_fdb_notify_one(nb, vxlan,
4c59b7d1
PM		 529 f, rdst,
530 extack);
4f89f5b5 531 if (rc)
fe1e0713 532 goto unlock;
4f89f5b5
PM		 533 }
534 }
535 }
fe1e0713 536 spin_unlock_bh(&vxlan->hash_lock[h]);
4f89f5b5 537 }
fe1e0713 538 return 0;
4f89f5b5 539
fe1e0713
L		 540unlock:
541 spin_unlock_bh(&vxlan->hash_lock[h]);
4f89f5b5
PM		 542 return rc;
543}
544EXPORT_SYMBOL_GPL(vxlan_fdb_replay);
545
e5ff4b19
PM		 546void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni)
547{
548 struct vxlan_dev *vxlan;
549 struct vxlan_rdst *rdst;
550 struct vxlan_fdb *f;
551 unsigned int h;
552
553 if (!netif_is_vxlan(dev))
554 return;
555 vxlan = netdev_priv(dev);
556
e5ff4b19 557 for (h = 0; h < FDB_HASH_SIZE; ++h) {
fe1e0713 558 spin_lock_bh(&vxlan->hash_lock[h]);
e5ff4b19
PM		 559 hlist_for_each_entry(f, &vxlan->fdb_head[h], hlist)
560 if (f->vni == vni)
561 list_for_each_entry(rdst, &f->remotes, list)
562 rdst->offloaded = false;
fe1e0713 563 spin_unlock_bh(&vxlan->hash_lock[h]);
e5ff4b19 564 }
fe1e0713 565
e5ff4b19
PM		 566}
567EXPORT_SYMBOL_GPL(vxlan_fdb_clear_offload);
568
906dc186
TR		 569/* Replace destination of unicast mac */
570static int vxlan_fdb_replace(struct vxlan_fdb *f,
54bfd872 571 union vxlan_addr *ip, __be16 port, __be32 vni,
ccdfd4f7 572 __u32 ifindex, struct vxlan_rdst *oldrd)
906dc186
TR		 573{
574 struct vxlan_rdst *rd;
575
576 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
577 if (rd)
578 return 0;
579
580 rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list);
581 if (!rd)
582 return 0;
0c1d70af 583
ccdfd4f7 584 *oldrd = *rd;
0c1d70af 585 dst_cache_reset(&rd->dst_cache);
e4c7ed41 586 rd->remote_ip = *ip;
906dc186
TR		 587 rd->remote_port = port;
588 rd->remote_vni = vni;
589 rd->remote_ifindex = ifindex;
6ad0b5a4 590 rd->offloaded = false;
906dc186
TR		 591 return 1;
592}
593
a5e7c10a
MR		 594/* Add/update destinations for multicast */
595static int vxlan_fdb_append(struct vxlan_fdb *f,
54bfd872 596 union vxlan_addr *ip, __be16 port, __be32 vni,
9e4b93f9 597 __u32 ifindex, struct vxlan_rdst **rdp)
a5e7c10a
MR		 598{
599 struct vxlan_rdst *rd;
600
601 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
602 if (rd)
603 return 0;
604
6681712d
DS		 605 rd = kmalloc(sizeof(*rd), GFP_ATOMIC);
606 if (rd == NULL)
7cea5560 607 return -ENOMEM;
0c1d70af
PA		 608
609 if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) {
610 kfree(rd);
7cea5560 611 return -ENOMEM;
0c1d70af
PA		 612 }
613
e4c7ed41 614 rd->remote_ip = *ip;
6681712d 615 rd->remote_port = port;
0efe1173 616 rd->offloaded = false;
6681712d
DS		 617 rd->remote_vni = vni;
618 rd->remote_ifindex = ifindex;
3e61aa8f
SH		 619
620 list_add_tail_rcu(&rd->list, &f->remotes);
621
9e4b93f9 622 *rdp = rd;
6681712d
DS		 623 return 1;
624}
625
17a0a644
JB		 626static bool vxlan_parse_gpe_proto(struct vxlanhdr *hdr, __be16 *protocol)
627{
628 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)hdr;
629
630 /* Need to have Next Protocol set for interfaces in GPE mode. */
631 if (!gpe->np_applied)
632 return false;
633 /* "The initial version is 0. If a receiver does not support the
634 * version indicated it MUST drop the packet.
635 */
636 if (gpe->version != 0)
637 return false;
638 /* "When the O bit is set to 1, the packet is an OAM packet and OAM
639 * processing MUST occur." However, we don't implement OAM
640 * processing, thus drop the packet.
641 */
642 if (gpe->oam_flag)
643 return false;
644
645 *protocol = tun_p_to_eth_p(gpe->next_protocol);
646 if (!*protocol)
647 return false;
648
649 return true;
650}
651
dfd8645e
TH		 652static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb,
653 unsigned int off,
654 struct vxlanhdr *vh, size_t hdrlen,
54bfd872
JB		 655 __be32 vni_field,
656 struct gro_remcsum *grc,
0ace2ca8 657 bool nopartial)
dfd8645e 658{
b7fe10e5 659 size_t start, offset;
dfd8645e
TH		 660
661 if (skb->remcsum_offload)
b7fe10e5 662 return vh;
dfd8645e
TH		 663
664 if (!NAPI_GRO_CB(skb)->csum_valid)
665 return NULL;
666
54bfd872
JB		 667 start = vxlan_rco_start(vni_field);
668 offset = start + vxlan_rco_offset(vni_field);
dfd8645e 669
b7fe10e5
TH		 670 vh = skb_gro_remcsum_process(skb, (void *)vh, off, hdrlen,
671 start, offset, grc, nopartial);
dfd8645e
TH		 672
673 skb->remcsum_offload = 1;
674
675 return vh;
676}
677
b0b672c4
JB		 678static struct vxlanhdr *vxlan_gro_prepare_receive(struct sock *sk,
679 struct list_head *head,
680 struct sk_buff *skb,
681 struct gro_remcsum *grc)
dc01e7d3 682{
d4546c25 683 struct sk_buff *p;
dc01e7d3 684 struct vxlanhdr *vh, *vh2;
9b174d88 685 unsigned int hlen, off_vx;
5602c48c 686 struct vxlan_sock *vs = rcu_dereference_sk_user_data(sk);
54bfd872 687 __be32 flags;
26c4f7da 688
b0b672c4 689 skb_gro_remcsum_init(grc);
dc01e7d3
OG		 690
691 off_vx = skb_gro_offset(skb);
692 hlen = off_vx + sizeof(*vh);
35ffb665
RG		 693 vh = skb_gro_header(skb, hlen, off_vx);
694 if (unlikely(!vh))
b0b672c4 695 return NULL;
dc01e7d3 696
dfd8645e
TH		 697 skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr));
698
54bfd872 699 flags = vh->vx_flags;
dfd8645e
TH		 700
701 if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) {
702 vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr),
b0b672c4 703 vh->vx_vni, grc,
0ace2ca8
TH		 704 !!(vs->flags &
705 VXLAN_F_REMCSUM_NOPARTIAL));
dfd8645e
TH		 706
707 if (!vh)
b0b672c4 708 return NULL;
dfd8645e
TH		 709 }
710
b7fe10e5
TH		 711 skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */
712
d4546c25 713 list_for_each_entry(p, head, list) {
dc01e7d3
OG		 714 if (!NAPI_GRO_CB(p)->same_flow)
715 continue;
716
717 vh2 = (struct vxlanhdr *)(p->data + off_vx);
3511494c
TG		 718 if (vh->vx_flags != vh2->vx_flags ||
719 vh->vx_vni != vh2->vx_vni) {
dc01e7d3
OG		 720 NAPI_GRO_CB(p)->same_flow = 0;
721 continue;
722 }
dc01e7d3
OG		 723 }
724
b0b672c4
JB		 725 return vh;
726}
dc01e7d3 727
b0b672c4
JB		 728static struct sk_buff *vxlan_gro_receive(struct sock *sk,
729 struct list_head *head,
730 struct sk_buff *skb)
731{
732 struct sk_buff *pp = NULL;
733 struct gro_remcsum grc;
734 int flush = 1;
dc01e7d3 735
b0b672c4
JB		 736 if (vxlan_gro_prepare_receive(sk, head, skb, &grc)) {
737 pp = call_gro_receive(eth_gro_receive, head, skb);
738 flush = 0;
739 }
603d4cf8 740 skb_gro_flush_final_remcsum(skb, pp, flush, &grc);
b0b672c4
JB		 741 return pp;
742}
dc01e7d3 743
b0b672c4
JB		 744static struct sk_buff *vxlan_gpe_gro_receive(struct sock *sk,
745 struct list_head *head,
746 struct sk_buff *skb)
747{
748 const struct packet_offload *ptype;
749 struct sk_buff *pp = NULL;
750 struct gro_remcsum grc;
751 struct vxlanhdr *vh;
752 __be16 protocol;
753 int flush = 1;
dc01e7d3 754
b0b672c4
JB		 755 vh = vxlan_gro_prepare_receive(sk, head, skb, &grc);
756 if (vh) {
757 if (!vxlan_parse_gpe_proto(vh, &protocol))
758 goto out;
759 ptype = gro_find_receive_by_type(protocol);
760 if (!ptype)
761 goto out;
762 pp = call_gro_receive(ptype->callbacks.gro_receive, head, skb);
763 flush = 0;
764 }
765out:
766 skb_gro_flush_final_remcsum(skb, pp, flush, &grc);
dc01e7d3
OG		 767 return pp;
768}
769
5602c48c 770static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
dc01e7d3 771{
229740c6
JR		 772 /* Sets 'skb->inner_mac_header' since we are always called with
773 * 'skb->encapsulation' set.
774 */
9b174d88 775 return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr));
dc01e7d3
OG		 776}
777
b0b672c4
JB		 778static int vxlan_gpe_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
779{
780 struct vxlanhdr *vh = (struct vxlanhdr *)(skb->data + nhoff);
781 const struct packet_offload *ptype;
782 int err = -ENOSYS;
783 __be16 protocol;
784
785 if (!vxlan_parse_gpe_proto(vh, &protocol))
786 return err;
787 ptype = gro_find_complete_by_type(protocol);
788 if (ptype)
789 err = ptype->callbacks.gro_complete(skb, nhoff + sizeof(struct vxlanhdr));
790 return err;
791}
792
c7cdbe2e
RP		 793static struct vxlan_fdb *vxlan_fdb_alloc(struct vxlan_dev *vxlan, const u8 *mac,
794 __u16 state, __be32 src_vni,
795 __u16 ndm_flags)
25e20e73
RP		 796{
797 struct vxlan_fdb *f;
798
799 f = kmalloc(sizeof(*f), GFP_ATOMIC);
800 if (!f)
801 return NULL;
802 f->state = state;
803 f->flags = ndm_flags;
804 f->updated = f->used = jiffies;
805 f->vni = src_vni;
1274e1cc 806 f->nh = NULL;
79472fe8 807 RCU_INIT_POINTER(f->vdev, vxlan);
1274e1cc 808 INIT_LIST_HEAD(&f->nh_list);
25e20e73
RP		 809 INIT_LIST_HEAD(&f->remotes);
810 memcpy(f->eth_addr, mac, ETH_ALEN);
811
812 return f;
813}
814
7c31e54a
TY		 815static void vxlan_fdb_insert(struct vxlan_dev *vxlan, const u8 *mac,
816 __be32 src_vni, struct vxlan_fdb *f)
817{
818 ++vxlan->addrcnt;
819 hlist_add_head_rcu(&f->hlist,
820 vxlan_fdb_head(vxlan, mac, src_vni));
821}
822
1274e1cc
RP		 823static int vxlan_fdb_nh_update(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
824 u32 nhid, struct netlink_ext_ack *extack)
825{
826 struct nexthop *old_nh = rtnl_dereference(fdb->nh);
1274e1cc
RP		 827 struct nexthop *nh;
828 int err = -EINVAL;
829
830 if (old_nh && old_nh->id == nhid)
831 return 0;
832
833 nh = nexthop_find_by_id(vxlan->net, nhid);
834 if (!nh) {
835 NL_SET_ERR_MSG(extack, "Nexthop id does not exist");
836 goto err_inval;
837 }
838
8daf4e75
DC		 839 if (!nexthop_get(nh)) {
840 NL_SET_ERR_MSG(extack, "Nexthop has been deleted");
841 nh = NULL;
842 goto err_inval;
843 }
844 if (!nexthop_is_fdb(nh)) {
845 NL_SET_ERR_MSG(extack, "Nexthop is not a fdb nexthop");
846 goto err_inval;
847 }
1274e1cc 848
8daf4e75
DC		 849 if (!nexthop_is_multipath(nh)) {
850 NL_SET_ERR_MSG(extack, "Nexthop is not a multipath group");
851 goto err_inval;
852 }
853
854 /* check nexthop group family */
855 switch (vxlan->default_dst.remote_ip.sa.sa_family) {
856 case AF_INET:
857 if (!nexthop_has_v4(nh)) {
858 err = -EAFNOSUPPORT;
859 NL_SET_ERR_MSG(extack, "Nexthop group family not supported");
1274e1cc
RP		 860 goto err_inval;
861 }
8daf4e75
DC		 862 break;
863 case AF_INET6:
864 if (nexthop_has_v4(nh)) {
865 err = -EAFNOSUPPORT;
866 NL_SET_ERR_MSG(extack, "Nexthop group family not supported");
867 goto err_inval;
1274e1cc
RP		 868 }
869 }
870
871 if (old_nh) {
872 list_del_rcu(&fdb->nh_list);
873 nexthop_put(old_nh);
874 }
875 rcu_assign_pointer(fdb->nh, nh);
876 list_add_tail_rcu(&fdb->nh_list, &nh->fdb_list);
877 return 1;
878
879err_inval:
880 if (nh)
881 nexthop_put(nh);
882 return err;
883}
884
c63053e0
RP		 885int vxlan_fdb_create(struct vxlan_dev *vxlan,
886 const u8 *mac, union vxlan_addr *ip,
887 __u16 state, __be16 port, __be32 src_vni,
888 __be32 vni, __u32 ifindex, __u16 ndm_flags,
889 u32 nhid, struct vxlan_fdb **fdb,
890 struct netlink_ext_ack *extack)
25e20e73
RP		 891{
892 struct vxlan_rdst *rd = NULL;
893 struct vxlan_fdb *f;
894 int rc;
895
896 if (vxlan->cfg.addrmax &&
897 vxlan->addrcnt >= vxlan->cfg.addrmax)
898 return -ENOSPC;
899
900 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
c7cdbe2e 901 f = vxlan_fdb_alloc(vxlan, mac, state, src_vni, ndm_flags);
25e20e73
RP		 902 if (!f)
903 return -ENOMEM;
904
1274e1cc
RP		 905 if (nhid)
906 rc = vxlan_fdb_nh_update(vxlan, f, nhid, extack);
907 else
908 rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
909 if (rc < 0)
910 goto errout;
25e20e73 911
25e20e73
RP		 912 *fdb = f;
913
914 return 0;
1274e1cc
RP		 915
916errout:
917 kfree(f);
918 return rc;
25e20e73
RP		 919}
920
7c31e54a 921static void __vxlan_fdb_free(struct vxlan_fdb *f)
c2b200e0 922{
c2b200e0 923 struct vxlan_rdst *rd, *nd;
1274e1cc
RP		 924 struct nexthop *nh;
925
926 nh = rcu_dereference_raw(f->nh);
927 if (nh) {
928 rcu_assign_pointer(f->nh, NULL);
79472fe8 929 rcu_assign_pointer(f->vdev, NULL);
1274e1cc
RP		 930 nexthop_put(nh);
931 }
c2b200e0
PM		 932
933 list_for_each_entry_safe(rd, nd, &f->remotes, list) {
934 dst_cache_destroy(&rd->dst_cache);
935 kfree(rd);
936 }
937 kfree(f);
938}
939
7c31e54a
TY		 940static void vxlan_fdb_free(struct rcu_head *head)
941{
942 struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu);
943
944 __vxlan_fdb_free(f);
945}
946
c2b200e0
PM		 947static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f,
948 bool do_notify, bool swdev_notify)
949{
950 struct vxlan_rdst *rd;
951
952 netdev_dbg(vxlan->dev, "delete %pM\n", f->eth_addr);
953
954 --vxlan->addrcnt;
1274e1cc
RP		 955 if (do_notify) {
956 if (rcu_access_pointer(f->nh))
957 vxlan_fdb_notify(vxlan, f, NULL, RTM_DELNEIGH,
4c59b7d1 958 swdev_notify, NULL);
1274e1cc
RP		 959 else
960 list_for_each_entry(rd, &f->remotes, list)
961 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH,
962 swdev_notify, NULL);
963 }
c2b200e0
PM		 964
965 hlist_del_rcu(&f->hlist);
79472fe8 966 list_del_rcu(&f->nh_list);
c2b200e0
PM		 967 call_rcu(&f->rcu, vxlan_fdb_free);
968}
969
fc4aa1ca
PM		 970static void vxlan_dst_free(struct rcu_head *head)
971{
972 struct vxlan_rdst *rd = container_of(head, struct vxlan_rdst, rcu);
973
974 dst_cache_destroy(&rd->dst_cache);
975 kfree(rd);
976}
977
a76d1ca2
PM		 978static int vxlan_fdb_update_existing(struct vxlan_dev *vxlan,
979 union vxlan_addr *ip,
980 __u16 state, __u16 flags,
981 __be16 port, __be32 vni,
982 __u32 ifindex, __u16 ndm_flags,
1274e1cc 983 struct vxlan_fdb *f, u32 nhid,
4c59b7d1
PM		 984 bool swdev_notify,
985 struct netlink_ext_ack *extack)
d342894c 986{
45598c1c 987 __u16 fdb_flags = (ndm_flags & ~NTF_USE);
9e4b93f9 988 struct vxlan_rdst *rd = NULL;
ccdfd4f7 989 struct vxlan_rdst oldrd;
d342894c 990 int notify = 0;
61f46fe8
PM		 991 int rc = 0;
992 int err;
d342894c 993
1274e1cc
RP		 994 if (nhid && !rcu_access_pointer(f->nh)) {
995 NL_SET_ERR_MSG(extack,
996 "Cannot replace an existing non nexthop fdb with a nexthop");
997 return -EOPNOTSUPP;
998 }
999
1000 if (nhid && (flags & NLM_F_APPEND)) {
1001 NL_SET_ERR_MSG(extack,
1002 "Cannot append to a nexthop fdb");
1003 return -EOPNOTSUPP;
1004 }
1005
a76d1ca2
PM		 1006 /* Do not allow an externally learned entry to take over an entry added
1007 * by the user.
1008 */
1009 if (!(fdb_flags & NTF_EXT_LEARNED) ||
1010 !(f->flags & NTF_VXLAN_ADDED_BY_USER)) {
1011 if (f->state != state) {
1012 f->state = state;
1013 f->updated = jiffies;
1014 notify = 1;
ae884082 1015 }
a76d1ca2
PM		 1016 if (f->flags != fdb_flags) {
1017 f->flags = fdb_flags;
1018 f->updated = jiffies;
1019 notify = 1;
906dc186 1020 }
a76d1ca2 1021 }
6681712d 1022
a76d1ca2
PM		 1023 if ((flags & NLM_F_REPLACE)) {
1024 /* Only change unicasts */
1025 if (!(is_multicast_ether_addr(f->eth_addr) ||
1026 is_zero_ether_addr(f->eth_addr))) {
1274e1cc
RP		 1027 if (nhid) {
1028 rc = vxlan_fdb_nh_update(vxlan, f, nhid, extack);
1029 if (rc < 0)
1030 return rc;
1031 } else {
1032 rc = vxlan_fdb_replace(f, ip, port, vni,
1033 ifindex, &oldrd);
1034 }
6681712d 1035 notify |= rc;
a76d1ca2 1036 } else {
1274e1cc 1037 NL_SET_ERR_MSG(extack, "Cannot replace non-unicast fdb entries");
906dc186 1038 return -EOPNOTSUPP;
a76d1ca2
PM		 1039 }
1040 }
1041 if ((flags & NLM_F_APPEND) &&
1042 (is_multicast_ether_addr(f->eth_addr) ||
1043 is_zero_ether_addr(f->eth_addr))) {
1044 rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
906dc186 1045
25e20e73 1046 if (rc < 0)
17b46365 1047 return rc;
a76d1ca2 1048 notify |= rc;
d342894c 1049 }
1050
a76d1ca2
PM		 1051 if (ndm_flags & NTF_USE)
1052 f->used = jiffies;
1053
9e4b93f9
ND		 1054 if (notify) {
1055 if (rd == NULL)
1056 rd = first_remote_rtnl(f);
a76d1ca2 1057
61f46fe8 1058 err = vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH,
4c59b7d1 1059 swdev_notify, extack);
61f46fe8
PM		 1060 if (err)
1061 goto err_notify;
9e4b93f9 1062 }
d342894c 1063
1064 return 0;
61f46fe8
PM		 1065
1066err_notify:
1274e1cc
RP		 1067 if (nhid)
1068 return err;
61f46fe8
PM		 1069 if ((flags & NLM_F_REPLACE) && rc)
1070 *rd = oldrd;
fc4aa1ca 1071 else if ((flags & NLM_F_APPEND) && rc) {
61f46fe8 1072 list_del_rcu(&rd->list);
fc4aa1ca
PM		 1073 call_rcu(&rd->rcu, vxlan_dst_free);
1074 }
61f46fe8 1075 return err;
d342894c 1076}
1077
a76d1ca2
PM		 1078static int vxlan_fdb_update_create(struct vxlan_dev *vxlan,
1079 const u8 *mac, union vxlan_addr *ip,
1080 __u16 state, __u16 flags,
1081 __be16 port, __be32 src_vni, __be32 vni,
1274e1cc 1082 __u32 ifindex, __u16 ndm_flags, u32 nhid,
4c59b7d1
PM		 1083 bool swdev_notify,
1084 struct netlink_ext_ack *extack)
a76d1ca2
PM		 1085{
1086 __u16 fdb_flags = (ndm_flags & ~NTF_USE);
1087 struct vxlan_fdb *f;
1088 int rc;
1089
1090 /* Disallow replace to add a multicast entry */
1091 if ((flags & NLM_F_REPLACE) &&
1092 (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac)))
1093 return -EOPNOTSUPP;
1094
1095 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
1096 rc = vxlan_fdb_create(vxlan, mac, ip, state, port, src_vni,
1274e1cc 1097 vni, ifindex, fdb_flags, nhid, &f, extack);
a76d1ca2
PM		 1098 if (rc < 0)
1099 return rc;
1100
7c31e54a 1101 vxlan_fdb_insert(vxlan, mac, src_vni, f);
61f46fe8 1102 rc = vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_NEWNEIGH,
4c59b7d1 1103 swdev_notify, extack);
61f46fe8
PM		 1104 if (rc)
1105 goto err_notify;
1106
a76d1ca2 1107 return 0;
61f46fe8
PM		 1108
1109err_notify:
1110 vxlan_fdb_destroy(vxlan, f, false, false);
1111 return rc;
a76d1ca2
PM		 1112}
1113
1114/* Add new entry to forwarding table -- assumes lock held */
c63053e0
RP		 1115int vxlan_fdb_update(struct vxlan_dev *vxlan,
1116 const u8 *mac, union vxlan_addr *ip,
1117 __u16 state, __u16 flags,
1118 __be16 port, __be32 src_vni, __be32 vni,
1119 __u32 ifindex, __u16 ndm_flags, u32 nhid,
1120 bool swdev_notify,
1121 struct netlink_ext_ack *extack)
a76d1ca2
PM		 1122{
1123 struct vxlan_fdb *f;
1124
1125 f = __vxlan_find_mac(vxlan, mac, src_vni);
1126 if (f) {
1127 if (flags & NLM_F_EXCL) {
1128 netdev_dbg(vxlan->dev,
1129 "lost race to create %pM\n", mac);
1130 return -EEXIST;
1131 }
1132
1133 return vxlan_fdb_update_existing(vxlan, ip, state, flags, port,
1134 vni, ifindex, ndm_flags, f,
1274e1cc 1135 nhid, swdev_notify, extack);
a76d1ca2
PM		 1136 } else {
1137 if (!(flags & NLM_F_CREATE))
1138 return -ENOENT;
1139
1140 return vxlan_fdb_update_create(vxlan, mac, ip, state, flags,
1141 port, src_vni, vni, ifindex,
1274e1cc
RP		 1142 ndm_flags, nhid, swdev_notify,
1143 extack);
a76d1ca2
PM		 1144 }
1145}
1146
35cf2845 1147static void vxlan_fdb_dst_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f,
0e6160f3 1148 struct vxlan_rdst *rd, bool swdev_notify)
35cf2845
LR		 1149{
1150 list_del_rcu(&rd->list);
4c59b7d1 1151 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH, swdev_notify, NULL);
35cf2845
LR		 1152 call_rcu(&rd->rcu, vxlan_dst_free);
1153}
1154
f0b074be 1155static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan,
3ad7a4b1 1156 union vxlan_addr *ip, __be16 *port, __be32 *src_vni,
e92695e5
AM		 1157 __be32 *vni, u32 *ifindex, u32 *nhid,
1158 struct netlink_ext_ack *extack)
d342894c 1159{
6681712d 1160 struct net *net = dev_net(vxlan->dev);
e4c7ed41 1161 int err;
d342894c 1162
c2e10f53
AM		 1163 if (tb[NDA_NH_ID] &&
1164 (tb[NDA_DST] || tb[NDA_VNI] || tb[NDA_IFINDEX] || tb[NDA_PORT])) {
1165 NL_SET_ERR_MSG(extack, "DST, VNI, ifindex and port are mutually exclusive with NH_ID");
1166 return -EINVAL;
1167 }
72b48682 1168
f0b074be 1169 if (tb[NDA_DST]) {
e4c7ed41 1170 err = vxlan_nla_get_addr(ip, tb[NDA_DST]);
e92695e5
AM		 1171 if (err) {
1172 NL_SET_ERR_MSG(extack, "Unsupported address family");
e4c7ed41 1173 return err;
e92695e5 1174 }
f0b074be 1175 } else {
e4c7ed41 1176 union vxlan_addr *remote = &vxlan->default_dst.remote_ip;
98c81476 1177
e4c7ed41
CW		 1178 if (remote->sa.sa_family == AF_INET) {
1179 ip->sin.sin_addr.s_addr = htonl(INADDR_ANY);
1180 ip->sa.sa_family = AF_INET;
1181#if IS_ENABLED(CONFIG_IPV6)
1182 } else {
1183 ip->sin6.sin6_addr = in6addr_any;
1184 ip->sa.sa_family = AF_INET6;
1185#endif
1186 }
f0b074be 1187 }
d342894c 1188
6681712d 1189 if (tb[NDA_PORT]) {
e92695e5
AM		 1190 if (nla_len(tb[NDA_PORT]) != sizeof(__be16)) {
1191 NL_SET_ERR_MSG(extack, "Invalid vxlan port");
6681712d 1192 return -EINVAL;
e92695e5 1193 }
f0b074be
MR		 1194 *port = nla_get_be16(tb[NDA_PORT]);
1195 } else {
0dfbdf41 1196 *port = vxlan->cfg.dst_port;
f0b074be 1197 }
6681712d
DS		 1198
1199 if (tb[NDA_VNI]) {
e92695e5
AM		 1200 if (nla_len(tb[NDA_VNI]) != sizeof(u32)) {
1201 NL_SET_ERR_MSG(extack, "Invalid vni");
6681712d 1202 return -EINVAL;
e92695e5 1203 }
54bfd872 1204 *vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI]));
f0b074be
MR		 1205 } else {
1206 *vni = vxlan->default_dst.remote_vni;
1207 }
6681712d 1208
3ad7a4b1 1209 if (tb[NDA_SRC_VNI]) {
e92695e5
AM		 1210 if (nla_len(tb[NDA_SRC_VNI]) != sizeof(u32)) {
1211 NL_SET_ERR_MSG(extack, "Invalid src vni");
3ad7a4b1 1212 return -EINVAL;
e92695e5 1213 }
3ad7a4b1
RP		 1214 *src_vni = cpu_to_be32(nla_get_u32(tb[NDA_SRC_VNI]));
1215 } else {
1216 *src_vni = vxlan->default_dst.remote_vni;
1217 }
1218
6681712d 1219 if (tb[NDA_IFINDEX]) {
5abb0029 1220 struct net_device *tdev;
6681712d 1221
e92695e5
AM		 1222 if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32)) {
1223 NL_SET_ERR_MSG(extack, "Invalid ifindex");
6681712d 1224 return -EINVAL;
e92695e5 1225 }
f0b074be 1226 *ifindex = nla_get_u32(tb[NDA_IFINDEX]);
73763949 1227 tdev = __dev_get_by_index(net, *ifindex);
e92695e5
AM		 1228 if (!tdev) {
1229 NL_SET_ERR_MSG(extack, "Device not found");
6681712d 1230 return -EADDRNOTAVAIL;
e92695e5 1231 }
f0b074be
MR		 1232 } else {
1233 *ifindex = 0;
1234 }
1235
1274e1cc
RP		 1236 if (tb[NDA_NH_ID])
1237 *nhid = nla_get_u32(tb[NDA_NH_ID]);
1238 else
1239 *nhid = 0;
1240
f0b074be
MR		 1241 return 0;
1242}
1243
1244/* Add static entry (via netlink) */
1245static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[],
1246 struct net_device *dev,
87b0984e
PM		 1247 const unsigned char *addr, u16 vid, u16 flags,
1248 struct netlink_ext_ack *extack)
f0b074be
MR		 1249{
1250 struct vxlan_dev *vxlan = netdev_priv(dev);
1251 /* struct net *net = dev_net(vxlan->dev); */
e4c7ed41 1252 union vxlan_addr ip;
f0b074be 1253 __be16 port;
3ad7a4b1 1254 __be32 src_vni, vni;
1274e1cc 1255 u32 ifindex, nhid;
fe1e0713 1256 u32 hash_index;
f0b074be
MR		 1257 int err;
1258
1259 if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) {
1260 pr_info("RTM_NEWNEIGH with invalid state %#x\n",
1261 ndm->ndm_state);
1262 return -EINVAL;
1263 }
1264
1274e1cc 1265 if (!tb || (!tb[NDA_DST] && !tb[NDA_NH_ID]))
f0b074be
MR		 1266 return -EINVAL;
1267
1274e1cc 1268 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &src_vni, &vni, &ifindex,
e92695e5 1269 &nhid, extack);
f0b074be
MR		 1270 if (err)
1271 return err;
6681712d 1272
5933a7bb
MR		 1273 if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family)
1274 return -EAFNOSUPPORT;
1275
fe1e0713
L		 1276 hash_index = fdb_head_index(vxlan, addr, src_vni);
1277 spin_lock_bh(&vxlan->hash_lock[hash_index]);
25e20e73 1278 err = vxlan_fdb_update(vxlan, addr, &ip, ndm->ndm_state, flags,
45598c1c
PM		 1279 port, src_vni, vni, ifindex,
1280 ndm->ndm_flags | NTF_VXLAN_ADDED_BY_USER,
1274e1cc 1281 nhid, true, extack);
fe1e0713 1282 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
d342894c 1283
1284 return err;
1285}
1286
c63053e0
RP		 1287int __vxlan_fdb_delete(struct vxlan_dev *vxlan,
1288 const unsigned char *addr, union vxlan_addr ip,
1289 __be16 port, __be32 src_vni, __be32 vni,
1290 u32 ifindex, bool swdev_notify)
d342894c 1291{
bc7892ba 1292 struct vxlan_rdst *rd = NULL;
1274e1cc 1293 struct vxlan_fdb *f;
3ad7a4b1 1294 int err = -ENOENT;
bc7892ba 1295
3ad7a4b1 1296 f = vxlan_find_mac(vxlan, addr, src_vni);
bc7892ba 1297 if (!f)
3ad7a4b1 1298 return err;
bc7892ba 1299
e4c7ed41
CW		 1300 if (!vxlan_addr_any(&ip)) {
1301 rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex);
bc7892ba
MR		 1302 if (!rd)
1303 goto out;
1304 }
1305
bc7892ba
MR		 1306 /* remove a destination if it's not the only one on the list,
1307 * otherwise destroy the fdb entry
1308 */
1309 if (rd && !list_is_singular(&f->remotes)) {
0e6160f3 1310 vxlan_fdb_dst_destroy(vxlan, f, rd, swdev_notify);
bc7892ba 1311 goto out;
d342894c 1312 }
bc7892ba 1313
0e6160f3 1314 vxlan_fdb_destroy(vxlan, f, true, swdev_notify);
bc7892ba
MR		 1315
1316out:
3ad7a4b1
RP		 1317 return 0;
1318}
1319
1320/* Delete entry (via netlink) */
1321static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[],
1322 struct net_device *dev,
ca4567f1 1323 const unsigned char *addr, u16 vid,
c2e10f53 1324 struct netlink_ext_ack *extack)
3ad7a4b1
RP		 1325{
1326 struct vxlan_dev *vxlan = netdev_priv(dev);
1327 union vxlan_addr ip;
1328 __be32 src_vni, vni;
1274e1cc 1329 u32 ifindex, nhid;
fe1e0713 1330 u32 hash_index;
1274e1cc 1331 __be16 port;
3ad7a4b1
RP		 1332 int err;
1333
1274e1cc 1334 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &src_vni, &vni, &ifindex,
e92695e5 1335 &nhid, extack);
3ad7a4b1
RP		 1336 if (err)
1337 return err;
1338
fe1e0713
L		 1339 hash_index = fdb_head_index(vxlan, addr, src_vni);
1340 spin_lock_bh(&vxlan->hash_lock[hash_index]);
0e6160f3
PM		 1341 err = __vxlan_fdb_delete(vxlan, addr, ip, port, src_vni, vni, ifindex,
1342 true);
fe1e0713 1343 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
d342894c 1344
1345 return err;
1346}
1347
1348/* Dump forwarding table */
1349static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb,
5d5eacb3 1350 struct net_device *dev,
d297653d 1351 struct net_device *filter_dev, int *idx)
d342894c 1352{
1353 struct vxlan_dev *vxlan = netdev_priv(dev);
1354 unsigned int h;
d297653d 1355 int err = 0;
d342894c 1356
1357 for (h = 0; h < FDB_HASH_SIZE; ++h) {
1358 struct vxlan_fdb *f;
d342894c 1359
b5141915 1360 rcu_read_lock();
b67bfe0d 1361 hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) {
6681712d 1362 struct vxlan_rdst *rd;
6681712d 1363
1274e1cc 1364 if (rcu_access_pointer(f->nh)) {
b18e9834
RP		 1365 if (*idx < cb->args[2])
1366 goto skip_nh;
1274e1cc
RP		 1367 err = vxlan_fdb_info(skb, vxlan, f,
1368 NETLINK_CB(cb->skb).portid,
1369 cb->nlh->nlmsg_seq,
1370 RTM_NEWNEIGH,
1371 NLM_F_MULTI, NULL);
b5141915
IS		 1372 if (err < 0) {
1373 rcu_read_unlock();
1274e1cc 1374 goto out;
b5141915 1375 }
b18e9834
RP		 1376skip_nh:
1377 *idx += 1;
1274e1cc
RP		 1378 continue;
1379 }
1380
3e61aa8f 1381 list_for_each_entry_rcu(rd, &f->remotes, list) {
d297653d 1382 if (*idx < cb->args[2])
07a51cd3
AW		 1383 goto skip;
1384
6681712d
DS		 1385 err = vxlan_fdb_info(skb, vxlan, f,
1386 NETLINK_CB(cb->skb).portid,
1387 cb->nlh->nlmsg_seq,
1388 RTM_NEWNEIGH,
1389 NLM_F_MULTI, rd);
b5141915
IS		 1390 if (err < 0) {
1391 rcu_read_unlock();
3e61aa8f 1392 goto out;
b5141915 1393 }
3e61aa8f 1394skip:
d297653d 1395 *idx += 1;
07a51cd3 1396 }
d342894c 1397 }
b5141915 1398 rcu_read_unlock();
d342894c 1399 }
3e61aa8f 1400out:
d297653d 1401 return err;
d342894c 1402}
1403
474c3c89
RP		 1404static int vxlan_fdb_get(struct sk_buff *skb,
1405 struct nlattr *tb[],
1406 struct net_device *dev,
1407 const unsigned char *addr,
1408 u16 vid, u32 portid, u32 seq,
1409 struct netlink_ext_ack *extack)
1410{
1411 struct vxlan_dev *vxlan = netdev_priv(dev);
1412 struct vxlan_fdb *f;
1413 __be32 vni;
1414 int err;
1415
1416 if (tb[NDA_VNI])
1417 vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI]));
1418 else
1419 vni = vxlan->default_dst.remote_vni;
1420
1421 rcu_read_lock();
1422
1423 f = __vxlan_find_mac(vxlan, addr, vni);
1424 if (!f) {
1425 NL_SET_ERR_MSG(extack, "Fdb entry not found");
1426 err = -ENOENT;
1427 goto errout;
1428 }
1429
1430 err = vxlan_fdb_info(skb, vxlan, f, portid, seq,
1431 RTM_NEWNEIGH, 0, first_remote_rcu(f));
1432errout:
1433 rcu_read_unlock();
1434 return err;
1435}
1436
d342894c 1437/* Watch incoming packets to learn mapping between Ethernet address
1438 * and Tunnel endpoint.
c4b49512 1439 * Return true if packet is bogus and should be dropped.
d342894c 1440 */
26a41ae6 1441static bool vxlan_snoop(struct net_device *dev,
3ad7a4b1 1442 union vxlan_addr *src_ip, const u8 *src_mac,
87613de9 1443 u32 src_ifindex, __be32 vni)
d342894c 1444{
1445 struct vxlan_dev *vxlan = netdev_priv(dev);
1446 struct vxlan_fdb *f;
87613de9
MS		 1447 u32 ifindex = 0;
1448
1449#if IS_ENABLED(CONFIG_IPV6)
1450 if (src_ip->sa.sa_family == AF_INET6 &&
1451 (ipv6_addr_type(&src_ip->sin6.sin6_addr) & IPV6_ADDR_LINKLOCAL))
1452 ifindex = src_ifindex;
1453#endif
d342894c 1454
3ad7a4b1 1455 f = vxlan_find_mac(vxlan, src_mac, vni);
d342894c 1456 if (likely(f)) {
5ca5461c 1457 struct vxlan_rdst *rdst = first_remote_rcu(f);
3e61aa8f 1458
87613de9
MS		 1459 if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip) &&
1460 rdst->remote_ifindex == ifindex))
26a41ae6 1461 return false;
1462
1463 /* Don't migrate static entries, drop packets */
e0090a9e 1464 if (f->state & (NUD_PERMANENT | NUD_NOARP))
26a41ae6 1465 return true;
d342894c 1466
1274e1cc
RP		 1467 /* Don't override an fdb with nexthop with a learnt entry */
1468 if (rcu_access_pointer(f->nh))
1469 return true;
1470
d342894c 1471 if (net_ratelimit())
1472 netdev_info(dev,
e4c7ed41 1473 "%pM migrated from %pIS to %pIS\n",
a4870f79 1474 src_mac, &rdst->remote_ip.sa, &src_ip->sa);
d342894c 1475
e4c7ed41 1476 rdst->remote_ip = *src_ip;
d342894c 1477 f->updated = jiffies;
4c59b7d1 1478 vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH, true, NULL);
d342894c 1479 } else {
fe1e0713
L		 1480 u32 hash_index = fdb_head_index(vxlan, src_mac, vni);
1481
d342894c 1482 /* learned new entry */
fe1e0713 1483 spin_lock(&vxlan->hash_lock[hash_index]);
3bf74b1a 1484
1485 /* close off race between vxlan_flush and incoming packets */
1486 if (netif_running(dev))
25e20e73 1487 vxlan_fdb_update(vxlan, src_mac, src_ip,
3bf74b1a 1488 NUD_REACHABLE,
1489 NLM_F_EXCL|NLM_F_CREATE,
0dfbdf41 1490 vxlan->cfg.dst_port,
3ad7a4b1 1491 vni,
3bf74b1a 1492 vxlan->default_dst.remote_vni,
1274e1cc 1493 ifindex, NTF_SELF, 0, true, NULL);
fe1e0713 1494 spin_unlock(&vxlan->hash_lock[hash_index]);
d342894c 1495 }
26a41ae6 1496
1497 return false;
d342894c 1498}
1499
544a773a 1500static bool __vxlan_sock_release_prep(struct vxlan_sock *vs)
7c47cedf 1501{
b1be00a6 1502 struct vxlan_net *vn;
012a5729 1503
b1be00a6 1504 if (!vs)
544a773a 1505 return false;
66af846f 1506 if (!refcount_dec_and_test(&vs->refcnt))
544a773a 1507 return false;
d342894c 1508
b1be00a6 1509 vn = net_generic(sock_net(vs->sock->sk), vxlan_net_id);
1c51a915 1510 spin_lock(&vn->sock_lock);
7c47cedf 1511 hlist_del_rcu(&vs->hlist);
e7b3db5e 1512 udp_tunnel_notify_del_rx_port(vs->sock,
b9adcd69
AD		 1513 (vs->flags & VXLAN_F_GPE) ?
1514 UDP_TUNNEL_TYPE_VXLAN_GPE :
e7b3db5e 1515 UDP_TUNNEL_TYPE_VXLAN);
1c51a915
SH		 1516 spin_unlock(&vn->sock_lock);
1517
544a773a 1518 return true;
d342894c 1519}
1520
b1be00a6
JB		 1521static void vxlan_sock_release(struct vxlan_dev *vxlan)
1522{
c6fcc4fc 1523 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
b1be00a6 1524#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 1525 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1526
57d88182 1527 RCU_INIT_POINTER(vxlan->vn6_sock, NULL);
544a773a
HFS		 1528#endif
1529
57d88182 1530 RCU_INIT_POINTER(vxlan->vn4_sock, NULL);
544a773a
HFS		 1531 synchronize_net();
1532
f9c4bb0b
RP		 1533 if (vxlan->cfg.flags & VXLAN_F_VNIFILTER)
1534 vxlan_vs_del_vnigrp(vxlan);
1535 else
1536 vxlan_vs_del_dev(vxlan);
a53cb29b 1537
c6fcc4fc 1538 if (__vxlan_sock_release_prep(sock4)) {
1539 udp_tunnel_sock_release(sock4->sock);
1540 kfree(sock4);
544a773a
HFS		 1541 }
1542
1543#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 1544 if (__vxlan_sock_release_prep(sock6)) {
1545 udp_tunnel_sock_release(sock6->sock);
1546 kfree(sock6);
544a773a 1547 }
b1be00a6
JB		 1548#endif
1549}
1550
f14ecebb
JB		 1551static bool vxlan_remcsum(struct vxlanhdr *unparsed,
1552 struct sk_buff *skb, u32 vxflags)
dfd8645e 1553{
7d34fa75 1554 size_t start, offset;
dfd8645e 1555
f14ecebb
JB		 1556 if (!(unparsed->vx_flags & VXLAN_HF_RCO) || skb->remcsum_offload)
1557 goto out;
b7fe10e5 1558
f14ecebb
JB		 1559 start = vxlan_rco_start(unparsed->vx_vni);
1560 offset = start + vxlan_rco_offset(unparsed->vx_vni);
dfd8645e 1561
7d34fa75 1562 if (!pskb_may_pull(skb, offset + sizeof(u16)))
be5cfeab 1563 return false;
dfd8645e 1564
be5cfeab
JB		 1565 skb_remcsum_process(skb, (void *)(vxlan_hdr(skb) + 1), start, offset,
1566 !!(vxflags & VXLAN_F_REMCSUM_NOPARTIAL));
f14ecebb
JB		 1567out:
1568 unparsed->vx_flags &= ~VXLAN_HF_RCO;
1569 unparsed->vx_vni &= VXLAN_VNI_MASK;
be5cfeab 1570 return true;
dfd8645e
TH		 1571}
1572
f14ecebb 1573static void vxlan_parse_gbp_hdr(struct vxlanhdr *unparsed,
64f87d36 1574 struct sk_buff *skb, u32 vxflags,
10a5af23 1575 struct vxlan_metadata *md)
3288af08 1576{
f14ecebb 1577 struct vxlanhdr_gbp *gbp = (struct vxlanhdr_gbp *)unparsed;
10a5af23 1578 struct metadata_dst *tun_dst;
f14ecebb
JB		 1579
1580 if (!(unparsed->vx_flags & VXLAN_HF_GBP))
1581 goto out;
3288af08 1582
3288af08
JB		 1583 md->gbp = ntohs(gbp->policy_id);
1584
10a5af23 1585 tun_dst = (struct metadata_dst *)skb_dst(skb);
810813c4 1586 if (tun_dst) {
3288af08 1587 tun_dst->u.tun_info.key.tun_flags |= TUNNEL_VXLAN_OPT;
810813c4
DM		 1588 tun_dst->u.tun_info.options_len = sizeof(*md);
1589 }
3288af08
JB		 1590 if (gbp->dont_learn)
1591 md->gbp |= VXLAN_GBP_DONT_LEARN;
1592
1593 if (gbp->policy_applied)
1594 md->gbp |= VXLAN_GBP_POLICY_APPLIED;
f14ecebb 1595
64f87d36
JB		 1596 /* In flow-based mode, GBP is carried in dst_metadata */
1597 if (!(vxflags & VXLAN_F_COLLECT_METADATA))
1598 skb->mark = md->gbp;
f14ecebb
JB		 1599out:
1600 unparsed->vx_flags &= ~VXLAN_GBP_USED_BITS;
3288af08
JB		 1601}
1602
1ab016e2
JB		 1603static bool vxlan_set_mac(struct vxlan_dev *vxlan,
1604 struct vxlan_sock *vs,
3ad7a4b1 1605 struct sk_buff *skb, __be32 vni)
614732ea 1606{
614732ea 1607 union vxlan_addr saddr;
87613de9 1608 u32 ifindex = skb->dev->ifindex;
614732ea 1609
614732ea 1610 skb_reset_mac_header(skb);
614732ea
TG		 1611 skb->protocol = eth_type_trans(skb, vxlan->dev);
1612 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1613
1614 /* Ignore packet loops (and multicast echo) */
1615 if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr))
1ab016e2 1616 return false;
614732ea 1617
760c6805 1618 /* Get address from the outer IP header */
ce212d0f 1619 if (vxlan_get_sk_family(vs) == AF_INET) {
1ab016e2 1620 saddr.sin.sin_addr.s_addr = ip_hdr(skb)->saddr;
614732ea
TG		 1621 saddr.sa.sa_family = AF_INET;
1622#if IS_ENABLED(CONFIG_IPV6)
1623 } else {
1ab016e2 1624 saddr.sin6.sin6_addr = ipv6_hdr(skb)->saddr;
614732ea
TG		 1625 saddr.sa.sa_family = AF_INET6;
1626#endif
1627 }
1628
dc5321d7 1629 if ((vxlan->cfg.flags & VXLAN_F_LEARN) &&
87613de9 1630 vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source, ifindex, vni))
1ab016e2
JB		 1631 return false;
1632
1633 return true;
1634}
1635
760c6805
JB		 1636static bool vxlan_ecn_decapsulate(struct vxlan_sock *vs, void *oiph,
1637 struct sk_buff *skb)
1638{
1639 int err = 0;
1640
1641 if (vxlan_get_sk_family(vs) == AF_INET)
1642 err = IP_ECN_decapsulate(oiph, skb);
1643#if IS_ENABLED(CONFIG_IPV6)
1644 else
1645 err = IP6_ECN_decapsulate(oiph, skb);
1646#endif
1647
1648 if (unlikely(err) && log_ecn_error) {
1649 if (vxlan_get_sk_family(vs) == AF_INET)
1650 net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n",
1651 &((struct iphdr *)oiph)->saddr,
1652 ((struct iphdr *)oiph)->tos);
1653 else
1654 net_info_ratelimited("non-ECT from %pI6\n",
1655 &((struct ipv6hdr *)oiph)->saddr);
1656 }
1657 return err <= 1;
1658}
1659
d342894c 1660/* Callback from net/ipv4/udp.c to receive packets */
f2d1968e 1661static int vxlan_rcv(struct sock *sk, struct sk_buff *skb)
d342894c 1662{
4095e0e1 1663 struct vxlan_vni_node *vninode = NULL;
c9e78efb 1664 struct vxlan_dev *vxlan;
5cfccc5a 1665 struct vxlan_sock *vs;
f14ecebb 1666 struct vxlanhdr unparsed;
ee122c79
TG		 1667 struct vxlan_metadata _md;
1668 struct vxlan_metadata *md = &_md;
61618eea 1669 __be16 protocol = htons(ETH_P_TEB);
e1e5314d 1670 bool raw_proto = false;
f2d1968e 1671 void *oiph;
3ad7a4b1 1672 __be32 vni = 0;
d342894c 1673
e1e5314d 1674 /* Need UDP and VXLAN header to be present */
7ce04758 1675 if (!pskb_may_pull(skb, VXLAN_HLEN))
e5aed006 1676 goto drop;
d342894c 1677
f14ecebb 1678 unparsed = *vxlan_hdr(skb);
288b01c8
JB		 1679 /* VNI flag always required to be set */
1680 if (!(unparsed.vx_flags & VXLAN_HF_VNI)) {
1681 netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n",
1682 ntohl(vxlan_hdr(skb)->vx_flags),
1683 ntohl(vxlan_hdr(skb)->vx_vni));
1684 /* Return non vxlan pkt */
e5aed006 1685 goto drop;
d342894c 1686 }
288b01c8
JB		 1687 unparsed.vx_flags &= ~VXLAN_HF_VNI;
1688 unparsed.vx_vni &= ~VXLAN_VNI_MASK;
d342894c 1689
559835ea 1690 vs = rcu_dereference_sk_user_data(sk);
5cfccc5a 1691 if (!vs)
d342894c 1692 goto drop;
d342894c 1693
3ad7a4b1
RP		 1694 vni = vxlan_vni(vxlan_hdr(skb)->vx_vni);
1695
4095e0e1 1696 vxlan = vxlan_vs_find_vni(vs, skb->dev->ifindex, vni, &vninode);
c9e78efb
JB		 1697 if (!vxlan)
1698 goto drop;
1699
e1e5314d
JB		 1700 /* For backwards compatibility, only allow reserved fields to be
1701 * used by VXLAN extensions if explicitly requested.
1702 */
1703 if (vs->flags & VXLAN_F_GPE) {
17a0a644 1704 if (!vxlan_parse_gpe_proto(&unparsed, &protocol))
e1e5314d 1705 goto drop;
17a0a644 1706 unparsed.vx_flags &= ~VXLAN_GPE_USED_BITS;
e1e5314d
JB		 1707 raw_proto = true;
1708 }
1709
1710 if (__iptunnel_pull_header(skb, VXLAN_HLEN, protocol, raw_proto,
1711 !net_eq(vxlan->net, dev_net(vxlan->dev))))
98c81476 1712 goto drop;
c9e78efb 1713
2ae2904b 1714 if (vs->flags & VXLAN_F_REMCSUM_RX)
0189399c 1715 if (unlikely(!vxlan_remcsum(&unparsed, skb, vs->flags)))
2ae2904b
FF		 1716 goto drop;
1717
ee122c79 1718 if (vxlan_collect_metadata(vs)) {
10a5af23 1719 struct metadata_dst *tun_dst;
07dabf20 1720
c29a70d2 1721 tun_dst = udp_tun_rx_dst(skb, vxlan_get_sk_family(vs), TUNNEL_KEY,
d817f432 1722 key32_to_tunnel_id(vni), sizeof(*md));
c29a70d2 1723
ee122c79
TG		 1724 if (!tun_dst)
1725 goto drop;
1726
0f1b7354 1727 md = ip_tunnel_info_opts(&tun_dst->u.tun_info);
10a5af23
JB		 1728
1729 skb_dst_set(skb, (struct dst_entry *)tun_dst);
ee122c79
TG		 1730 } else {
1731 memset(md, 0, sizeof(*md));
1732 }
1733
f14ecebb 1734 if (vs->flags & VXLAN_F_GBP)
10a5af23 1735 vxlan_parse_gbp_hdr(&unparsed, skb, vs->flags, md);
e1e5314d
JB		 1736 /* Note that GBP and GPE can never be active together. This is
1737 * ensured in vxlan_dev_configure.
1738 */
3511494c 1739
f14ecebb 1740 if (unparsed.vx_flags || unparsed.vx_vni) {
3bf39475
TH		 1741 /* If there are any unprocessed flags remaining treat
1742 * this as a malformed packet. This behavior diverges from
1743 * VXLAN RFC (RFC7348) which stipulates that bits in reserved
1744 * in reserved fields are to be ignored. The approach here
c4b49512 1745 * maintains compatibility with previous stack code, and also
3bf39475
TH		 1746 * is more robust and provides a little more security in
1747 * adding extensions to VXLAN.
1748 */
288b01c8 1749 goto drop;
3bf39475
TH		 1750 }
1751
e1e5314d 1752 if (!raw_proto) {
3ad7a4b1 1753 if (!vxlan_set_mac(vxlan, vs, skb, vni))
e1e5314d
JB		 1754 goto drop;
1755 } else {
8be0cfa4 1756 skb_reset_mac_header(skb);
e1e5314d
JB		 1757 skb->dev = vxlan->dev;
1758 skb->pkt_type = PACKET_HOST;
1759 }
f2d1968e 1760
f2d1968e
JB		 1761 oiph = skb_network_header(skb);
1762 skb_reset_network_header(skb);
1763
1764 if (!vxlan_ecn_decapsulate(vs, oiph, skb)) {
1765 ++vxlan->dev->stats.rx_frame_errors;
1766 ++vxlan->dev->stats.rx_errors;
4095e0e1
NA		 1767 vxlan_vnifilter_count(vxlan, vni, vninode,
1768 VXLAN_VNI_STATS_RX_ERRORS, 0);
f2d1968e
JB		 1769 goto drop;
1770 }
1771
59cbf56f
ED		 1772 rcu_read_lock();
1773
1774 if (unlikely(!(vxlan->dev->flags & IFF_UP))) {
1775 rcu_read_unlock();
625788b5 1776 dev_core_stats_rx_dropped_inc(vxlan->dev);
4095e0e1
NA		 1777 vxlan_vnifilter_count(vxlan, vni, vninode,
1778 VXLAN_VNI_STATS_RX_DROPS, 0);
59cbf56f
ED		 1779 goto drop;
1780 }
1781
1f8dda1d 1782 dev_sw_netstats_rx_add(vxlan->dev, skb->len);
4095e0e1 1783 vxlan_vnifilter_count(vxlan, vni, vninode, VXLAN_VNI_STATS_RX, skb->len);
f2d1968e 1784 gro_cells_receive(&vxlan->gro_cells, skb);
59cbf56f
ED		 1785
1786 rcu_read_unlock();
1787
5cfccc5a
PS		 1788 return 0;
1789
1790drop:
288b01c8
JB		 1791 /* Consume bad packet */
1792 kfree_skb(skb);
1793 return 0;
5cfccc5a
PS		 1794}
1795
c3a43b9f
SB		 1796/* Callback from net/ipv{4,6}/udp.c to check that we have a VNI for errors */
1797static int vxlan_err_lookup(struct sock *sk, struct sk_buff *skb)
1798{
1799 struct vxlan_dev *vxlan;
1800 struct vxlan_sock *vs;
1801 struct vxlanhdr *hdr;
1802 __be32 vni;
1803
8399a693 1804 if (!pskb_may_pull(skb, skb_transport_offset(skb) + VXLAN_HLEN))
c3a43b9f
SB		 1805 return -EINVAL;
1806
1807 hdr = vxlan_hdr(skb);
1808
1809 if (!(hdr->vx_flags & VXLAN_HF_VNI))
1810 return -EINVAL;
1811
1812 vs = rcu_dereference_sk_user_data(sk);
1813 if (!vs)
1814 return -ENOENT;
1815
1816 vni = vxlan_vni(hdr->vx_vni);
4095e0e1 1817 vxlan = vxlan_vs_find_vni(vs, skb->dev->ifindex, vni, NULL);
c3a43b9f
SB		 1818 if (!vxlan)
1819 return -ENOENT;
1820
1821 return 0;
1822}
1823
3ad7a4b1 1824static int arp_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni)
e4f67add
DS		 1825{
1826 struct vxlan_dev *vxlan = netdev_priv(dev);
1827 struct arphdr *parp;
1828 u8 *arpptr, *sha;
1829 __be32 sip, tip;
1830 struct neighbour *n;
1831
1832 if (dev->flags & IFF_NOARP)
1833 goto out;
1834
1835 if (!pskb_may_pull(skb, arp_hdr_len(dev))) {
1836 dev->stats.tx_dropped++;
1837 goto out;
1838 }
1839 parp = arp_hdr(skb);
1840
1841 if ((parp->ar_hrd != htons(ARPHRD_ETHER) &&
1842 parp->ar_hrd != htons(ARPHRD_IEEE802)) ||
1843 parp->ar_pro != htons(ETH_P_IP) ||
1844 parp->ar_op != htons(ARPOP_REQUEST) ||
1845 parp->ar_hln != dev->addr_len ||
1846 parp->ar_pln != 4)
1847 goto out;
1848 arpptr = (u8 *)parp + sizeof(struct arphdr);
1849 sha = arpptr;
1850 arpptr += dev->addr_len; /* sha */
1851 memcpy(&sip, arpptr, sizeof(sip));
1852 arpptr += sizeof(sip);
1853 arpptr += dev->addr_len; /* tha */
1854 memcpy(&tip, arpptr, sizeof(tip));
1855
1856 if (ipv4_is_loopback(tip) ||
1857 ipv4_is_multicast(tip))
1858 goto out;
1859
1860 n = neigh_lookup(&arp_tbl, &tip, dev);
1861
1862 if (n) {
e4f67add
DS		 1863 struct vxlan_fdb *f;
1864 struct sk_buff *reply;
1865
b071af52 1866 if (!(READ_ONCE(n->nud_state) & NUD_CONNECTED)) {
e4f67add
DS		 1867 neigh_release(n);
1868 goto out;
1869 }
1870
3ad7a4b1 1871 f = vxlan_find_mac(vxlan, n->ha, vni);
e4c7ed41 1872 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
e4f67add
DS		 1873 /* bridge-local neighbor */
1874 neigh_release(n);
1875 goto out;
1876 }
1877
1878 reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
1879 n->ha, sha);
1880
1881 neigh_release(n);
1882
7346135d
DS		 1883 if (reply == NULL)
1884 goto out;
1885
e4f67add
DS		 1886 skb_reset_mac_header(reply);
1887 __skb_pull(reply, skb_network_offset(reply));
1888 reply->ip_summed = CHECKSUM_UNNECESSARY;
1889 reply->pkt_type = PACKET_HOST;
1890
3d391f65 1891 if (netif_rx(reply) == NET_RX_DROP) {
e4f67add 1892 dev->stats.rx_dropped++;
4095e0e1
NA		 1893 vxlan_vnifilter_count(vxlan, vni, NULL,
1894 VXLAN_VNI_STATS_RX_DROPS, 0);
1895 }
1896
dc5321d7 1897 } else if (vxlan->cfg.flags & VXLAN_F_L3MISS) {
e4c7ed41
CW		 1898 union vxlan_addr ipa = {
1899 .sin.sin_addr.s_addr = tip,
a45e92a5 1900 .sin.sin_family = AF_INET,
e4c7ed41
CW		 1901 };
1902
1903 vxlan_ip_miss(dev, &ipa);
1904 }
e4f67add
DS		 1905out:
1906 consume_skb(skb);
1907 return NETDEV_TX_OK;
1908}
1909
f564f45c 1910#if IS_ENABLED(CONFIG_IPV6)
4b29dba9
DS		 1911static struct sk_buff *vxlan_na_create(struct sk_buff *request,
1912 struct neighbour *n, bool isrouter)
1913{
1914 struct net_device *dev = request->dev;
1915 struct sk_buff *reply;
1916 struct nd_msg *ns, *na;
1917 struct ipv6hdr *pip6;
1918 u8 *daddr;
1919 int na_olen = 8; /* opt hdr + ETH_ALEN for target */
1920 int ns_olen;
1921 int i, len;
1922
f1fb08f6 1923 if (dev == NULL || !pskb_may_pull(request, request->len))
4b29dba9
DS		 1924 return NULL;
1925
1926 len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) +
1927 sizeof(*na) + na_olen + dev->needed_tailroom;
1928 reply = alloc_skb(len, GFP_ATOMIC);
1929 if (reply == NULL)
1930 return NULL;
1931
1932 reply->protocol = htons(ETH_P_IPV6);
1933 reply->dev = dev;
1934 skb_reserve(reply, LL_RESERVED_SPACE(request->dev));
1935 skb_push(reply, sizeof(struct ethhdr));
6297b91c 1936 skb_reset_mac_header(reply);
4b29dba9 1937
f1fb08f6 1938 ns = (struct nd_msg *)(ipv6_hdr(request) + 1);
4b29dba9
DS		 1939
1940 daddr = eth_hdr(request)->h_source;
f1fb08f6
VB		 1941 ns_olen = request->len - skb_network_offset(request) -
1942 sizeof(struct ipv6hdr) - sizeof(*ns);
4b29dba9 1943 for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) {
8066e6b4
IS		 1944 if (!ns->opt[i + 1]) {
1945 kfree_skb(reply);
1946 return NULL;
1947 }
4b29dba9
DS		 1948 if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) {
1949 daddr = ns->opt + i + sizeof(struct nd_opt_hdr);
1950 break;
1951 }
1952 }
1953
1954 /* Ethernet header */
1955 ether_addr_copy(eth_hdr(reply)->h_dest, daddr);
1956 ether_addr_copy(eth_hdr(reply)->h_source, n->ha);
1957 eth_hdr(reply)->h_proto = htons(ETH_P_IPV6);
1958 reply->protocol = htons(ETH_P_IPV6);
1959
1960 skb_pull(reply, sizeof(struct ethhdr));
6297b91c 1961 skb_reset_network_header(reply);
4b29dba9
DS		 1962 skb_put(reply, sizeof(struct ipv6hdr));
1963
1964 /* IPv6 header */
1965
1966 pip6 = ipv6_hdr(reply);
1967 memset(pip6, 0, sizeof(struct ipv6hdr));
1968 pip6->version = 6;
1969 pip6->priority = ipv6_hdr(request)->priority;
1970 pip6->nexthdr = IPPROTO_ICMPV6;
1971 pip6->hop_limit = 255;
1972 pip6->daddr = ipv6_hdr(request)->saddr;
1973 pip6->saddr = *(struct in6_addr *)n->primary_key;
1974
1975 skb_pull(reply, sizeof(struct ipv6hdr));
6297b91c 1976 skb_reset_transport_header(reply);
4b29dba9 1977
4b29dba9 1978 /* Neighbor Advertisement */
b080db58 1979 na = skb_put_zero(reply, sizeof(*na) + na_olen);
4b29dba9
DS		 1980 na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT;
1981 na->icmph.icmp6_router = isrouter;
1982 na->icmph.icmp6_override = 1;
1983 na->icmph.icmp6_solicited = 1;
1984 na->target = ns->target;
1985 ether_addr_copy(&na->opt[2], n->ha);
1986 na->opt[0] = ND_OPT_TARGET_LL_ADDR;
1987 na->opt[1] = na_olen >> 3;
1988
1989 na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr,
1990 &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6,
1991 csum_partial(na, sizeof(*na)+na_olen, 0));
1992
1993 pip6->payload_len = htons(sizeof(*na)+na_olen);
1994
1995 skb_push(reply, sizeof(struct ipv6hdr));
1996
1997 reply->ip_summed = CHECKSUM_UNNECESSARY;
1998
1999 return reply;
2000}
2001
3ad7a4b1 2002static int neigh_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni)
f564f45c
CW		 2003{
2004 struct vxlan_dev *vxlan = netdev_priv(dev);
8dcd81a9 2005 const struct in6_addr *daddr;
8bff3685 2006 const struct ipv6hdr *iphdr;
4b29dba9 2007 struct inet6_dev *in6_dev;
8bff3685
XL		 2008 struct neighbour *n;
2009 struct nd_msg *msg;
f564f45c 2010
85e8b032 2011 rcu_read_lock();
f564f45c
CW		 2012 in6_dev = __in6_dev_get(dev);
2013 if (!in6_dev)
2014 goto out;
2015
f564f45c 2016 iphdr = ipv6_hdr(skb);
f564f45c 2017 daddr = &iphdr->daddr;
f1fb08f6 2018 msg = (struct nd_msg *)(iphdr + 1);
f564f45c 2019
4b29dba9
DS		 2020 if (ipv6_addr_loopback(daddr) ||
2021 ipv6_addr_is_multicast(&msg->target))
2022 goto out;
2023
2024 n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev);
f564f45c
CW		 2025
2026 if (n) {
2027 struct vxlan_fdb *f;
4b29dba9 2028 struct sk_buff *reply;
f564f45c 2029
b071af52 2030 if (!(READ_ONCE(n->nud_state) & NUD_CONNECTED)) {
f564f45c
CW		 2031 neigh_release(n);
2032 goto out;
2033 }
2034
3ad7a4b1 2035 f = vxlan_find_mac(vxlan, n->ha, vni);
f564f45c
CW		 2036 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
2037 /* bridge-local neighbor */
2038 neigh_release(n);
2039 goto out;
2040 }
2041
4b29dba9
DS		 2042 reply = vxlan_na_create(skb, n,
2043 !!(f ? f->flags & NTF_ROUTER : 0));
2044
f564f45c 2045 neigh_release(n);
4b29dba9
DS		 2046
2047 if (reply == NULL)
2048 goto out;
2049
3d391f65 2050 if (netif_rx(reply) == NET_RX_DROP) {
4b29dba9 2051 dev->stats.rx_dropped++;
4095e0e1
NA		 2052 vxlan_vnifilter_count(vxlan, vni, NULL,
2053 VXLAN_VNI_STATS_RX_DROPS, 0);
2054 }
dc5321d7 2055 } else if (vxlan->cfg.flags & VXLAN_F_L3MISS) {
4b29dba9
DS		 2056 union vxlan_addr ipa = {
2057 .sin6.sin6_addr = msg->target,
a45e92a5 2058 .sin6.sin6_family = AF_INET6,
4b29dba9
DS		 2059 };
2060
f564f45c
CW		 2061 vxlan_ip_miss(dev, &ipa);
2062 }
2063
2064out:
85e8b032 2065 rcu_read_unlock();
f564f45c
CW		 2066 consume_skb(skb);
2067 return NETDEV_TX_OK;
2068}
2069#endif
2070
e4f67add
DS		 2071static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb)
2072{
2073 struct vxlan_dev *vxlan = netdev_priv(dev);
2074 struct neighbour *n;
e4f67add
DS		 2075
2076 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
2077 return false;
2078
2079 n = NULL;
2080 switch (ntohs(eth_hdr(skb)->h_proto)) {
2081 case ETH_P_IP:
e15a00aa
CW		 2082 {
2083 struct iphdr *pip;
2084
e4f67add
DS		 2085 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
2086 return false;
2087 pip = ip_hdr(skb);
2088 n = neigh_lookup(&arp_tbl, &pip->daddr, dev);
dc5321d7 2089 if (!n && (vxlan->cfg.flags & VXLAN_F_L3MISS)) {
e4c7ed41
CW		 2090 union vxlan_addr ipa = {
2091 .sin.sin_addr.s_addr = pip->daddr,
a45e92a5 2092 .sin.sin_family = AF_INET,
e4c7ed41
CW		 2093 };
2094
2095 vxlan_ip_miss(dev, &ipa);
2096 return false;
2097 }
2098
e4f67add 2099 break;
e15a00aa
CW		 2100 }
2101#if IS_ENABLED(CONFIG_IPV6)
2102 case ETH_P_IPV6:
2103 {
2104 struct ipv6hdr *pip6;
2105
2106 if (!pskb_may_pull(skb, sizeof(struct ipv6hdr)))
2107 return false;
2108 pip6 = ipv6_hdr(skb);
2109 n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev);
dc5321d7 2110 if (!n && (vxlan->cfg.flags & VXLAN_F_L3MISS)) {
e15a00aa
CW		 2111 union vxlan_addr ipa = {
2112 .sin6.sin6_addr = pip6->daddr,
a45e92a5 2113 .sin6.sin6_family = AF_INET6,
e15a00aa
CW		 2114 };
2115
2116 vxlan_ip_miss(dev, &ipa);
2117 return false;
2118 }
2119
2120 break;
2121 }
2122#endif
e4f67add
DS		 2123 default:
2124 return false;
2125 }
2126
2127 if (n) {
2128 bool diff;
2129
7367d0b5 2130 diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha);
e4f67add
DS		 2131 if (diff) {
2132 memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest,
2133 dev->addr_len);
2134 memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len);
2135 }
2136 neigh_release(n);
2137 return diff;
e4c7ed41
CW		 2138 }
2139
e4f67add
DS		 2140 return false;
2141}
2142
df5e87f1 2143static int vxlan_build_gpe_hdr(struct vxlanhdr *vxh, __be16 protocol)
e1e5314d
JB		 2144{
2145 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)vxh;
2146
2147 gpe->np_applied = 1;
fa20e0e3
JB		 2148 gpe->next_protocol = tun_p_from_eth_p(protocol);
2149 if (!gpe->next_protocol)
2150 return -EPFNOSUPPORT;
2151 return 0;
e1e5314d
JB		 2152}
2153
f491e56d
JB		 2154static int vxlan_build_skb(struct sk_buff *skb, struct dst_entry *dst,
2155 int iphdr_len, __be32 vni,
2156 struct vxlan_metadata *md, u32 vxflags,
b4ed5cad 2157 bool udp_sum)
e4c7ed41 2158{
e4c7ed41 2159 struct vxlanhdr *vxh;
e4c7ed41
CW		 2160 int min_headroom;
2161 int err;
dfd8645e 2162 int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL;
e1e5314d 2163 __be16 inner_protocol = htons(ETH_P_TEB);
dfd8645e 2164
af33c1ad 2165 if ((vxflags & VXLAN_F_REMCSUM_TX) &&
dfd8645e
TH		 2166 skb->ip_summed == CHECKSUM_PARTIAL) {
2167 int csum_start = skb_checksum_start_offset(skb);
2168
2169 if (csum_start <= VXLAN_MAX_REMCSUM_START &&
2170 !(csum_start & VXLAN_RCO_SHIFT_MASK) &&
2171 (skb->csum_offset == offsetof(struct udphdr, check) ||
b5708501 2172 skb->csum_offset == offsetof(struct tcphdr, check)))
dfd8645e 2173 type |= SKB_GSO_TUNNEL_REMCSUM;
dfd8645e 2174 }
e4c7ed41 2175
e4c7ed41 2176 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
4a4f86cc 2177 + VXLAN_HLEN + iphdr_len;
649c5b8b
PS		 2178
2179 /* Need space for new headers (invalidates iph ptr) */
2180 err = skb_cow_head(skb, min_headroom);
e1e5314d 2181 if (unlikely(err))
c46b7897 2182 return err;
649c5b8b 2183
aed069df
AD		 2184 err = iptunnel_handle_offloads(skb, type);
2185 if (err)
c46b7897 2186 return err;
b736a623 2187
d58ff351 2188 vxh = __skb_push(skb, sizeof(*vxh));
54bfd872
JB		 2189 vxh->vx_flags = VXLAN_HF_VNI;
2190 vxh->vx_vni = vxlan_vni_field(vni);
49560532 2191
dfd8645e 2192 if (type & SKB_GSO_TUNNEL_REMCSUM) {
54bfd872 2193 unsigned int start;
dfd8645e 2194
54bfd872
JB		 2195 start = skb_checksum_start_offset(skb) - sizeof(struct vxlanhdr);
2196 vxh->vx_vni |= vxlan_compute_rco(start, skb->csum_offset);
2197 vxh->vx_flags |= VXLAN_HF_RCO;
dfd8645e
TH		 2198
2199 if (!skb_is_gso(skb)) {
2200 skb->ip_summed = CHECKSUM_NONE;
2201 skb->encapsulation = 0;
2202 }
2203 }
2204
af33c1ad 2205 if (vxflags & VXLAN_F_GBP)
df5e87f1 2206 vxlan_build_gbp_hdr(vxh, md);
e1e5314d 2207 if (vxflags & VXLAN_F_GPE) {
df5e87f1 2208 err = vxlan_build_gpe_hdr(vxh, skb->protocol);
e1e5314d 2209 if (err < 0)
c46b7897 2210 return err;
e1e5314d
JB		 2211 inner_protocol = skb->protocol;
2212 }
3511494c 2213
e1e5314d 2214 skb_set_inner_protocol(skb, inner_protocol);
039f5062 2215 return 0;
49560532 2216}
49560532 2217
655c3de1 2218static struct rtable *vxlan_get_route(struct vxlan_dev *vxlan, struct net_device *dev,
2219 struct vxlan_sock *sock4,
1a8496ba 2220 struct sk_buff *skb, int oif, u8 tos,
4ecb1d83 2221 __be32 daddr, __be32 *saddr, __be16 dport, __be16 sport,
7e2fb8bc 2222 __u8 flow_flags, struct dst_cache *dst_cache,
db3c6139 2223 const struct ip_tunnel_info *info)
1a8496ba 2224{
db3c6139 2225 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
1a8496ba
JB		 2226 struct rtable *rt = NULL;
2227 struct flowi4 fl4;
2228
655c3de1 2229 if (!sock4)
2230 return ERR_PTR(-EIO);
2231
db3c6139
DB		 2232 if (tos && !info)
2233 use_cache = false;
2234 if (use_cache) {
0c1d70af
PA		 2235 rt = dst_cache_get_ip4(dst_cache, saddr);
2236 if (rt)
2237 return rt;
2238 }
2239
1a8496ba
JB		 2240 memset(&fl4, 0, sizeof(fl4));
2241 fl4.flowi4_oif = oif;
2242 fl4.flowi4_tos = RT_TOS(tos);
2243 fl4.flowi4_mark = skb->mark;
2244 fl4.flowi4_proto = IPPROTO_UDP;
2245 fl4.daddr = daddr;
272d96a5 2246 fl4.saddr = *saddr;
4ecb1d83
MP		 2247 fl4.fl4_dport = dport;
2248 fl4.fl4_sport = sport;
7e2fb8bc 2249 fl4.flowi4_flags = flow_flags;
1a8496ba
JB		 2250
2251 rt = ip_route_output_key(vxlan->net, &fl4);
478db1f1 2252 if (!IS_ERR(rt)) {
655c3de1 2253 if (rt->dst.dev == dev) {
2254 netdev_dbg(dev, "circular route to %pI4\n", &daddr);
2255 ip_rt_put(rt);
2256 return ERR_PTR(-ELOOP);
2257 }
2258
1a8496ba 2259 *saddr = fl4.saddr;
0c1d70af
PA		 2260 if (use_cache)
2261 dst_cache_set_ip4(dst_cache, &rt->dst, fl4.saddr);
655c3de1 2262 } else {
2263 netdev_dbg(dev, "no route to %pI4\n", &daddr);
2264 return ERR_PTR(-ENETUNREACH);
0c1d70af 2265 }
1a8496ba
JB		 2266 return rt;
2267}
2268
e5d4b29f
JB		 2269#if IS_ENABLED(CONFIG_IPV6)
2270static struct dst_entry *vxlan6_get_route(struct vxlan_dev *vxlan,
655c3de1 2271 struct net_device *dev,
03dc52a8 2272 struct vxlan_sock *sock6,
1400615d 2273 struct sk_buff *skb, int oif, u8 tos,
e7f70af1 2274 __be32 label,
e5d4b29f 2275 const struct in6_addr *daddr,
0c1d70af 2276 struct in6_addr *saddr,
4ecb1d83 2277 __be16 dport, __be16 sport,
db3c6139
DB		 2278 struct dst_cache *dst_cache,
2279 const struct ip_tunnel_info *info)
e5d4b29f 2280{
db3c6139 2281 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
e5d4b29f
JB		 2282 struct dst_entry *ndst;
2283 struct flowi6 fl6;
e5d4b29f 2284
c6fcc4fc 2285 if (!sock6)
2286 return ERR_PTR(-EIO);
2287
1400615d
DB		 2288 if (tos && !info)
2289 use_cache = false;
db3c6139 2290 if (use_cache) {
0c1d70af
PA		 2291 ndst = dst_cache_get_ip6(dst_cache, saddr);
2292 if (ndst)
2293 return ndst;
2294 }
2295
e5d4b29f
JB		 2296 memset(&fl6, 0, sizeof(fl6));
2297 fl6.flowi6_oif = oif;
2298 fl6.daddr = *daddr;
272d96a5 2299 fl6.saddr = *saddr;
e488d4f5 2300 fl6.flowlabel = ip6_make_flowinfo(tos, label);
e5d4b29f
JB		 2301 fl6.flowi6_mark = skb->mark;
2302 fl6.flowi6_proto = IPPROTO_UDP;
4ecb1d83
MP		 2303 fl6.fl6_dport = dport;
2304 fl6.fl6_sport = sport;
e5d4b29f 2305
6c8991f4
SD		 2306 ndst = ipv6_stub->ipv6_dst_lookup_flow(vxlan->net, sock6->sock->sk,
2307 &fl6, NULL);
a10b24b8 2308 if (IS_ERR(ndst)) {
655c3de1 2309 netdev_dbg(dev, "no route to %pI6\n", daddr);
2310 return ERR_PTR(-ENETUNREACH);
2311 }
2312
2313 if (unlikely(ndst->dev == dev)) {
2314 netdev_dbg(dev, "circular route to %pI6\n", daddr);
2315 dst_release(ndst);
2316 return ERR_PTR(-ELOOP);
2317 }
e5d4b29f
JB		 2318
2319 *saddr = fl6.saddr;
db3c6139 2320 if (use_cache)
0c1d70af 2321 dst_cache_set_ip6(dst_cache, ndst, saddr);
e5d4b29f
JB		 2322 return ndst;
2323}
2324#endif
2325
9dcc71e1
SS		 2326/* Bypass encapsulation if the destination is local */
2327static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan,
fc68c995
SB		 2328 struct vxlan_dev *dst_vxlan, __be32 vni,
2329 bool snoop)
9dcc71e1 2330{
e4c7ed41
CW		 2331 union vxlan_addr loopback;
2332 union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip;
4179cb5a 2333 struct net_device *dev;
ce6502a8 2334 int len = skb->len;
9dcc71e1
SS		 2335
2336 skb->pkt_type = PACKET_HOST;
2337 skb->encapsulation = 0;
2338 skb->dev = dst_vxlan->dev;
2339 __skb_pull(skb, skb_network_offset(skb));
2340
e4c7ed41
CW		 2341 if (remote_ip->sa.sa_family == AF_INET) {
2342 loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2343 loopback.sa.sa_family = AF_INET;
2344#if IS_ENABLED(CONFIG_IPV6)
2345 } else {
2346 loopback.sin6.sin6_addr = in6addr_loopback;
2347 loopback.sa.sa_family = AF_INET6;
2348#endif
2349 }
2350
4179cb5a
ED		 2351 rcu_read_lock();
2352 dev = skb->dev;
2353 if (unlikely(!(dev->flags & IFF_UP))) {
2354 kfree_skb(skb);
2355 goto drop;
2356 }
2357
fc68c995 2358 if ((dst_vxlan->cfg.flags & VXLAN_F_LEARN) && snoop)
4179cb5a 2359 vxlan_snoop(dev, &loopback, eth_hdr(skb)->h_source, 0, vni);
9dcc71e1 2360
3c0930b4 2361 dev_sw_netstats_tx_add(src_vxlan->dev, 1, len);
4095e0e1 2362 vxlan_vnifilter_count(src_vxlan, vni, NULL, VXLAN_VNI_STATS_TX, len);
9dcc71e1 2363
baebdf48 2364 if (__netif_rx(skb) == NET_RX_SUCCESS) {
3c0930b4 2365 dev_sw_netstats_rx_add(dst_vxlan->dev, len);
4095e0e1
NA		 2366 vxlan_vnifilter_count(dst_vxlan, vni, NULL, VXLAN_VNI_STATS_RX,
2367 len);
9dcc71e1 2368 } else {
4179cb5a 2369drop:
ce6502a8 2370 dev->stats.rx_dropped++;
4095e0e1
NA		 2371 vxlan_vnifilter_count(dst_vxlan, vni, NULL,
2372 VXLAN_VNI_STATS_RX_DROPS, 0);
9dcc71e1 2373 }
4179cb5a 2374 rcu_read_unlock();
9dcc71e1
SS		 2375}
2376
fee1fad7 2377static int encap_bypass_if_local(struct sk_buff *skb, struct net_device *dev,
49f810f0
MS		 2378 struct vxlan_dev *vxlan,
2379 union vxlan_addr *daddr,
2380 __be16 dst_port, int dst_ifindex, __be32 vni,
2381 struct dst_entry *dst,
fee1fad7 2382 u32 rt_flags)
2383{
2384#if IS_ENABLED(CONFIG_IPV6)
2385 /* IPv6 rt-flags are checked against RTF_LOCAL, but the value of
2386 * RTF_LOCAL is equal to RTCF_LOCAL. So to keep code simple
2387 * we can use RTCF_LOCAL which works for ipv4 and ipv6 route entry.
2388 */
2389 BUILD_BUG_ON(RTCF_LOCAL != RTF_LOCAL);
2390#endif
2391 /* Bypass encapsulation if the destination is local */
2392 if (rt_flags & RTCF_LOCAL &&
69474a8a
VN		 2393 !(rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST)) &&
2394 vxlan->cfg.flags & VXLAN_F_LOCALBYPASS) {
fee1fad7 2395 struct vxlan_dev *dst_vxlan;
2396
2397 dst_release(dst);
49f810f0 2398 dst_vxlan = vxlan_find_vni(vxlan->net, dst_ifindex, vni,
fee1fad7 2399 daddr->sa.sa_family, dst_port,
dc5321d7 2400 vxlan->cfg.flags);
fee1fad7 2401 if (!dst_vxlan) {
2402 dev->stats.tx_errors++;
4095e0e1
NA		 2403 vxlan_vnifilter_count(vxlan, vni, NULL,
2404 VXLAN_VNI_STATS_TX_ERRORS, 0);
fee1fad7 2405 kfree_skb(skb);
2406
2407 return -ENOENT;
2408 }
fc68c995 2409 vxlan_encap_bypass(skb, vxlan, dst_vxlan, vni, true);
fee1fad7 2410 return 1;
2411 }
2412
2413 return 0;
2414}
2415
6ab271aa
IS		 2416void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev,
2417 __be32 default_vni, struct vxlan_rdst *rdst, bool did_rsc)
d342894c 2418{
d71785ff 2419 struct dst_cache *dst_cache;
3093fbe7 2420 struct ip_tunnel_info *info;
d342894c 2421 struct vxlan_dev *vxlan = netdev_priv(dev);
0770b53b 2422 const struct iphdr *old_iph = ip_hdr(skb);
e4c7ed41 2423 union vxlan_addr *dst;
272d96a5 2424 union vxlan_addr remote_ip, local_ip;
ee122c79
TG		 2425 struct vxlan_metadata _md;
2426 struct vxlan_metadata *md = &_md;
4095e0e1 2427 unsigned int pkt_len = skb->len;
e4c7ed41 2428 __be16 src_port = 0, dst_port;
655c3de1 2429 struct dst_entry *ndst = NULL;
7e2fb8bc 2430 __u8 tos, ttl, flow_flags = 0;
49f810f0 2431 int ifindex;
0e6fbc5b 2432 int err;
dc5321d7 2433 u32 flags = vxlan->cfg.flags;
b4ed5cad 2434 bool udp_sum = false;
f491e56d 2435 bool xnet = !net_eq(vxlan->net, dev_net(vxlan->dev));
fba55a66
RP		 2436 __be32 vni = 0;
2437#if IS_ENABLED(CONFIG_IPV6)
2438 __be32 label;
2439#endif
d342894c 2440
61adedf3 2441 info = skb_tunnel_info(skb);
3093fbe7 2442
ee122c79 2443 if (rdst) {
0770b53b 2444 dst = &rdst->remote_ip;
2445 if (vxlan_addr_any(dst)) {
2446 if (did_rsc) {
2447 /* short-circuited back to local bridge */
fc68c995
SB		 2448 vxlan_encap_bypass(skb, vxlan, vxlan,
2449 default_vni, true);
0770b53b 2450 return;
2451 }
2452 goto drop;
2453 }
2454
0dfbdf41 2455 dst_port = rdst->remote_port ? rdst->remote_port : vxlan->cfg.dst_port;
3ad7a4b1 2456 vni = (rdst->remote_vni) ? : default_vni;
49f810f0 2457 ifindex = rdst->remote_ifindex;
1158632b 2458 local_ip = vxlan->cfg.saddr;
d71785ff 2459 dst_cache = &rdst->dst_cache;
0770b53b 2460 md->gbp = skb->mark;
72f6d71e
HL		 2461 if (flags & VXLAN_F_TTL_INHERIT) {
2462 ttl = ip_tunnel_get_ttl(old_iph, skb);
2463 } else {
2464 ttl = vxlan->cfg.ttl;
2465 if (!ttl && vxlan_addr_multicast(dst))
2466 ttl = 1;
2467 }
0770b53b 2468
2469 tos = vxlan->cfg.tos;
2470 if (tos == 1)
2471 tos = ip_tunnel_get_dsfield(old_iph, skb);
2472
2473 if (dst->sa.sa_family == AF_INET)
2474 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM_TX);
2475 else
2476 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM6_TX);
fba55a66 2477#if IS_ENABLED(CONFIG_IPV6)
0770b53b 2478 label = vxlan->cfg.label;
fba55a66 2479#endif
ee122c79 2480 } else {
435be28b
JK		 2481 if (!info) {
2482 WARN_ONCE(1, "%s: Missing encapsulation instructions\n",
2483 dev->name);
2484 goto drop;
2485 }
b1be00a6 2486 remote_ip.sa.sa_family = ip_tunnel_info_af(info);
272d96a5 2487 if (remote_ip.sa.sa_family == AF_INET) {
a725e514 2488 remote_ip.sin.sin_addr.s_addr = info->key.u.ipv4.dst;
272d96a5 2489 local_ip.sin.sin_addr.s_addr = info->key.u.ipv4.src;
2490 } else {
a725e514 2491 remote_ip.sin6.sin6_addr = info->key.u.ipv6.dst;
272d96a5 2492 local_ip.sin6.sin6_addr = info->key.u.ipv6.src;
2493 }
ee122c79 2494 dst = &remote_ip;
0770b53b 2495 dst_port = info->key.tp_dst ? : vxlan->cfg.dst_port;
7e2fb8bc 2496 flow_flags = info->key.flow_flags;
0770b53b 2497 vni = tunnel_id_to_key32(info->key.tun_id);
49f810f0 2498 ifindex = 0;
d71785ff 2499 dst_cache = &info->dst_cache;
eadf52cf
XL		 2500 if (info->key.tun_flags & TUNNEL_VXLAN_OPT) {
2501 if (info->options_len < sizeof(*md))
2502 goto drop;
0770b53b 2503 md = ip_tunnel_info_opts(info);
eadf52cf 2504 }
a725e514
JB		 2505 ttl = info->key.ttl;
2506 tos = info->key.tos;
fba55a66 2507#if IS_ENABLED(CONFIG_IPV6)
e7f70af1 2508 label = info->key.label;
fba55a66 2509#endif
b4ed5cad 2510 udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
a725e514 2511 }
0770b53b 2512 src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
2513 vxlan->cfg.port_max, true);
a725e514 2514
56de859e 2515 rcu_read_lock();
e4c7ed41 2516 if (dst->sa.sa_family == AF_INET) {
c6fcc4fc 2517 struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock);
c46b7897 2518 struct rtable *rt;
0770b53b 2519 __be16 df = 0;
c6fcc4fc 2520
aab8cc36
AB		 2521 if (!ifindex)
2522 ifindex = sock4->sock->sk->sk_bound_dev_if;
2523
49f810f0 2524 rt = vxlan_get_route(vxlan, dev, sock4, skb, ifindex, tos,
272d96a5 2525 dst->sin.sin_addr.s_addr,
1158632b 2526 &local_ip.sin.sin_addr.s_addr,
7e2fb8bc 2527 dst_port, src_port, flow_flags,
d71785ff 2528 dst_cache, info);
8ebd115b
DM		 2529 if (IS_ERR(rt)) {
2530 err = PTR_ERR(rt);
c46b7897 2531 goto tx_error;
8ebd115b 2532 }
e4c7ed41 2533
fee1fad7 2534 if (!info) {
b4d30697 2535 /* Bypass encapsulation if the destination is local */
fee1fad7 2536 err = encap_bypass_if_local(skb, dev, vxlan, dst,
49f810f0
MS		 2537 dst_port, ifindex, vni,
2538 &rt->dst, rt->rt_flags);
fee1fad7 2539 if (err)
56de859e 2540 goto out_unlock;
b4d30697
SB		 2541
2542 if (vxlan->cfg.df == VXLAN_DF_SET) {
2543 df = htons(IP_DF);
2544 } else if (vxlan->cfg.df == VXLAN_DF_INHERIT) {
2545 struct ethhdr *eth = eth_hdr(skb);
2546
2547 if (ntohs(eth->h_proto) == ETH_P_IPV6 ||
2548 (ntohs(eth->h_proto) == ETH_P_IP &&
2549 old_iph->frag_off & htons(IP_DF)))
2550 df = htons(IP_DF);
2551 }
fee1fad7 2552 } else if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT) {
6ceb31ca 2553 df = htons(IP_DF);
fee1fad7 2554 }
6ceb31ca 2555
c46b7897 2556 ndst = &rt->dst;
94d166c5 2557 err = skb_tunnel_check_pmtu(skb, ndst, vxlan_headroom(flags & VXLAN_F_GPE),
fc68c995
SB		 2558 netif_is_any_bridge_port(dev));
2559 if (err < 0) {
2560 goto tx_error;
2561 } else if (err) {
2562 if (info) {
30a93d2b 2563 struct ip_tunnel_info *unclone;
fc68c995
SB		 2564 struct in_addr src, dst;
2565
30a93d2b
AT		 2566 unclone = skb_tunnel_info_unclone(skb);
2567 if (unlikely(!unclone))
2568 goto tx_error;
2569
fc68c995
SB		 2570 src = remote_ip.sin.sin_addr;
2571 dst = local_ip.sin.sin_addr;
30a93d2b
AT		 2572 unclone->key.u.ipv4.src = src.s_addr;
2573 unclone->key.u.ipv4.dst = dst.s_addr;
fc68c995
SB		 2574 }
2575 vxlan_encap_bypass(skb, vxlan, vxlan, vni, false);
2576 dst_release(ndst);
2577 goto out_unlock;
2578 }
a93bf0ff 2579
a0dced17 2580 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
e4c7ed41 2581 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
c46b7897 2582 err = vxlan_build_skb(skb, ndst, sizeof(struct iphdr),
54bfd872 2583 vni, md, flags, udp_sum);
f491e56d 2584 if (err < 0)
c46b7897 2585 goto tx_error;
f491e56d 2586
1158632b 2587 udp_tunnel_xmit_skb(rt, sock4->sock->sk, skb, local_ip.sin.sin_addr.s_addr,
f491e56d
JB		 2588 dst->sin.sin_addr.s_addr, tos, ttl, df,
2589 src_port, dst_port, xnet, !udp_sum);
e4c7ed41
CW		 2590#if IS_ENABLED(CONFIG_IPV6)
2591 } else {
c6fcc4fc 2592 struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock);
e4c7ed41 2593
aab8cc36
AB		 2594 if (!ifindex)
2595 ifindex = sock6->sock->sk->sk_bound_dev_if;
2596
49f810f0 2597 ndst = vxlan6_get_route(vxlan, dev, sock6, skb, ifindex, tos,
272d96a5 2598 label, &dst->sin6.sin6_addr,
1158632b 2599 &local_ip.sin6.sin6_addr,
4ecb1d83 2600 dst_port, src_port,
db3c6139 2601 dst_cache, info);
e5d4b29f 2602 if (IS_ERR(ndst)) {
8ebd115b 2603 err = PTR_ERR(ndst);
c46b7897 2604 ndst = NULL;
9dcc71e1 2605 goto tx_error;
e4c7ed41 2606 }
655c3de1 2607
fee1fad7 2608 if (!info) {
2609 u32 rt6i_flags = ((struct rt6_info *)ndst)->rt6i_flags;
49560532 2610
fee1fad7 2611 err = encap_bypass_if_local(skb, dev, vxlan, dst,
49f810f0
MS		 2612 dst_port, ifindex, vni,
2613 ndst, rt6i_flags);
fee1fad7 2614 if (err)
56de859e 2615 goto out_unlock;
fee1fad7 2616 }
35e2d115 2617
94d166c5
JB		 2618 err = skb_tunnel_check_pmtu(skb, ndst,
2619 vxlan_headroom((flags & VXLAN_F_GPE) | VXLAN_F_IPV6),
fc68c995
SB		 2620 netif_is_any_bridge_port(dev));
2621 if (err < 0) {
2622 goto tx_error;
2623 } else if (err) {
2624 if (info) {
30a93d2b 2625 struct ip_tunnel_info *unclone;
fc68c995
SB		 2626 struct in6_addr src, dst;
2627
30a93d2b
AT		 2628 unclone = skb_tunnel_info_unclone(skb);
2629 if (unlikely(!unclone))
2630 goto tx_error;
2631
fc68c995
SB		 2632 src = remote_ip.sin6.sin6_addr;
2633 dst = local_ip.sin6.sin6_addr;
30a93d2b
AT		 2634 unclone->key.u.ipv6.src = src;
2635 unclone->key.u.ipv6.dst = dst;
fc68c995
SB		 2636 }
2637
2638 vxlan_encap_bypass(skb, vxlan, vxlan, vni, false);
2639 dst_release(ndst);
2640 goto out_unlock;
2641 }
a93bf0ff 2642
a0dced17 2643 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
e4c7ed41 2644 ttl = ttl ? : ip6_dst_hoplimit(ndst);
f491e56d
JB		 2645 skb_scrub_packet(skb, xnet);
2646 err = vxlan_build_skb(skb, ndst, sizeof(struct ipv6hdr),
54bfd872 2647 vni, md, flags, udp_sum);
c46b7897 2648 if (err < 0)
2649 goto tx_error;
2650
0770b53b 2651 udp_tunnel6_xmit_skb(ndst, sock6->sock->sk, skb, dev,
1158632b 2652 &local_ip.sin6.sin6_addr,
272d96a5 2653 &dst->sin6.sin6_addr, tos, ttl,
e7f70af1 2654 label, src_port, dst_port, !udp_sum);
e4c7ed41
CW		 2655#endif
2656 }
4095e0e1 2657 vxlan_vnifilter_count(vxlan, vni, NULL, VXLAN_VNI_STATS_TX, pkt_len);
56de859e
JK		 2658out_unlock:
2659 rcu_read_unlock();
4ad16930 2660 return;
d342894c 2661
2662drop:
2663 dev->stats.tx_dropped++;
4095e0e1 2664 vxlan_vnifilter_count(vxlan, vni, NULL, VXLAN_VNI_STATS_TX_DROPS, 0);
c46b7897 2665 dev_kfree_skb(skb);
2666 return;
d342894c 2667
2668tx_error:
56de859e 2669 rcu_read_unlock();
655c3de1 2670 if (err == -ELOOP)
2671 dev->stats.collisions++;
2672 else if (err == -ENETUNREACH)
2673 dev->stats.tx_carrier_errors++;
c46b7897 2674 dst_release(ndst);
d342894c 2675 dev->stats.tx_errors++;
4095e0e1 2676 vxlan_vnifilter_count(vxlan, vni, NULL, VXLAN_VNI_STATS_TX_ERRORS, 0);
c46b7897 2677 kfree_skb(skb);
d342894c 2678}
2679
1274e1cc
RP		 2680static void vxlan_xmit_nh(struct sk_buff *skb, struct net_device *dev,
2681 struct vxlan_fdb *f, __be32 vni, bool did_rsc)
2682{
2683 struct vxlan_rdst nh_rdst;
2684 struct nexthop *nh;
2685 bool do_xmit;
2686 u32 hash;
2687
2688 memset(&nh_rdst, 0, sizeof(struct vxlan_rdst));
2689 hash = skb_get_hash(skb);
2690
2691 rcu_read_lock();
2692 nh = rcu_dereference(f->nh);
2693 if (!nh) {
2694 rcu_read_unlock();
2695 goto drop;
2696 }
2697 do_xmit = vxlan_fdb_nh_path_select(nh, hash, &nh_rdst);
2698 rcu_read_unlock();
2699
2700 if (likely(do_xmit))
2701 vxlan_xmit_one(skb, dev, vni, &nh_rdst, did_rsc);
2702 else
2703 goto drop;
2704
2705 return;
2706
2707drop:
2708 dev->stats.tx_dropped++;
4095e0e1
NA		 2709 vxlan_vnifilter_count(netdev_priv(dev), vni, NULL,
2710 VXLAN_VNI_STATS_TX_DROPS, 0);
1274e1cc
RP		 2711 dev_kfree_skb(skb);
2712}
2713
d977e1c8
IS		 2714static netdev_tx_t vxlan_xmit_nhid(struct sk_buff *skb, struct net_device *dev,
2715 u32 nhid, __be32 vni)
2716{
2717 struct vxlan_dev *vxlan = netdev_priv(dev);
2718 struct vxlan_rdst nh_rdst;
2719 struct nexthop *nh;
2720 bool do_xmit;
2721 u32 hash;
2722
2723 memset(&nh_rdst, 0, sizeof(struct vxlan_rdst));
2724 hash = skb_get_hash(skb);
2725
2726 rcu_read_lock();
2727 nh = nexthop_find_by_id(dev_net(dev), nhid);
2728 if (unlikely(!nh || !nexthop_is_fdb(nh) || !nexthop_is_multipath(nh))) {
2729 rcu_read_unlock();
2730 goto drop;
2731 }
2732 do_xmit = vxlan_fdb_nh_path_select(nh, hash, &nh_rdst);
2733 rcu_read_unlock();
2734
2735 if (vxlan->cfg.saddr.sa.sa_family != nh_rdst.remote_ip.sa.sa_family)
2736 goto drop;
2737
2738 if (likely(do_xmit))
2739 vxlan_xmit_one(skb, dev, vni, &nh_rdst, false);
2740 else
2741 goto drop;
2742
2743 return NETDEV_TX_OK;
2744
2745drop:
2746 dev->stats.tx_dropped++;
2747 vxlan_vnifilter_count(netdev_priv(dev), vni, NULL,
2748 VXLAN_VNI_STATS_TX_DROPS, 0);
2749 dev_kfree_skb(skb);
2750 return NETDEV_TX_OK;
2751}
2752
6681712d
DS		 2753/* Transmit local packets over Vxlan
2754 *
2755 * Outer IP header inherits ECN and DF from inner header.
2756 * Outer UDP destination is the VXLAN assigned port.
2757 * source port is based on hash of flow
2758 */
2759static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev)
2760{
2761 struct vxlan_dev *vxlan = netdev_priv(dev);
8bff3685 2762 struct vxlan_rdst *rdst, *fdst = NULL;
3093fbe7 2763 const struct ip_tunnel_info *info;
6681712d 2764 bool did_rsc = false;
6681712d 2765 struct vxlan_fdb *f;
8bff3685 2766 struct ethhdr *eth;
3ad7a4b1 2767 __be32 vni = 0;
d977e1c8 2768 u32 nhid = 0;
6681712d 2769
61adedf3 2770 info = skb_tunnel_info(skb);
3093fbe7 2771
6681712d 2772 skb_reset_mac_header(skb);
6681712d 2773
dc5321d7 2774 if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) {
3ad7a4b1
RP		 2775 if (info && info->mode & IP_TUNNEL_INFO_BRIDGE &&
2776 info->mode & IP_TUNNEL_INFO_TX) {
2777 vni = tunnel_id_to_key32(info->key.tun_id);
d977e1c8 2778 nhid = info->key.nhid;
3ad7a4b1
RP		 2779 } else {
2780 if (info && info->mode & IP_TUNNEL_INFO_TX)
2781 vxlan_xmit_one(skb, dev, vni, NULL, false);
2782 else
2783 kfree_skb(skb);
2784 return NETDEV_TX_OK;
2785 }
47e5d1b0
JB		 2786 }
2787
dc5321d7 2788 if (vxlan->cfg.flags & VXLAN_F_PROXY) {
47e5d1b0 2789 eth = eth_hdr(skb);
f564f45c 2790 if (ntohs(eth->h_proto) == ETH_P_ARP)
3ad7a4b1 2791 return arp_reduce(dev, skb, vni);
f564f45c 2792#if IS_ENABLED(CONFIG_IPV6)
8bff3685
XL		 2793 else if (ntohs(eth->h_proto) == ETH_P_IPV6 &&
2794 pskb_may_pull(skb, sizeof(struct ipv6hdr) +
2795 sizeof(struct nd_msg)) &&
2796 ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) {
2797 struct nd_msg *m = (struct nd_msg *)(ipv6_hdr(skb) + 1);
2798
2799 if (m->icmph.icmp6_code == 0 &&
2800 m->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION)
f1fb08f6 2801 return neigh_reduce(dev, skb, vni);
f564f45c
CW		 2802 }
2803#endif
2804 }
6681712d 2805
d977e1c8
IS		 2806 if (nhid)
2807 return vxlan_xmit_nhid(skb, dev, nhid, vni);
2808
0f83e69f
IS		 2809 if (vxlan->cfg.flags & VXLAN_F_MDB) {
2810 struct vxlan_mdb_entry *mdb_entry;
2811
2812 rcu_read_lock();
2813 mdb_entry = vxlan_mdb_entry_skb_get(vxlan, skb, vni);
2814 if (mdb_entry) {
2815 netdev_tx_t ret;
2816
2817 ret = vxlan_mdb_xmit(vxlan, mdb_entry, skb);
2818 rcu_read_unlock();
2819 return ret;
2820 }
2821 rcu_read_unlock();
2822 }
2823
47e5d1b0 2824 eth = eth_hdr(skb);
3ad7a4b1 2825 f = vxlan_find_mac(vxlan, eth->h_dest, vni);
ae884082
DS		 2826 did_rsc = false;
2827
dc5321d7 2828 if (f && (f->flags & NTF_ROUTER) && (vxlan->cfg.flags & VXLAN_F_RSC) &&
e15a00aa
CW		 2829 (ntohs(eth->h_proto) == ETH_P_IP ||
2830 ntohs(eth->h_proto) == ETH_P_IPV6)) {
ae884082
DS		 2831 did_rsc = route_shortcircuit(dev, skb);
2832 if (did_rsc)
3ad7a4b1 2833 f = vxlan_find_mac(vxlan, eth->h_dest, vni);
ae884082
DS		 2834 }
2835
6681712d 2836 if (f == NULL) {
3ad7a4b1 2837 f = vxlan_find_mac(vxlan, all_zeros_mac, vni);
afbd8bae 2838 if (f == NULL) {
dc5321d7 2839 if ((vxlan->cfg.flags & VXLAN_F_L2MISS) &&
afbd8bae
MR		 2840 !is_multicast_ether_addr(eth->h_dest))
2841 vxlan_fdb_miss(vxlan, eth->h_dest);
2842
2843 dev->stats.tx_dropped++;
4095e0e1
NA		 2844 vxlan_vnifilter_count(vxlan, vni, NULL,
2845 VXLAN_VNI_STATS_TX_DROPS, 0);
8f646c92 2846 kfree_skb(skb);
afbd8bae
MR		 2847 return NETDEV_TX_OK;
2848 }
2849 }
6681712d 2850
1274e1cc
RP		 2851 if (rcu_access_pointer(f->nh)) {
2852 vxlan_xmit_nh(skb, dev, f,
2853 (vni ? : vxlan->default_dst.remote_vni), did_rsc);
2854 } else {
2855 list_for_each_entry_rcu(rdst, &f->remotes, list) {
2856 struct sk_buff *skb1;
6681712d 2857
1274e1cc
RP		 2858 if (!fdst) {
2859 fdst = rdst;
2860 continue;
2861 }
2862 skb1 = skb_clone(skb, GFP_ATOMIC);
2863 if (skb1)
2864 vxlan_xmit_one(skb1, dev, vni, rdst, did_rsc);
8f646c92 2865 }
1274e1cc
RP		 2866 if (fdst)
2867 vxlan_xmit_one(skb, dev, vni, fdst, did_rsc);
2868 else
2869 kfree_skb(skb);
6681712d
DS		 2870 }
2871
4ad16930 2872 return NETDEV_TX_OK;
6681712d
DS		 2873}
2874
d342894c 2875/* Walk the forwarding table and purge stale entries */
df7e828c 2876static void vxlan_cleanup(struct timer_list *t)
d342894c 2877{
df7e828c 2878 struct vxlan_dev *vxlan = from_timer(vxlan, t, age_timer);
d342894c 2879 unsigned long next_timer = jiffies + FDB_AGE_INTERVAL;
2880 unsigned int h;
2881
2882 if (!netif_running(vxlan->dev))
2883 return;
2884
d342894c 2885 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2886 struct hlist_node *p, *n;
14e1d0fa 2887
fe1e0713 2888 spin_lock(&vxlan->hash_lock[h]);
d342894c 2889 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2890 struct vxlan_fdb *f
2891 = container_of(p, struct vxlan_fdb, hlist);
2892 unsigned long timeout;
2893
efb5f68f 2894 if (f->state & (NUD_PERMANENT | NUD_NOARP))
d342894c 2895 continue;
2896
def499c9
RP		 2897 if (f->flags & NTF_EXT_LEARNED)
2898 continue;
2899
0dfbdf41 2900 timeout = f->used + vxlan->cfg.age_interval * HZ;
d342894c 2901 if (time_before_eq(timeout, jiffies)) {
2902 netdev_dbg(vxlan->dev,
2903 "garbage collect %pM\n",
2904 f->eth_addr);
2905 f->state = NUD_STALE;
0e6160f3 2906 vxlan_fdb_destroy(vxlan, f, true, true);
d342894c 2907 } else if (time_before(timeout, next_timer))
2908 next_timer = timeout;
2909 }
fe1e0713 2910 spin_unlock(&vxlan->hash_lock[h]);
d342894c 2911 }
d342894c 2912
2913 mod_timer(&vxlan->age_timer, next_timer);
2914}
2915
a53cb29b
MB		 2916static void vxlan_vs_del_dev(struct vxlan_dev *vxlan)
2917{
2918 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
2919
2920 spin_lock(&vn->sock_lock);
69e76661
JB		 2921 hlist_del_init_rcu(&vxlan->hlist4.hlist);
2922#if IS_ENABLED(CONFIG_IPV6)
2923 hlist_del_init_rcu(&vxlan->hlist6.hlist);
2924#endif
a53cb29b
MB		 2925 spin_unlock(&vn->sock_lock);
2926}
2927
69e76661
JB		 2928static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan,
2929 struct vxlan_dev_node *node)
9c2e24e1 2930{
56ef9c90 2931 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
54bfd872 2932 __be32 vni = vxlan->default_dst.remote_vni;
9c2e24e1 2933
69e76661 2934 node->vxlan = vxlan;
56ef9c90 2935 spin_lock(&vn->sock_lock);
69e76661 2936 hlist_add_head_rcu(&node->hlist, vni_head(vs, vni));
56ef9c90 2937 spin_unlock(&vn->sock_lock);
9c2e24e1
PS		 2938}
2939
d342894c 2940/* Setup stats when device is created */
2941static int vxlan_init(struct net_device *dev)
2942{
384d91c2
TY		 2943 struct vxlan_dev *vxlan = netdev_priv(dev);
2944 int err;
2945
f9c4bb0b
RP		 2946 if (vxlan->cfg.flags & VXLAN_F_VNIFILTER)
2947 vxlan_vnigroup_init(vxlan);
2948
1c213bd2 2949 dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
06bf6294
IS		 2950 if (!dev->tstats) {
2951 err = -ENOMEM;
2952 goto err_vnigroup_uninit;
2953 }
d342894c 2954
384d91c2 2955 err = gro_cells_init(&vxlan->gro_cells, dev);
06bf6294
IS		 2956 if (err)
2957 goto err_free_percpu;
384d91c2 2958
a3a48de5
IS		 2959 err = vxlan_mdb_init(vxlan);
2960 if (err)
2961 goto err_gro_cells_destroy;
2962
d342894c 2963 return 0;
06bf6294 2964
a3a48de5
IS		 2965err_gro_cells_destroy:
2966 gro_cells_destroy(&vxlan->gro_cells);
06bf6294
IS		 2967err_free_percpu:
2968 free_percpu(dev->tstats);
2969err_vnigroup_uninit:
2970 if (vxlan->cfg.flags & VXLAN_F_VNIFILTER)
2971 vxlan_vnigroup_uninit(vxlan);
2972 return err;
d342894c 2973}
2974
3ad7a4b1 2975static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan, __be32 vni)
afbd8bae
MR		 2976{
2977 struct vxlan_fdb *f;
fe1e0713 2978 u32 hash_index = fdb_head_index(vxlan, all_zeros_mac, vni);
afbd8bae 2979
fe1e0713 2980 spin_lock_bh(&vxlan->hash_lock[hash_index]);
3ad7a4b1 2981 f = __vxlan_find_mac(vxlan, all_zeros_mac, vni);
afbd8bae 2982 if (f)
0e6160f3 2983 vxlan_fdb_destroy(vxlan, f, true, true);
fe1e0713 2984 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
afbd8bae
MR		 2985}
2986
ebf4063e
SH		 2987static void vxlan_uninit(struct net_device *dev)
2988{
2989 struct vxlan_dev *vxlan = netdev_priv(dev);
ebf4063e 2990
a3a48de5
IS		 2991 vxlan_mdb_fini(vxlan);
2992
f9c4bb0b
RP		 2993 if (vxlan->cfg.flags & VXLAN_F_VNIFILTER)
2994 vxlan_vnigroup_uninit(vxlan);
2995
ad6c9986
SB		 2996 gro_cells_destroy(&vxlan->gro_cells);
2997
3ad7a4b1 2998 vxlan_fdb_delete_default(vxlan, vxlan->cfg.vni);
afbd8bae 2999
ebf4063e
SH		 3000 free_percpu(dev->tstats);
3001}
3002
d342894c 3003/* Start ageing timer and join group when device is brought up */
3004static int vxlan_open(struct net_device *dev)
3005{
3006 struct vxlan_dev *vxlan = netdev_priv(dev);
205f356d 3007 int ret;
56ef9c90 3008
205f356d
JB		 3009 ret = vxlan_sock_add(vxlan);
3010 if (ret < 0)
3011 return ret;
d342894c 3012
f9c4bb0b
RP		 3013 ret = vxlan_multicast_join(vxlan);
3014 if (ret) {
3015 vxlan_sock_release(vxlan);
3016 return ret;
d342894c 3017 }
3018
0dfbdf41 3019 if (vxlan->cfg.age_interval)
d342894c 3020 mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL);
3021
56ef9c90 3022 return ret;
d342894c 3023}
3024
3025/* Purge the forwarding table */
8b3f9337 3026static void vxlan_flush(struct vxlan_dev *vxlan, bool do_all)
d342894c 3027{
31fec5aa 3028 unsigned int h;
d342894c 3029
d342894c 3030 for (h = 0; h < FDB_HASH_SIZE; ++h) {
3031 struct hlist_node *p, *n;
fe1e0713
L		 3032
3033 spin_lock_bh(&vxlan->hash_lock[h]);
d342894c 3034 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
3035 struct vxlan_fdb *f
3036 = container_of(p, struct vxlan_fdb, hlist);
8b3f9337
RP		 3037 if (!do_all && (f->state & (NUD_PERMANENT | NUD_NOARP)))
3038 continue;
afbd8bae 3039 /* the all_zeros_mac entry is deleted at vxlan_uninit */
fda2ec62
TY		 3040 if (is_zero_ether_addr(f->eth_addr) &&
3041 f->vni == vxlan->cfg.vni)
3042 continue;
3043 vxlan_fdb_destroy(vxlan, f, true, true);
d342894c 3044 }
fe1e0713 3045 spin_unlock_bh(&vxlan->hash_lock[h]);
d342894c 3046 }
d342894c 3047}
3048
3049/* Cleanup timer and forwarding table on shutdown */
3050static int vxlan_stop(struct net_device *dev)
3051{
3052 struct vxlan_dev *vxlan = netdev_priv(dev);
3053
f9c4bb0b 3054 vxlan_multicast_leave(vxlan);
d342894c 3055
3056 del_timer_sync(&vxlan->age_timer);
3057
8b3f9337 3058 vxlan_flush(vxlan, false);
205f356d 3059 vxlan_sock_release(vxlan);
d342894c 3060
e58bc864 3061 return 0;
d342894c 3062}
3063
d342894c 3064/* Stub, nothing needs to be done. */
3065static void vxlan_set_multicast_list(struct net_device *dev)
3066{
3067}
3068
91572088 3069static int vxlan_change_mtu(struct net_device *dev, int new_mtu)
345010b5 3070{
91572088
JW		 3071 struct vxlan_dev *vxlan = netdev_priv(dev);
3072 struct vxlan_rdst *dst = &vxlan->default_dst;
3073 struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
3074 dst->remote_ifindex);
345010b5 3075
91572088
JW		 3076 /* This check is different than dev->max_mtu, because it looks at
3077 * the lowerdev->mtu, rather than the static dev->max_mtu
3078 */
3079 if (lowerdev) {
94d166c5 3080 int max_mtu = lowerdev->mtu - vxlan_headroom(vxlan->cfg.flags);
91572088 3081 if (new_mtu > max_mtu)
72564b59 3082 return -EINVAL;
72564b59
DW		 3083 }
3084
345010b5
DB		 3085 dev->mtu = new_mtu;
3086 return 0;
3087}
3088
fc4099f1
PS		 3089static int vxlan_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
3090{
3091 struct vxlan_dev *vxlan = netdev_priv(dev);
3092 struct ip_tunnel_info *info = skb_tunnel_info(skb);
3093 __be16 sport, dport;
3094
3095 sport = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
3096 vxlan->cfg.port_max, true);
3097 dport = info->key.tp_dst ? : vxlan->cfg.dst_port;
3098
239e944f 3099 if (ip_tunnel_info_af(info) == AF_INET) {
c6fcc4fc 3100 struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock);
1a8496ba
JB		 3101 struct rtable *rt;
3102
655c3de1 3103 rt = vxlan_get_route(vxlan, dev, sock4, skb, 0, info->key.tos,
1a8496ba 3104 info->key.u.ipv4.dst,
22f0708a 3105 &info->key.u.ipv4.src, dport, sport,
7e2fb8bc
PC		 3106 info->key.flow_flags, &info->dst_cache,
3107 info);
1a8496ba
JB		 3108 if (IS_ERR(rt))
3109 return PTR_ERR(rt);
3110 ip_rt_put(rt);
239e944f
JB		 3111 } else {
3112#if IS_ENABLED(CONFIG_IPV6)
03dc52a8 3113 struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock);
239e944f
JB		 3114 struct dst_entry *ndst;
3115
655c3de1 3116 ndst = vxlan6_get_route(vxlan, dev, sock6, skb, 0, info->key.tos,
e7f70af1 3117 info->key.label, &info->key.u.ipv6.dst,
22f0708a
PA		 3118 &info->key.u.ipv6.src, dport, sport,
3119 &info->dst_cache, info);
239e944f
JB		 3120 if (IS_ERR(ndst))
3121 return PTR_ERR(ndst);
3122 dst_release(ndst);
239e944f
JB		 3123#else /* !CONFIG_IPV6 */
3124 return -EPFNOSUPPORT;
3125#endif
3126 }
1a8496ba
JB		 3127 info->key.tp_src = sport;
3128 info->key.tp_dst = dport;
239e944f 3129 return 0;
fc4099f1
PS		 3130}
3131
0c867c9b 3132static const struct net_device_ops vxlan_netdev_ether_ops = {
d342894c 3133 .ndo_init = vxlan_init,
ebf4063e 3134 .ndo_uninit = vxlan_uninit,
d342894c 3135 .ndo_open = vxlan_open,
3136 .ndo_stop = vxlan_stop,
3137 .ndo_start_xmit = vxlan_xmit,
b220a4a7 3138 .ndo_get_stats64 = dev_get_tstats64,
d342894c 3139 .ndo_set_rx_mode = vxlan_set_multicast_list,
345010b5 3140 .ndo_change_mtu = vxlan_change_mtu,
d342894c 3141 .ndo_validate_addr = eth_validate_addr,
3142 .ndo_set_mac_address = eth_mac_addr,
3143 .ndo_fdb_add = vxlan_fdb_add,
3144 .ndo_fdb_del = vxlan_fdb_delete,
3145 .ndo_fdb_dump = vxlan_fdb_dump,
474c3c89 3146 .ndo_fdb_get = vxlan_fdb_get,
08f876a7
IS		 3147 .ndo_mdb_add = vxlan_mdb_add,
3148 .ndo_mdb_del = vxlan_mdb_del,
3149 .ndo_mdb_dump = vxlan_mdb_dump,
fc4099f1 3150 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
d342894c 3151};
3152
e1e5314d
JB		 3153static const struct net_device_ops vxlan_netdev_raw_ops = {
3154 .ndo_init = vxlan_init,
3155 .ndo_uninit = vxlan_uninit,
3156 .ndo_open = vxlan_open,
3157 .ndo_stop = vxlan_stop,
3158 .ndo_start_xmit = vxlan_xmit,
b220a4a7 3159 .ndo_get_stats64 = dev_get_tstats64,
e1e5314d
JB		 3160 .ndo_change_mtu = vxlan_change_mtu,
3161 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
3162};
3163
d342894c 3164/* Info for udev, that this is a virtual tunnel endpoint */
3165static struct device_type vxlan_type = {
3166 .name = "vxlan",
3167};
3168
e5de25dc 3169/* Calls the ndo_udp_tunnel_add of the caller in order to
35e42379 3170 * supply the listening VXLAN udp ports. Callers are expected
e5de25dc 3171 * to implement the ndo_udp_tunnel_add.
53cf5275 3172 */
2d2b13fc 3173static void vxlan_offload_rx_ports(struct net_device *dev, bool push)
53cf5275
JG		 3174{
3175 struct vxlan_sock *vs;
3176 struct net *net = dev_net(dev);
3177 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
35e42379 3178 unsigned int i;
53cf5275
JG		 3179
3180 spin_lock(&vn->sock_lock);
3181 for (i = 0; i < PORT_HASH_SIZE; ++i) {
2d2b13fc
SD		 3182 hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) {
3183 unsigned short type;
3184
3185 if (vs->flags & VXLAN_F_GPE)
3186 type = UDP_TUNNEL_TYPE_VXLAN_GPE;
3187 else
3188 type = UDP_TUNNEL_TYPE_VXLAN;
3189
3190 if (push)
3191 udp_tunnel_push_rx_port(dev, vs->sock, type);
3192 else
3193 udp_tunnel_drop_rx_port(dev, vs->sock, type);
3194 }
53cf5275
JG		 3195 }
3196 spin_unlock(&vn->sock_lock);
3197}
53cf5275 3198
d342894c 3199/* Initialize the device structure. */
3200static void vxlan_setup(struct net_device *dev)
3201{
3202 struct vxlan_dev *vxlan = netdev_priv(dev);
31fec5aa 3203 unsigned int h;
d342894c 3204
65226ef8
JB		 3205 eth_hw_addr_random(dev);
3206 ether_setup(dev);
3207
cf124db5 3208 dev->needs_free_netdev = true;
d342894c 3209 SET_NETDEV_DEVTYPE(dev, &vxlan_type);
3210
d342894c 3211 dev->features |= NETIF_F_LLTX;
cb2c5711 3212 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_FRAGLIST;
0afb1666 3213 dev->features |= NETIF_F_RXCSUM;
05c0db08 3214 dev->features |= NETIF_F_GSO_SOFTWARE;
0afb1666 3215
1eaa8178 3216 dev->vlan_features = dev->features;
cb2c5711
XL		 3217 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_FRAGLIST;
3218 dev->hw_features |= NETIF_F_RXCSUM;
05c0db08 3219 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
02875878 3220 netif_keep_dst(dev);
2106efda 3221 dev->priv_flags |= IFF_NO_QUEUE | IFF_CHANGE_PROTO_DOWN;
d342894c 3222
a985343b
MS		 3223 /* MTU range: 68 - 65535 */
3224 dev->min_mtu = ETH_MIN_MTU;
3225 dev->max_mtu = ETH_MAX_MTU;
3226
553675fb 3227 INIT_LIST_HEAD(&vxlan->next);
d342894c 3228
df7e828c 3229 timer_setup(&vxlan->age_timer, vxlan_cleanup, TIMER_DEFERRABLE);
d342894c 3230
3231 vxlan->dev = dev;
3232
fe1e0713
L		 3233 for (h = 0; h < FDB_HASH_SIZE; ++h) {
3234 spin_lock_init(&vxlan->hash_lock[h]);
d342894c 3235 INIT_HLIST_HEAD(&vxlan->fdb_head[h]);
fe1e0713 3236 }
d342894c 3237}
3238
0c867c9b
JB		 3239static void vxlan_ether_setup(struct net_device *dev)
3240{
0c867c9b
JB		 3241 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
3242 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
3243 dev->netdev_ops = &vxlan_netdev_ether_ops;
3244}
3245
e1e5314d
JB		 3246static void vxlan_raw_setup(struct net_device *dev)
3247{
65226ef8 3248 dev->header_ops = NULL;
e1e5314d
JB		 3249 dev->type = ARPHRD_NONE;
3250 dev->hard_header_len = 0;
3251 dev->addr_len = 0;
e1e5314d
JB		 3252 dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
3253 dev->netdev_ops = &vxlan_netdev_raw_ops;
3254}
3255
d342894c 3256static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = {
69474a8a 3257 [IFLA_VXLAN_UNSPEC] = { .strict_start_type = IFLA_VXLAN_LOCALBYPASS },
d342894c 3258 [IFLA_VXLAN_ID] = { .type = NLA_U32 },
c593642c 3259 [IFLA_VXLAN_GROUP] = { .len = sizeof_field(struct iphdr, daddr) },
e4c7ed41 3260 [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) },
d342894c 3261 [IFLA_VXLAN_LINK] = { .type = NLA_U32 },
c593642c 3262 [IFLA_VXLAN_LOCAL] = { .len = sizeof_field(struct iphdr, saddr) },
e4c7ed41 3263 [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) },
d342894c 3264 [IFLA_VXLAN_TOS] = { .type = NLA_U8 },
3265 [IFLA_VXLAN_TTL] = { .type = NLA_U8 },
e7f70af1 3266 [IFLA_VXLAN_LABEL] = { .type = NLA_U32 },
d342894c 3267 [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 },
3268 [IFLA_VXLAN_AGEING] = { .type = NLA_U32 },
3269 [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 },
05f47d69 3270 [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) },
e4f67add
DS		 3271 [IFLA_VXLAN_PROXY] = { .type = NLA_U8 },
3272 [IFLA_VXLAN_RSC] = { .type = NLA_U8 },
3273 [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 },
3274 [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 },
f8a9b1bc 3275 [IFLA_VXLAN_COLLECT_METADATA] = { .type = NLA_U8 },
823aa873 3276 [IFLA_VXLAN_PORT] = { .type = NLA_U16 },
5c91ae08
TH		 3277 [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 },
3278 [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 },
3279 [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 },
dfd8645e
TH		 3280 [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 },
3281 [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 },
3511494c 3282 [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, },
e1e5314d 3283 [IFLA_VXLAN_GPE] = { .type = NLA_FLAG, },
0ace2ca8 3284 [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG },
72f6d71e 3285 [IFLA_VXLAN_TTL_INHERIT] = { .type = NLA_FLAG },
b4d30697 3286 [IFLA_VXLAN_DF] = { .type = NLA_U8 },
f9c4bb0b 3287 [IFLA_VXLAN_VNIFILTER] = { .type = NLA_U8 },
69474a8a 3288 [IFLA_VXLAN_LOCALBYPASS] = NLA_POLICY_MAX(NLA_U8, 1),
d342894c 3289};
3290
a8b8a889
MS		 3291static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[],
3292 struct netlink_ext_ack *extack)
d342894c 3293{
3294 if (tb[IFLA_ADDRESS]) {
3295 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) {
653ef6a3
GM		 3296 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_ADDRESS],
3297 "Provided link layer address is not Ethernet");
d342894c 3298 return -EINVAL;
3299 }
3300
3301 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) {
653ef6a3
GM		 3302 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_ADDRESS],
3303 "Provided Ethernet address is not unicast");
d342894c 3304 return -EADDRNOTAVAIL;
3305 }
3306 }
3307
a985343b 3308 if (tb[IFLA_MTU]) {
019b13ae 3309 u32 mtu = nla_get_u32(tb[IFLA_MTU]);
a985343b 3310
653ef6a3
GM		 3311 if (mtu < ETH_MIN_MTU || mtu > ETH_MAX_MTU) {
3312 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_MTU],
3313 "MTU must be between 68 and 65535");
a985343b 3314 return -EINVAL;
653ef6a3 3315 }
a985343b
MS		 3316 }
3317
653ef6a3
GM		 3318 if (!data) {
3319 NL_SET_ERR_MSG(extack,
3320 "Required attributes not provided to perform the operation");
d342894c 3321 return -EINVAL;
653ef6a3 3322 }
d342894c 3323
3324 if (data[IFLA_VXLAN_ID]) {
a985343b
MS		 3325 u32 id = nla_get_u32(data[IFLA_VXLAN_ID]);
3326
653ef6a3 3327 if (id >= VXLAN_N_VID) {
cc8e7c69 3328 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_ID],
653ef6a3 3329 "VXLAN ID must be lower than 16777216");
d342894c 3330 return -ERANGE;
653ef6a3 3331 }
d342894c 3332 }
3333
05f47d69 3334 if (data[IFLA_VXLAN_PORT_RANGE]) {
3335 const struct ifla_vxlan_port_range *p
3336 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
3337
3338 if (ntohs(p->high) < ntohs(p->low)) {
cc8e7c69 3339 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_PORT_RANGE],
653ef6a3 3340 "Invalid source port range");
05f47d69 3341 return -EINVAL;
3342 }
3343 }
3344
b4d30697
SB		 3345 if (data[IFLA_VXLAN_DF]) {
3346 enum ifla_vxlan_df df = nla_get_u8(data[IFLA_VXLAN_DF]);
3347
3348 if (df < 0 || df > VXLAN_DF_MAX) {
cc8e7c69 3349 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_DF],
b4d30697
SB		 3350 "Invalid DF attribute");
3351 return -EINVAL;
3352 }
3353 }
3354
d342894c 3355 return 0;
3356}
3357
1b13c97f
YB		 3358static void vxlan_get_drvinfo(struct net_device *netdev,
3359 struct ethtool_drvinfo *drvinfo)
3360{
fb3ceec1
WS		 3361 strscpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version));
3362 strscpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver));
1b13c97f
YB		 3363}
3364
36fe3a61
MS		 3365static int vxlan_get_link_ksettings(struct net_device *dev,
3366 struct ethtool_link_ksettings *cmd)
3367{
3368 struct vxlan_dev *vxlan = netdev_priv(dev);
3369 struct vxlan_rdst *dst = &vxlan->default_dst;
3370 struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
3371 dst->remote_ifindex);
3372
3373 if (!lowerdev) {
3374 cmd->base.duplex = DUPLEX_UNKNOWN;
3375 cmd->base.port = PORT_OTHER;
3376 cmd->base.speed = SPEED_UNKNOWN;
3377
3378 return 0;
3379 }
3380
3381 return __ethtool_get_link_ksettings(lowerdev, cmd);
3382}
3383
1b13c97f 3384static const struct ethtool_ops vxlan_ethtool_ops = {
36fe3a61
MS		 3385 .get_drvinfo = vxlan_get_drvinfo,
3386 .get_link = ethtool_op_get_link,
3387 .get_link_ksettings = vxlan_get_link_ksettings,
1b13c97f
YB		 3388};
3389
3ee64f39 3390static struct socket *vxlan_create_sock(struct net *net, bool ipv6,
aab8cc36 3391 __be16 port, u32 flags, int ifindex)
553675fb 3392{
e4c7ed41 3393 struct socket *sock;
3ee64f39
TH		 3394 struct udp_port_cfg udp_conf;
3395 int err;
e4c7ed41 3396
3ee64f39 3397 memset(&udp_conf, 0, sizeof(udp_conf));
e4c7ed41 3398
3ee64f39
TH		 3399 if (ipv6) {
3400 udp_conf.family = AF_INET6;
3ee64f39 3401 udp_conf.use_udp6_rx_checksums =
3dc2b6a8 3402 !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX);
a43a9ef6 3403 udp_conf.ipv6_v6only = 1;
3ee64f39
TH		 3404 } else {
3405 udp_conf.family = AF_INET;
e4c7ed41 3406 }
e4c7ed41 3407
3ee64f39 3408 udp_conf.local_udp_port = port;
aab8cc36 3409 udp_conf.bind_ifindex = ifindex;
553675fb 3410
3ee64f39
TH		 3411 /* Open UDP socket */
3412 err = udp_sock_create(net, &udp_conf, &sock);
3413 if (err < 0)
3414 return ERR_PTR(err);
e4c7ed41 3415
d18931a9 3416 udp_allow_gso(sock->sk);
39deb2c7 3417 return sock;
e4c7ed41
CW		 3418}
3419
3420/* Create new listen socket if needed */
b1be00a6 3421static struct vxlan_sock *vxlan_socket_create(struct net *net, bool ipv6,
aab8cc36
AB		 3422 __be16 port, u32 flags,
3423 int ifindex)
e4c7ed41
CW		 3424{
3425 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
3426 struct vxlan_sock *vs;
3427 struct socket *sock;
e4c7ed41 3428 unsigned int h;
acbf74a7 3429 struct udp_tunnel_sock_cfg tunnel_cfg;
e4c7ed41 3430
dc01e7d3 3431 vs = kzalloc(sizeof(*vs), GFP_KERNEL);
e4c7ed41
CW		 3432 if (!vs)
3433 return ERR_PTR(-ENOMEM);
3434
3435 for (h = 0; h < VNI_HASH_SIZE; ++h)
3436 INIT_HLIST_HEAD(&vs->vni_list[h]);
3437
aab8cc36 3438 sock = vxlan_create_sock(net, ipv6, port, flags, ifindex);
39deb2c7 3439 if (IS_ERR(sock)) {
553675fb 3440 kfree(vs);
e50fddc8 3441 return ERR_CAST(sock);
553675fb 3442 }
e4c7ed41
CW		 3443
3444 vs->sock = sock;
66af846f 3445 refcount_set(&vs->refcnt, 1);
af33c1ad 3446 vs->flags = (flags & VXLAN_F_RCV_FLAGS);
553675fb 3447
9c2e24e1
PS		 3448 spin_lock(&vn->sock_lock);
3449 hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
e7b3db5e 3450 udp_tunnel_notify_add_rx_port(sock,
b9adcd69
AD		 3451 (vs->flags & VXLAN_F_GPE) ?
3452 UDP_TUNNEL_TYPE_VXLAN_GPE :
e7b3db5e 3453 UDP_TUNNEL_TYPE_VXLAN);
9c2e24e1 3454 spin_unlock(&vn->sock_lock);
553675fb 3455
3456 /* Mark socket as an encapsulation socket. */
5602c48c 3457 memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
acbf74a7
AZ		 3458 tunnel_cfg.sk_user_data = vs;
3459 tunnel_cfg.encap_type = 1;
f2d1968e 3460 tunnel_cfg.encap_rcv = vxlan_rcv;
c3a43b9f 3461 tunnel_cfg.encap_err_lookup = vxlan_err_lookup;
acbf74a7 3462 tunnel_cfg.encap_destroy = NULL;
b0b672c4
JB		 3463 if (vs->flags & VXLAN_F_GPE) {
3464 tunnel_cfg.gro_receive = vxlan_gpe_gro_receive;
3465 tunnel_cfg.gro_complete = vxlan_gpe_gro_complete;
3466 } else {
3467 tunnel_cfg.gro_receive = vxlan_gro_receive;
3468 tunnel_cfg.gro_complete = vxlan_gro_complete;
3469 }
acbf74a7
AZ		 3470
3471 setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
e4c7ed41 3472
9c2e24e1
PS		 3473 return vs;
3474}
3475
b1be00a6 3476static int __vxlan_sock_add(struct vxlan_dev *vxlan, bool ipv6)
9c2e24e1 3477{
205f356d 3478 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
f9c4bb0b 3479 bool metadata = vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA;
205f356d 3480 struct vxlan_sock *vs = NULL;
69e76661 3481 struct vxlan_dev_node *node;
aab8cc36
AB		 3482 int l3mdev_index = 0;
3483
3484 if (vxlan->cfg.remote_ifindex)
3485 l3mdev_index = l3mdev_master_upper_ifindex_by_index(
3486 vxlan->net, vxlan->cfg.remote_ifindex);
9c2e24e1 3487
205f356d 3488 if (!vxlan->cfg.no_share) {
56ef9c90 3489 spin_lock(&vn->sock_lock);
205f356d 3490 vs = vxlan_find_sock(vxlan->net, ipv6 ? AF_INET6 : AF_INET,
aab8cc36
AB		 3491 vxlan->cfg.dst_port, vxlan->cfg.flags,
3492 l3mdev_index);
66af846f 3493 if (vs && !refcount_inc_not_zero(&vs->refcnt)) {
56ef9c90 3494 spin_unlock(&vn->sock_lock);
205f356d 3495 return -EBUSY;
56ef9c90
MRL		 3496 }
3497 spin_unlock(&vn->sock_lock);
3498 }
205f356d 3499 if (!vs)
b1be00a6 3500 vs = vxlan_socket_create(vxlan->net, ipv6,
aab8cc36
AB		 3501 vxlan->cfg.dst_port, vxlan->cfg.flags,
3502 l3mdev_index);
205f356d
JB		 3503 if (IS_ERR(vs))
3504 return PTR_ERR(vs);
b1be00a6 3505#if IS_ENABLED(CONFIG_IPV6)
69e76661 3506 if (ipv6) {
c6fcc4fc 3507 rcu_assign_pointer(vxlan->vn6_sock, vs);
69e76661
JB		 3508 node = &vxlan->hlist6;
3509 } else
b1be00a6 3510#endif
69e76661 3511 {
c6fcc4fc 3512 rcu_assign_pointer(vxlan->vn4_sock, vs);
69e76661
JB		 3513 node = &vxlan->hlist4;
3514 }
f9c4bb0b
RP		 3515
3516 if (metadata && (vxlan->cfg.flags & VXLAN_F_VNIFILTER))
3517 vxlan_vs_add_vnigrp(vxlan, vs, ipv6);
3518 else
3519 vxlan_vs_add_dev(vs, vxlan, node);
3520
205f356d 3521 return 0;
553675fb 3522}
3523
b1be00a6
JB		 3524static int vxlan_sock_add(struct vxlan_dev *vxlan)
3525{
dc5321d7
MS		 3526 bool metadata = vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA;
3527 bool ipv6 = vxlan->cfg.flags & VXLAN_F_IPV6 || metadata;
d074bf96 3528 bool ipv4 = !ipv6 || metadata;
b1be00a6
JB		 3529 int ret = 0;
3530
c6fcc4fc 3531 RCU_INIT_POINTER(vxlan->vn4_sock, NULL);
b1be00a6 3532#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 3533 RCU_INIT_POINTER(vxlan->vn6_sock, NULL);
d074bf96 3534 if (ipv6) {
b1be00a6 3535 ret = __vxlan_sock_add(vxlan, true);
d074bf96
JB		 3536 if (ret < 0 && ret != -EAFNOSUPPORT)
3537 ipv4 = false;
3538 }
b1be00a6 3539#endif
d074bf96 3540 if (ipv4)
b1be00a6
JB		 3541 ret = __vxlan_sock_add(vxlan, false);
3542 if (ret < 0)
3543 vxlan_sock_release(vxlan);
3544 return ret;
3545}
3546
f9c4bb0b
RP		 3547int vxlan_vni_in_use(struct net *src_net, struct vxlan_dev *vxlan,
3548 struct vxlan_config *conf, __be32 vni)
efe0f94b
RP		 3549{
3550 struct vxlan_net *vn = net_generic(src_net, vxlan_net_id);
3551 struct vxlan_dev *tmp;
3552
3553 list_for_each_entry(tmp, &vn->vxlan_list, next) {
3554 if (tmp == vxlan)
3555 continue;
f9c4bb0b
RP		 3556 if (tmp->cfg.flags & VXLAN_F_VNIFILTER) {
3557 if (!vxlan_vnifilter_lookup(tmp, vni))
3558 continue;
3559 } else if (tmp->cfg.vni != vni) {
efe0f94b 3560 continue;
f9c4bb0b 3561 }
efe0f94b
RP		 3562 if (tmp->cfg.dst_port != conf->dst_port)
3563 continue;
3564 if ((tmp->cfg.flags & (VXLAN_F_RCV_FLAGS | VXLAN_F_IPV6)) !=
3565 (conf->flags & (VXLAN_F_RCV_FLAGS | VXLAN_F_IPV6)))
3566 continue;
3567
3568 if ((conf->flags & VXLAN_F_IPV6_LINKLOCAL) &&
3569 tmp->cfg.remote_ifindex != conf->remote_ifindex)
3570 continue;
3571
3572 return -EEXIST;
3573 }
3574
3575 return 0;
3576}
3577
a985343b
MS		 3578static int vxlan_config_validate(struct net *src_net, struct vxlan_config *conf,
3579 struct net_device **lower,
653ef6a3
GM		 3580 struct vxlan_dev *old,
3581 struct netlink_ext_ack *extack)
d342894c 3582{
e4c7ed41 3583 bool use_ipv6 = false;
d342894c 3584
a985343b
MS		 3585 if (conf->flags & VXLAN_F_GPE) {
3586 /* For now, allow GPE only together with
3587 * COLLECT_METADATA. This can be relaxed later; in such
3588 * case, the other side of the PtP link will have to be
3589 * provided.
3590 */
3591 if ((conf->flags & ~VXLAN_F_ALLOWED_GPE) ||
3592 !(conf->flags & VXLAN_F_COLLECT_METADATA)) {
653ef6a3
GM		 3593 NL_SET_ERR_MSG(extack,
3594 "VXLAN GPE does not support this combination of attributes");
a985343b 3595 return -EINVAL;
3555621d 3596 }
e1e5314d 3597 }
0c867c9b 3598
ce44a4ae
MS		 3599 if (!conf->remote_ip.sa.sa_family && !conf->saddr.sa.sa_family) {
3600 /* Unless IPv6 is explicitly requested, assume IPv4 */
a985343b 3601 conf->remote_ip.sa.sa_family = AF_INET;
ce44a4ae
MS		 3602 conf->saddr.sa.sa_family = AF_INET;
3603 } else if (!conf->remote_ip.sa.sa_family) {
3604 conf->remote_ip.sa.sa_family = conf->saddr.sa.sa_family;
3605 } else if (!conf->saddr.sa.sa_family) {
3606 conf->saddr.sa.sa_family = conf->remote_ip.sa.sa_family;
3607 }
3608
653ef6a3
GM		 3609 if (conf->saddr.sa.sa_family != conf->remote_ip.sa.sa_family) {
3610 NL_SET_ERR_MSG(extack,
3611 "Local and remote address must be from the same family");
ce44a4ae 3612 return -EINVAL;
653ef6a3 3613 }
e4c7ed41 3614
653ef6a3
GM		 3615 if (vxlan_addr_multicast(&conf->saddr)) {
3616 NL_SET_ERR_MSG(extack, "Local address cannot be multicast");
0f22a3c6 3617 return -EINVAL;
653ef6a3 3618 }
0f22a3c6 3619
ce44a4ae 3620 if (conf->saddr.sa.sa_family == AF_INET6) {
653ef6a3
GM		 3621 if (!IS_ENABLED(CONFIG_IPV6)) {
3622 NL_SET_ERR_MSG(extack,
3623 "IPv6 support not enabled in the kernel");
057ba29b 3624 return -EPFNOSUPPORT;
653ef6a3 3625 }
e4c7ed41 3626 use_ipv6 = true;
a985343b 3627 conf->flags |= VXLAN_F_IPV6;
0f22a3c6
MS		 3628
3629 if (!(conf->flags & VXLAN_F_COLLECT_METADATA)) {
3630 int local_type =
3631 ipv6_addr_type(&conf->saddr.sin6.sin6_addr);
3632 int remote_type =
3633 ipv6_addr_type(&conf->remote_ip.sin6.sin6_addr);
3634
3635 if (local_type & IPV6_ADDR_LINKLOCAL) {
3636 if (!(remote_type & IPV6_ADDR_LINKLOCAL) &&
653ef6a3
GM		 3637 (remote_type != IPV6_ADDR_ANY)) {
3638 NL_SET_ERR_MSG(extack,
3639 "Invalid combination of local and remote address scopes");
0f22a3c6 3640 return -EINVAL;
653ef6a3 3641 }
0f22a3c6
MS		 3642
3643 conf->flags |= VXLAN_F_IPV6_LINKLOCAL;
3644 } else {
3645 if (remote_type ==
653ef6a3
GM		 3646 (IPV6_ADDR_UNICAST | IPV6_ADDR_LINKLOCAL)) {
3647 NL_SET_ERR_MSG(extack,
3648 "Invalid combination of local and remote address scopes");
0f22a3c6 3649 return -EINVAL;
653ef6a3 3650 }
0f22a3c6
MS		 3651
3652 conf->flags &= ~VXLAN_F_IPV6_LINKLOCAL;
3653 }
3654 }
057ba29b 3655 }
d342894c 3656
653ef6a3
GM		 3657 if (conf->label && !use_ipv6) {
3658 NL_SET_ERR_MSG(extack,
3659 "Label attribute only applies to IPv6 VXLAN devices");
e7f70af1 3660 return -EINVAL;
653ef6a3 3661 }
e7f70af1 3662
a985343b
MS		 3663 if (conf->remote_ifindex) {
3664 struct net_device *lowerdev;
34e02aa1 3665
a985343b 3666 lowerdev = __dev_get_by_index(src_net, conf->remote_ifindex);
653ef6a3
GM		 3667 if (!lowerdev) {
3668 NL_SET_ERR_MSG(extack,
3669 "Invalid local interface, device not found");
34e02aa1 3670 return -ENODEV;
653ef6a3 3671 }
d342894c 3672
e4c7ed41
CW		 3673#if IS_ENABLED(CONFIG_IPV6)
3674 if (use_ipv6) {
3675 struct inet6_dev *idev = __in6_dev_get(lowerdev);
6fadbdd6 3676
653ef6a3
GM		 3677 if (idev && idev->cnf.disable_ipv6) {
3678 NL_SET_ERR_MSG(extack,
3679 "IPv6 support disabled by administrator");
e4c7ed41 3680 return -EPERM;
653ef6a3 3681 }
e4c7ed41
CW		 3682 }
3683#endif
3684
a985343b
MS		 3685 *lower = lowerdev;
3686 } else {
653ef6a3
GM		 3687 if (vxlan_addr_multicast(&conf->remote_ip)) {
3688 NL_SET_ERR_MSG(extack,
3689 "Local interface required for multicast remote destination");
3690
a985343b 3691 return -EINVAL;
653ef6a3 3692 }
1ba56fb4 3693
0f22a3c6 3694#if IS_ENABLED(CONFIG_IPV6)
653ef6a3
GM		 3695 if (conf->flags & VXLAN_F_IPV6_LINKLOCAL) {
3696 NL_SET_ERR_MSG(extack,
3697 "Local interface required for link-local local/remote addresses");
0f22a3c6 3698 return -EINVAL;
653ef6a3 3699 }
0f22a3c6
MS		 3700#endif
3701
a985343b 3702 *lower = NULL;
9dc2ad10 3703 }
d342894c 3704
a985343b
MS		 3705 if (!conf->dst_port) {
3706 if (conf->flags & VXLAN_F_GPE)
ed618bd8 3707 conf->dst_port = htons(IANA_VXLAN_GPE_UDP_PORT);
a985343b
MS		 3708 else
3709 conf->dst_port = htons(vxlan_port);
d6acfeb1
FM		 3710 }
3711
a985343b
MS		 3712 if (!conf->age_interval)
3713 conf->age_interval = FDB_AGE_DEFAULT;
91572088 3714
efe0f94b 3715 if (vxlan_vni_in_use(src_net, old, conf, conf->vni)) {
653ef6a3
GM		 3716 NL_SET_ERR_MSG(extack,
3717 "A VXLAN device with the specified VNI already exists");
49f810f0 3718 return -EEXIST;
a985343b 3719 }
91572088 3720
a985343b
MS		 3721 return 0;
3722}
3723
3724static void vxlan_config_apply(struct net_device *dev,
3725 struct vxlan_config *conf,
889ce937
SD		 3726 struct net_device *lowerdev,
3727 struct net *src_net,
3728 bool changelink)
a985343b
MS		 3729{
3730 struct vxlan_dev *vxlan = netdev_priv(dev);
3731 struct vxlan_rdst *dst = &vxlan->default_dst;
3732 unsigned short needed_headroom = ETH_HLEN;
a985343b 3733 int max_mtu = ETH_MAX_MTU;
94d166c5 3734 u32 flags = conf->flags;
a985343b
MS		 3735
3736 if (!changelink) {
94d166c5 3737 if (flags & VXLAN_F_GPE)
a985343b
MS		 3738 vxlan_raw_setup(dev);
3739 else
3740 vxlan_ether_setup(dev);
3741
3742 if (conf->mtu)
3743 dev->mtu = conf->mtu;
889ce937
SD		 3744
3745 vxlan->net = src_net;
a985343b
MS		 3746 }
3747
3748 dst->remote_vni = conf->vni;
3749
3750 memcpy(&dst->remote_ip, &conf->remote_ip, sizeof(conf->remote_ip));
3751
3752 if (lowerdev) {
3753 dst->remote_ifindex = conf->remote_ifindex;
3754
6df6398f 3755 netif_inherit_tso_max(dev, lowerdev);
91572088 3756
a985343b 3757 needed_headroom = lowerdev->hard_header_len;
0a35dc41 3758 needed_headroom += lowerdev->needed_headroom;
91572088 3759
a5e74021 3760 dev->needed_tailroom = lowerdev->needed_tailroom;
91572088 3761
94d166c5 3762 max_mtu = lowerdev->mtu - vxlan_headroom(flags);
f870c1ff
AK		 3763 if (max_mtu < ETH_MIN_MTU)
3764 max_mtu = ETH_MIN_MTU;
3765
3766 if (!changelink && !conf->mtu)
3767 dev->mtu = max_mtu;
7e059158
DW		 3768 }
3769
a985343b
MS		 3770 if (dev->mtu > max_mtu)
3771 dev->mtu = max_mtu;
3772
94d166c5
JB		 3773 if (flags & VXLAN_F_COLLECT_METADATA)
3774 flags |= VXLAN_F_IPV6;
3775 needed_headroom += vxlan_headroom(flags);
b1be00a6
JB		 3776 dev->needed_headroom = needed_headroom;
3777
0dfbdf41 3778 memcpy(&vxlan->cfg, conf, sizeof(*conf));
a985343b 3779}
0dfbdf41 3780
a985343b 3781static int vxlan_dev_configure(struct net *src_net, struct net_device *dev,
653ef6a3
GM		 3782 struct vxlan_config *conf, bool changelink,
3783 struct netlink_ext_ack *extack)
a985343b
MS		 3784{
3785 struct vxlan_dev *vxlan = netdev_priv(dev);
3786 struct net_device *lowerdev;
3787 int ret;
0dfbdf41 3788
653ef6a3 3789 ret = vxlan_config_validate(src_net, conf, &lowerdev, vxlan, extack);
a985343b
MS		 3790 if (ret)
3791 return ret;
8bcdc4f3 3792
889ce937 3793 vxlan_config_apply(dev, conf, lowerdev, src_net, changelink);
0dfbdf41 3794
0dfbdf41
TG		 3795 return 0;
3796}
3797
c80498e3 3798static int __vxlan_dev_create(struct net *net, struct net_device *dev,
653ef6a3
GM		 3799 struct vxlan_config *conf,
3800 struct netlink_ext_ack *extack)
c80498e3
ND		 3801{
3802 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
3803 struct vxlan_dev *vxlan = netdev_priv(dev);
0ce1822c 3804 struct net_device *remote_dev = NULL;
0241b836 3805 struct vxlan_fdb *f = NULL;
6db92468 3806 bool unregister = false;
0ce1822c 3807 struct vxlan_rdst *dst;
c80498e3
ND		 3808 int err;
3809
0ce1822c 3810 dst = &vxlan->default_dst;
653ef6a3 3811 err = vxlan_dev_configure(net, dev, conf, false, extack);
c80498e3
ND		 3812 if (err)
3813 return err;
3814
3815 dev->ethtool_ops = &vxlan_ethtool_ops;
3816
3817 /* create an fdb entry for a valid default destination */
0ce1822c 3818 if (!vxlan_addr_any(&dst->remote_ip)) {
0241b836 3819 err = vxlan_fdb_create(vxlan, all_zeros_mac,
0ce1822c 3820 &dst->remote_ip,
c80498e3 3821 NUD_REACHABLE | NUD_PERMANENT,
c80498e3 3822 vxlan->cfg.dst_port,
0ce1822c
TY		 3823 dst->remote_vni,
3824 dst->remote_vni,
3825 dst->remote_ifindex,
1274e1cc 3826 NTF_SELF, 0, &f, extack);
c80498e3
ND		 3827 if (err)
3828 return err;
3829 }
3830
3831 err = register_netdevice(dev);
0241b836
RP		 3832 if (err)
3833 goto errout;
6db92468 3834 unregister = true;
0241b836 3835
0ce1822c
TY		 3836 if (dst->remote_ifindex) {
3837 remote_dev = __dev_get_by_index(net, dst->remote_ifindex);
832e0979
ZC		 3838 if (!remote_dev) {
3839 err = -ENODEV;
0ce1822c 3840 goto errout;
832e0979 3841 }
0ce1822c
TY		 3842
3843 err = netdev_upper_dev_link(remote_dev, dev, extack);
3844 if (err)
3845 goto errout;
3846 }
3847
1d997f10 3848 err = rtnl_configure_link(dev, NULL, 0, NULL);
2eabcb8a 3849 if (err < 0)
0ce1822c 3850 goto unlink;
c80498e3 3851
61f46fe8 3852 if (f) {
0ce1822c 3853 vxlan_fdb_insert(vxlan, all_zeros_mac, dst->remote_vni, f);
7c31e54a
TY		 3854
3855 /* notify default fdb entry */
61f46fe8 3856 err = vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f),
4c59b7d1 3857 RTM_NEWNEIGH, true, extack);
7c31e54a
TY		 3858 if (err) {
3859 vxlan_fdb_destroy(vxlan, f, false, false);
0ce1822c
TY		 3860 if (remote_dev)
3861 netdev_upper_dev_unlink(remote_dev, dev);
7c31e54a
TY		 3862 goto unregister;
3863 }
61f46fe8 3864 }
0241b836 3865
c80498e3 3866 list_add(&vxlan->next, &vn->vxlan_list);
0ce1822c
TY		 3867 if (remote_dev)
3868 dst->remote_dev = remote_dev;
c80498e3 3869 return 0;
0ce1822c
TY		 3870unlink:
3871 if (remote_dev)
3872 netdev_upper_dev_unlink(remote_dev, dev);
0241b836 3873errout:
6db92468
PM		 3874 /* unregister_netdevice() destroys the default FDB entry with deletion
3875 * notification. But the addition notification was not sent yet, so
3876 * destroy the entry by hand here.
3877 */
0241b836 3878 if (f)
7c31e54a
TY		 3879 __vxlan_fdb_free(f);
3880unregister:
6db92468
PM		 3881 if (unregister)
3882 unregister_netdevice(dev);
0241b836 3883 return err;
c80498e3
ND		 3884}
3885
70fb0828
RP		 3886/* Set/clear flags based on attribute */
3887static int vxlan_nl2flag(struct vxlan_config *conf, struct nlattr *tb[],
3888 int attrtype, unsigned long mask, bool changelink,
3889 bool changelink_supported,
3890 struct netlink_ext_ack *extack)
3891{
3892 unsigned long flags;
3893
3894 if (!tb[attrtype])
3895 return 0;
3896
3897 if (changelink && !changelink_supported) {
3898 vxlan_flag_attr_error(attrtype, extack);
3899 return -EOPNOTSUPP;
3900 }
3901
3902 if (vxlan_policy[attrtype].type == NLA_FLAG)
3903 flags = conf->flags | mask;
3904 else if (nla_get_u8(tb[attrtype]))
3905 flags = conf->flags | mask;
3906 else
3907 flags = conf->flags & ~mask;
3908
3909 conf->flags = flags;
3910
3911 return 0;
3912}
3913
8bcdc4f3
RP		 3914static int vxlan_nl2conf(struct nlattr *tb[], struct nlattr *data[],
3915 struct net_device *dev, struct vxlan_config *conf,
70fb0828 3916 bool changelink, struct netlink_ext_ack *extack)
0dfbdf41 3917{
8bcdc4f3 3918 struct vxlan_dev *vxlan = netdev_priv(dev);
70fb0828 3919 int err = 0;
0dfbdf41 3920
8bcdc4f3 3921 memset(conf, 0, sizeof(*conf));
e277de5f 3922
8bcdc4f3
RP		 3923 /* if changelink operation, start with old existing cfg */
3924 if (changelink)
3925 memcpy(conf, &vxlan->cfg, sizeof(*conf));
3926
3927 if (data[IFLA_VXLAN_ID]) {
3928 __be32 vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID]));
3929
70fb0828
RP		 3930 if (changelink && (vni != conf->vni)) {
3931 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_ID], "Cannot change VNI");
8bcdc4f3 3932 return -EOPNOTSUPP;
70fb0828 3933 }
8bcdc4f3
RP		 3934 conf->vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID]));
3935 }
0dfbdf41
TG		 3936
3937 if (data[IFLA_VXLAN_GROUP]) {
70fb0828
RP		 3938 if (changelink && (conf->remote_ip.sa.sa_family != AF_INET)) {
3939 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_GROUP], "New group address family does not match old group");
ce44a4ae 3940 return -EOPNOTSUPP;
70fb0828 3941 }
ce44a4ae 3942
8bcdc4f3 3943 conf->remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]);
ce44a4ae 3944 conf->remote_ip.sa.sa_family = AF_INET;
0dfbdf41 3945 } else if (data[IFLA_VXLAN_GROUP6]) {
70fb0828
RP		 3946 if (!IS_ENABLED(CONFIG_IPV6)) {
3947 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_GROUP6], "IPv6 support not enabled in the kernel");
0dfbdf41 3948 return -EPFNOSUPPORT;
70fb0828 3949 }
0dfbdf41 3950
70fb0828
RP		 3951 if (changelink && (conf->remote_ip.sa.sa_family != AF_INET6)) {
3952 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_GROUP6], "New group address family does not match old group");
ce44a4ae 3953 return -EOPNOTSUPP;
70fb0828 3954 }
ce44a4ae 3955
8bcdc4f3
RP		 3956 conf->remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]);
3957 conf->remote_ip.sa.sa_family = AF_INET6;
0dfbdf41
TG		 3958 }
3959
3960 if (data[IFLA_VXLAN_LOCAL]) {
70fb0828
RP		 3961 if (changelink && (conf->saddr.sa.sa_family != AF_INET)) {
3962 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LOCAL], "New local address family does not match old");
ce44a4ae 3963 return -EOPNOTSUPP;
70fb0828 3964 }
ce44a4ae 3965
8bcdc4f3
RP		 3966 conf->saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]);
3967 conf->saddr.sa.sa_family = AF_INET;
0dfbdf41 3968 } else if (data[IFLA_VXLAN_LOCAL6]) {
70fb0828
RP		 3969 if (!IS_ENABLED(CONFIG_IPV6)) {
3970 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LOCAL6], "IPv6 support not enabled in the kernel");
0dfbdf41 3971 return -EPFNOSUPPORT;
70fb0828 3972 }
0dfbdf41 3973
70fb0828
RP		 3974 if (changelink && (conf->saddr.sa.sa_family != AF_INET6)) {
3975 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LOCAL6], "New local address family does not match old");
ce44a4ae 3976 return -EOPNOTSUPP;
70fb0828 3977 }
ce44a4ae 3978
0dfbdf41 3979 /* TODO: respect scope id */
8bcdc4f3
RP		 3980 conf->saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]);
3981 conf->saddr.sa.sa_family = AF_INET6;
0dfbdf41
TG		 3982 }
3983
3984 if (data[IFLA_VXLAN_LINK])
8bcdc4f3 3985 conf->remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]);
0dfbdf41 3986
d342894c 3987 if (data[IFLA_VXLAN_TOS])
8bcdc4f3 3988 conf->tos = nla_get_u8(data[IFLA_VXLAN_TOS]);
d342894c 3989
afb97186 3990 if (data[IFLA_VXLAN_TTL])
8bcdc4f3 3991 conf->ttl = nla_get_u8(data[IFLA_VXLAN_TTL]);
afb97186 3992
72f6d71e 3993 if (data[IFLA_VXLAN_TTL_INHERIT]) {
70fb0828
RP		 3994 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_TTL_INHERIT,
3995 VXLAN_F_TTL_INHERIT, changelink, false,
3996 extack);
3997 if (err)
3998 return err;
3999
72f6d71e
HL		 4000 }
4001
e7f70af1 4002 if (data[IFLA_VXLAN_LABEL])
8bcdc4f3 4003 conf->label = nla_get_be32(data[IFLA_VXLAN_LABEL]) &
e7f70af1
DB		 4004 IPV6_FLOWLABEL_MASK;
4005
8bcdc4f3 4006 if (data[IFLA_VXLAN_LEARNING]) {
70fb0828
RP		 4007 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_LEARNING,
4008 VXLAN_F_LEARN, changelink, true,
4009 extack);
4010 if (err)
4011 return err;
8bcdc4f3
RP		 4012 } else if (!changelink) {
4013 /* default to learn on a new device */
4014 conf->flags |= VXLAN_F_LEARN;
4015 }
d342894c 4016
40051c4d 4017 if (data[IFLA_VXLAN_AGEING])
8bcdc4f3 4018 conf->age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]);
d342894c 4019
8bcdc4f3 4020 if (data[IFLA_VXLAN_PROXY]) {
70fb0828
RP		 4021 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_PROXY,
4022 VXLAN_F_PROXY, changelink, false,
4023 extack);
4024 if (err)
4025 return err;
8bcdc4f3 4026 }
e4f67add 4027
8bcdc4f3 4028 if (data[IFLA_VXLAN_RSC]) {
70fb0828
RP		 4029 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_RSC,
4030 VXLAN_F_RSC, changelink, false,
4031 extack);
4032 if (err)
4033 return err;
8bcdc4f3 4034 }
e4f67add 4035
8bcdc4f3 4036 if (data[IFLA_VXLAN_L2MISS]) {
70fb0828
RP		 4037 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_L2MISS,
4038 VXLAN_F_L2MISS, changelink, false,
4039 extack);
4040 if (err)
4041 return err;
8bcdc4f3 4042 }
e4f67add 4043
8bcdc4f3 4044 if (data[IFLA_VXLAN_L3MISS]) {
70fb0828
RP		 4045 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_L3MISS,
4046 VXLAN_F_L3MISS, changelink, false,
4047 extack);
4048 if (err)
4049 return err;
8bcdc4f3 4050 }
e4f67add 4051
8bcdc4f3 4052 if (data[IFLA_VXLAN_LIMIT]) {
70fb0828
RP		 4053 if (changelink) {
4054 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LIMIT],
4055 "Cannot change limit");
8bcdc4f3 4056 return -EOPNOTSUPP;
70fb0828 4057 }
8bcdc4f3
RP		 4058 conf->addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]);
4059 }
d342894c 4060
8bcdc4f3 4061 if (data[IFLA_VXLAN_COLLECT_METADATA]) {
70fb0828
RP		 4062 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_COLLECT_METADATA,
4063 VXLAN_F_COLLECT_METADATA, changelink, false,
4064 extack);
4065 if (err)
4066 return err;
8bcdc4f3 4067 }
f8a9b1bc 4068
05f47d69 4069 if (data[IFLA_VXLAN_PORT_RANGE]) {
8bcdc4f3
RP		 4070 if (!changelink) {
4071 const struct ifla_vxlan_port_range *p
4072 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
4073 conf->port_min = ntohs(p->low);
4074 conf->port_max = ntohs(p->high);
4075 } else {
70fb0828
RP		 4076 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_PORT_RANGE],
4077 "Cannot change port range");
8bcdc4f3
RP		 4078 return -EOPNOTSUPP;
4079 }
05f47d69 4080 }
4081
8bcdc4f3 4082 if (data[IFLA_VXLAN_PORT]) {
70fb0828
RP		 4083 if (changelink) {
4084 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_PORT],
4085 "Cannot change port");
8bcdc4f3 4086 return -EOPNOTSUPP;
70fb0828 4087 }
8bcdc4f3
RP		 4088 conf->dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]);
4089 }
823aa873 4090
8bcdc4f3 4091 if (data[IFLA_VXLAN_UDP_CSUM]) {
70fb0828
RP		 4092 if (changelink) {
4093 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_UDP_CSUM],
4094 "Cannot change UDP_CSUM flag");
8bcdc4f3 4095 return -EOPNOTSUPP;
70fb0828 4096 }
8bcdc4f3
RP		 4097 if (!nla_get_u8(data[IFLA_VXLAN_UDP_CSUM]))
4098 conf->flags |= VXLAN_F_UDP_ZERO_CSUM_TX;
4099 }
359a0ea9 4100
69474a8a
VN		 4101 if (data[IFLA_VXLAN_LOCALBYPASS]) {
4102 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_LOCALBYPASS,
4103 VXLAN_F_LOCALBYPASS, changelink,
4104 true, extack);
4105 if (err)
4106 return err;
4107 } else if (!changelink) {
4108 /* default to local bypass on a new device */
4109 conf->flags |= VXLAN_F_LOCALBYPASS;
4110 }
4111
8bcdc4f3 4112 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX]) {
70fb0828
RP		 4113 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_UDP_ZERO_CSUM6_TX,
4114 VXLAN_F_UDP_ZERO_CSUM6_TX, changelink,
4115 false, extack);
4116 if (err)
4117 return err;
8bcdc4f3 4118 }
359a0ea9 4119
8bcdc4f3 4120 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX]) {
70fb0828
RP		 4121 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_UDP_ZERO_CSUM6_RX,
4122 VXLAN_F_UDP_ZERO_CSUM6_RX, changelink,
4123 false, extack);
4124 if (err)
4125 return err;
8bcdc4f3 4126 }
359a0ea9 4127
8bcdc4f3 4128 if (data[IFLA_VXLAN_REMCSUM_TX]) {
70fb0828
RP		 4129 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_REMCSUM_TX,
4130 VXLAN_F_REMCSUM_TX, changelink, false,
4131 extack);
4132 if (err)
4133 return err;
8bcdc4f3 4134 }
dfd8645e 4135
8bcdc4f3 4136 if (data[IFLA_VXLAN_REMCSUM_RX]) {
70fb0828
RP		 4137 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_REMCSUM_RX,
4138 VXLAN_F_REMCSUM_RX, changelink, false,
4139 extack);
4140 if (err)
4141 return err;
8bcdc4f3
RP		 4142 }
4143
4144 if (data[IFLA_VXLAN_GBP]) {
70fb0828
RP		 4145 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_GBP,
4146 VXLAN_F_GBP, changelink, false, extack);
4147 if (err)
4148 return err;
8bcdc4f3
RP		 4149 }
4150
4151 if (data[IFLA_VXLAN_GPE]) {
70fb0828
RP		 4152 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_GPE,
4153 VXLAN_F_GPE, changelink, false,
4154 extack);
4155 if (err)
4156 return err;
8bcdc4f3
RP		 4157 }
4158
4159 if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL]) {
70fb0828
RP		 4160 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_REMCSUM_NOPARTIAL,
4161 VXLAN_F_REMCSUM_NOPARTIAL, changelink,
4162 false, extack);
4163 if (err)
4164 return err;
8bcdc4f3
RP		 4165 }
4166
4167 if (tb[IFLA_MTU]) {
70fb0828
RP		 4168 if (changelink) {
4169 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_MTU],
4170 "Cannot change mtu");
8bcdc4f3 4171 return -EOPNOTSUPP;
70fb0828 4172 }
8bcdc4f3
RP		 4173 conf->mtu = nla_get_u32(tb[IFLA_MTU]);
4174 }
4175
b4d30697
SB		 4176 if (data[IFLA_VXLAN_DF])
4177 conf->df = nla_get_u8(data[IFLA_VXLAN_DF]);
4178
f9c4bb0b
RP		 4179 if (data[IFLA_VXLAN_VNIFILTER]) {
4180 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_VNIFILTER,
4181 VXLAN_F_VNIFILTER, changelink, false,
4182 extack);
4183 if (err)
4184 return err;
4185
4186 if ((conf->flags & VXLAN_F_VNIFILTER) &&
4187 !(conf->flags & VXLAN_F_COLLECT_METADATA)) {
4188 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_VNIFILTER],
4189 "vxlan vnifilter only valid in collect metadata mode");
4190 return -EINVAL;
4191 }
4192 }
4193
8bcdc4f3
RP		 4194 return 0;
4195}
4196
4197static int vxlan_newlink(struct net *src_net, struct net_device *dev,
7a3f4a18
MS		 4198 struct nlattr *tb[], struct nlattr *data[],
4199 struct netlink_ext_ack *extack)
8bcdc4f3 4200{
8bcdc4f3
RP		 4201 struct vxlan_config conf;
4202 int err;
4203
70fb0828 4204 err = vxlan_nl2conf(tb, data, dev, &conf, false, extack);
8bcdc4f3
RP		 4205 if (err)
4206 return err;
4207
653ef6a3 4208 return __vxlan_dev_create(src_net, dev, &conf, extack);
8bcdc4f3 4209}
dfd8645e 4210
8bcdc4f3 4211static int vxlan_changelink(struct net_device *dev, struct nlattr *tb[],
ad744b22
MS		 4212 struct nlattr *data[],
4213 struct netlink_ext_ack *extack)
8bcdc4f3
RP		 4214{
4215 struct vxlan_dev *vxlan = netdev_priv(dev);
8db9427d 4216 struct net_device *lowerdev;
8bcdc4f3 4217 struct vxlan_config conf;
0ce1822c 4218 struct vxlan_rdst *dst;
8bcdc4f3
RP		 4219 int err;
4220
0ce1822c 4221 dst = &vxlan->default_dst;
70fb0828 4222 err = vxlan_nl2conf(tb, data, dev, &conf, true, extack);
8bcdc4f3
RP		 4223 if (err)
4224 return err;
3511494c 4225
8db9427d
PM		 4226 err = vxlan_config_validate(vxlan->net, &conf, &lowerdev,
4227 vxlan, extack);
8bcdc4f3
RP		 4228 if (err)
4229 return err;
0ace2ca8 4230
c6761cf5
TY		 4231 if (dst->remote_dev == lowerdev)
4232 lowerdev = NULL;
4233
0ce1822c
TY		 4234 err = netdev_adjacent_change_prepare(dst->remote_dev, lowerdev, dev,
4235 extack);
4236 if (err)
4237 return err;
4238
8bcdc4f3 4239 /* handle default dst entry */
038a5a99 4240 if (!vxlan_addr_equal(&conf.remote_ip, &dst->remote_ip)) {
fe1e0713
L		 4241 u32 hash_index = fdb_head_index(vxlan, all_zeros_mac, conf.vni);
4242
4243 spin_lock_bh(&vxlan->hash_lock[hash_index]);
038a5a99 4244 if (!vxlan_addr_any(&conf.remote_ip)) {
ce5e098f 4245 err = vxlan_fdb_update(vxlan, all_zeros_mac,
038a5a99 4246 &conf.remote_ip,
8bcdc4f3 4247 NUD_REACHABLE | NUD_PERMANENT,
ce5e098f 4248 NLM_F_APPEND | NLM_F_CREATE,
8bcdc4f3 4249 vxlan->cfg.dst_port,
038a5a99
PM		 4250 conf.vni, conf.vni,
4251 conf.remote_ifindex,
1274e1cc 4252 NTF_SELF, 0, true, extack);
8bcdc4f3 4253 if (err) {
fe1e0713 4254 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
0ce1822c
TY		 4255 netdev_adjacent_change_abort(dst->remote_dev,
4256 lowerdev, dev);
8bcdc4f3
RP		 4257 return err;
4258 }
4259 }
1cdc98c2
PM		 4260 if (!vxlan_addr_any(&dst->remote_ip))
4261 __vxlan_fdb_delete(vxlan, all_zeros_mac,
4262 dst->remote_ip,
4263 vxlan->cfg.dst_port,
4264 dst->remote_vni,
4265 dst->remote_vni,
4266 dst->remote_ifindex,
4267 true);
fe1e0713 4268 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
f9c4bb0b
RP		 4269
4270 /* If vni filtering device, also update fdb entries of
4271 * all vnis that were using default remote ip
4272 */
4273 if (vxlan->cfg.flags & VXLAN_F_VNIFILTER) {
4274 err = vxlan_vnilist_update_group(vxlan, &dst->remote_ip,
4275 &conf.remote_ip, extack);
4276 if (err) {
4277 netdev_adjacent_change_abort(dst->remote_dev,
4278 lowerdev, dev);
4279 return err;
4280 }
4281 }
8bcdc4f3 4282 }
ce577668 4283
038a5a99
PM		 4284 if (conf.age_interval != vxlan->cfg.age_interval)
4285 mod_timer(&vxlan->age_timer, jiffies);
4286
0ce1822c 4287 netdev_adjacent_change_commit(dst->remote_dev, lowerdev, dev);
845e0ebb 4288 if (lowerdev && lowerdev != dst->remote_dev)
0ce1822c 4289 dst->remote_dev = lowerdev;
038a5a99 4290 vxlan_config_apply(dev, &conf, lowerdev, vxlan->net, true);
8bcdc4f3 4291 return 0;
d342894c 4292}
4293
4294static void vxlan_dellink(struct net_device *dev, struct list_head *head)
4295{
4296 struct vxlan_dev *vxlan = netdev_priv(dev);
4297
8b3f9337
RP		 4298 vxlan_flush(vxlan, true);
4299
553675fb 4300 list_del(&vxlan->next);
d342894c 4301 unregister_netdevice_queue(dev, head);
0ce1822c
TY		 4302 if (vxlan->default_dst.remote_dev)
4303 netdev_upper_dev_unlink(vxlan->default_dst.remote_dev, dev);
d342894c 4304}
4305
4306static size_t vxlan_get_size(const struct net_device *dev)
4307{
4308
4309 return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */
e4c7ed41 4310 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */
d342894c 4311 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */
e4c7ed41 4312 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */
d342894c 4313 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */
8fd78069 4314 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL_INHERIT */
d342894c 4315 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */
b4d30697 4316 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_DF */
e7f70af1 4317 nla_total_size(sizeof(__be32)) + /* IFLA_VXLAN_LABEL */
d342894c 4318 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */
e4f67add
DS		 4319 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */
4320 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */
4321 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */
4322 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */
da8b43c0 4323 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_COLLECT_METADATA */
d342894c 4324 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */
4325 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */
05f47d69 4326 nla_total_size(sizeof(struct ifla_vxlan_port_range)) +
359a0ea9
TH		 4327 nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */
4328 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */
4329 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */
4330 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */
dfd8645e
TH		 4331 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */
4332 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */
69474a8a 4333 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LOCALBYPASS */
4e4b1798
BP		 4334 nla_total_size(0) + /* IFLA_VXLAN_GBP */
4335 nla_total_size(0) + /* IFLA_VXLAN_GPE */
4336 nla_total_size(0) + /* IFLA_VXLAN_REMCSUM_NOPARTIAL */
4337 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_VNIFILTER */
d342894c 4338 0;
4339}
4340
4341static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev)
4342{
4343 const struct vxlan_dev *vxlan = netdev_priv(dev);
c7995c43 4344 const struct vxlan_rdst *dst = &vxlan->default_dst;
05f47d69 4345 struct ifla_vxlan_port_range ports = {
0dfbdf41
TG		 4346 .low = htons(vxlan->cfg.port_min),
4347 .high = htons(vxlan->cfg.port_max),
05f47d69 4348 };
d342894c 4349
54bfd872 4350 if (nla_put_u32(skb, IFLA_VXLAN_ID, be32_to_cpu(dst->remote_vni)))
d342894c 4351 goto nla_put_failure;
4352
e4c7ed41
CW		 4353 if (!vxlan_addr_any(&dst->remote_ip)) {
4354 if (dst->remote_ip.sa.sa_family == AF_INET) {
930345ea
JB		 4355 if (nla_put_in_addr(skb, IFLA_VXLAN_GROUP,
4356 dst->remote_ip.sin.sin_addr.s_addr))
e4c7ed41
CW		 4357 goto nla_put_failure;
4358#if IS_ENABLED(CONFIG_IPV6)
4359 } else {
930345ea
JB		 4360 if (nla_put_in6_addr(skb, IFLA_VXLAN_GROUP6,
4361 &dst->remote_ip.sin6.sin6_addr))
e4c7ed41
CW		 4362 goto nla_put_failure;
4363#endif
4364 }
4365 }
d342894c 4366
c7995c43 4367 if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex))
d342894c 4368 goto nla_put_failure;
4369
0dfbdf41
TG		 4370 if (!vxlan_addr_any(&vxlan->cfg.saddr)) {
4371 if (vxlan->cfg.saddr.sa.sa_family == AF_INET) {
930345ea 4372 if (nla_put_in_addr(skb, IFLA_VXLAN_LOCAL,
0dfbdf41 4373 vxlan->cfg.saddr.sin.sin_addr.s_addr))
e4c7ed41
CW		 4374 goto nla_put_failure;
4375#if IS_ENABLED(CONFIG_IPV6)
4376 } else {
930345ea 4377 if (nla_put_in6_addr(skb, IFLA_VXLAN_LOCAL6,
0dfbdf41 4378 &vxlan->cfg.saddr.sin6.sin6_addr))
e4c7ed41
CW		 4379 goto nla_put_failure;
4380#endif
4381 }
4382 }
d342894c 4383
0dfbdf41 4384 if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->cfg.ttl) ||
8fd78069
HL		 4385 nla_put_u8(skb, IFLA_VXLAN_TTL_INHERIT,
4386 !!(vxlan->cfg.flags & VXLAN_F_TTL_INHERIT)) ||
0dfbdf41 4387 nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->cfg.tos) ||
b4d30697 4388 nla_put_u8(skb, IFLA_VXLAN_DF, vxlan->cfg.df) ||
e7f70af1 4389 nla_put_be32(skb, IFLA_VXLAN_LABEL, vxlan->cfg.label) ||
e4f67add 4390 nla_put_u8(skb, IFLA_VXLAN_LEARNING,
98c81476 4391 !!(vxlan->cfg.flags & VXLAN_F_LEARN)) ||
e4f67add 4392 nla_put_u8(skb, IFLA_VXLAN_PROXY,
98c81476 4393 !!(vxlan->cfg.flags & VXLAN_F_PROXY)) ||
dc5321d7
MS		 4394 nla_put_u8(skb, IFLA_VXLAN_RSC,
4395 !!(vxlan->cfg.flags & VXLAN_F_RSC)) ||
e4f67add 4396 nla_put_u8(skb, IFLA_VXLAN_L2MISS,
98c81476 4397 !!(vxlan->cfg.flags & VXLAN_F_L2MISS)) ||
e4f67add 4398 nla_put_u8(skb, IFLA_VXLAN_L3MISS,
98c81476 4399 !!(vxlan->cfg.flags & VXLAN_F_L3MISS)) ||
da8b43c0 4400 nla_put_u8(skb, IFLA_VXLAN_COLLECT_METADATA,
dc5321d7 4401 !!(vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA)) ||
0dfbdf41
TG		 4402 nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->cfg.age_interval) ||
4403 nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->cfg.addrmax) ||
4404 nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->cfg.dst_port) ||
359a0ea9 4405 nla_put_u8(skb, IFLA_VXLAN_UDP_CSUM,
98c81476 4406 !(vxlan->cfg.flags & VXLAN_F_UDP_ZERO_CSUM_TX)) ||
359a0ea9 4407 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_TX,
98c81476 4408 !!(vxlan->cfg.flags & VXLAN_F_UDP_ZERO_CSUM6_TX)) ||
359a0ea9 4409 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_RX,
98c81476 4410 !!(vxlan->cfg.flags & VXLAN_F_UDP_ZERO_CSUM6_RX)) ||
dfd8645e 4411 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_TX,
98c81476 4412 !!(vxlan->cfg.flags & VXLAN_F_REMCSUM_TX)) ||
dfd8645e 4413 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_RX,
69474a8a
VN		 4414 !!(vxlan->cfg.flags & VXLAN_F_REMCSUM_RX)) ||
4415 nla_put_u8(skb, IFLA_VXLAN_LOCALBYPASS,
4416 !!(vxlan->cfg.flags & VXLAN_F_LOCALBYPASS)))
d342894c 4417 goto nla_put_failure;
4418
05f47d69 4419 if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports))
4420 goto nla_put_failure;
4421
dc5321d7 4422 if (vxlan->cfg.flags & VXLAN_F_GBP &&
3511494c
TG		 4423 nla_put_flag(skb, IFLA_VXLAN_GBP))
4424 goto nla_put_failure;
4425
dc5321d7 4426 if (vxlan->cfg.flags & VXLAN_F_GPE &&
e1e5314d
JB		 4427 nla_put_flag(skb, IFLA_VXLAN_GPE))
4428 goto nla_put_failure;
4429
dc5321d7 4430 if (vxlan->cfg.flags & VXLAN_F_REMCSUM_NOPARTIAL &&
0ace2ca8
TH		 4431 nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL))
4432 goto nla_put_failure;
4433
f9c4bb0b
RP		 4434 if (vxlan->cfg.flags & VXLAN_F_VNIFILTER &&
4435 nla_put_u8(skb, IFLA_VXLAN_VNIFILTER,
4436 !!(vxlan->cfg.flags & VXLAN_F_VNIFILTER)))
4437 goto nla_put_failure;
4438
d342894c 4439 return 0;
4440
4441nla_put_failure:
4442 return -EMSGSIZE;
4443}
4444
1728d4fa
ND		 4445static struct net *vxlan_get_link_net(const struct net_device *dev)
4446{
4447 struct vxlan_dev *vxlan = netdev_priv(dev);
4448
4449 return vxlan->net;
4450}
4451
d342894c 4452static struct rtnl_link_ops vxlan_link_ops __read_mostly = {
4453 .kind = "vxlan",
4454 .maxtype = IFLA_VXLAN_MAX,
4455 .policy = vxlan_policy,
4456 .priv_size = sizeof(struct vxlan_dev),
4457 .setup = vxlan_setup,
4458 .validate = vxlan_validate,
4459 .newlink = vxlan_newlink,
8bcdc4f3 4460 .changelink = vxlan_changelink,
d342894c 4461 .dellink = vxlan_dellink,
4462 .get_size = vxlan_get_size,
4463 .fill_info = vxlan_fill_info,
1728d4fa 4464 .get_link_net = vxlan_get_link_net,
d342894c 4465};
4466
cf5da330
ND		 4467struct net_device *vxlan_dev_create(struct net *net, const char *name,
4468 u8 name_assign_type,
4469 struct vxlan_config *conf)
4470{
4471 struct nlattr *tb[IFLA_MAX + 1];
4472 struct net_device *dev;
4473 int err;
4474
4475 memset(&tb, 0, sizeof(tb));
4476
4477 dev = rtnl_create_link(net, name, name_assign_type,
d0522f1c 4478 &vxlan_link_ops, tb, NULL);
cf5da330
ND		 4479 if (IS_ERR(dev))
4480 return dev;
4481
653ef6a3 4482 err = __vxlan_dev_create(net, dev, conf, NULL);
cf5da330
ND		 4483 if (err < 0) {
4484 free_netdev(dev);
4485 return ERR_PTR(err);
4486 }
4487
1d997f10 4488 err = rtnl_configure_link(dev, NULL, 0, NULL);
cf5da330
ND		 4489 if (err < 0) {
4490 LIST_HEAD(list_kill);
4491
4492 vxlan_dellink(dev, &list_kill);
4493 unregister_netdevice_many(&list_kill);
4494 return ERR_PTR(err);
4495 }
4496
4497 return dev;
4498}
4499EXPORT_SYMBOL_GPL(vxlan_dev_create);
4500
acaf4e70
DB		 4501static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn,
4502 struct net_device *dev)
4503{
4504 struct vxlan_dev *vxlan, *next;
4505 LIST_HEAD(list_kill);
4506
4507 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
4508 struct vxlan_rdst *dst = &vxlan->default_dst;
4509
4510 /* In case we created vxlan device with carrier
4511 * and we loose the carrier due to module unload
4512 * we also need to remove vxlan device. In other
4513 * cases, it's not necessary and remote_ifindex
4514 * is 0 here, so no matches.
4515 */
4516 if (dst->remote_ifindex == dev->ifindex)
4517 vxlan_dellink(vxlan->dev, &list_kill);
4518 }
4519
4520 unregister_netdevice_many(&list_kill);
4521}
4522
b7aade15
HFS		 4523static int vxlan_netdevice_event(struct notifier_block *unused,
4524 unsigned long event, void *ptr)
acaf4e70
DB		 4525{
4526 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
783c1463 4527 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
acaf4e70 4528
dedc33e7 4529 if (event == NETDEV_UNREGISTER)
acaf4e70 4530 vxlan_handle_lowerdev_unregister(vn, dev);
dedc33e7
JK		 4531 else if (event == NETDEV_UDP_TUNNEL_PUSH_INFO)
4532 vxlan_offload_rx_ports(dev, true);
4533 else if (event == NETDEV_UDP_TUNNEL_DROP_INFO)
4534 vxlan_offload_rx_ports(dev, false);
acaf4e70
DB		 4535
4536 return NOTIFY_DONE;
4537}
4538
4539static struct notifier_block vxlan_notifier_block __read_mostly = {
b7aade15 4540 .notifier_call = vxlan_netdevice_event,
acaf4e70
DB		 4541};
4542
0efe1173
PM		 4543static void
4544vxlan_fdb_offloaded_set(struct net_device *dev,
4545 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
4546{
4547 struct vxlan_dev *vxlan = netdev_priv(dev);
4548 struct vxlan_rdst *rdst;
4549 struct vxlan_fdb *f;
fe1e0713
L		 4550 u32 hash_index;
4551
4552 hash_index = fdb_head_index(vxlan, fdb_info->eth_addr, fdb_info->vni);
0efe1173 4553
fe1e0713 4554 spin_lock_bh(&vxlan->hash_lock[hash_index]);
0efe1173
PM		 4555
4556 f = vxlan_find_mac(vxlan, fdb_info->eth_addr, fdb_info->vni);
4557 if (!f)
4558 goto out;
4559
4560 rdst = vxlan_fdb_find_rdst(f, &fdb_info->remote_ip,
4561 fdb_info->remote_port,
4562 fdb_info->remote_vni,
4563 fdb_info->remote_ifindex);
4564 if (!rdst)
4565 goto out;
4566
4567 rdst->offloaded = fdb_info->offloaded;
4568
4569out:
fe1e0713 4570 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
0efe1173
PM		 4571}
4572
5728ae0d
PM		 4573static int
4574vxlan_fdb_external_learn_add(struct net_device *dev,
4575 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
4576{
4577 struct vxlan_dev *vxlan = netdev_priv(dev);
4c59b7d1 4578 struct netlink_ext_ack *extack;
fe1e0713 4579 u32 hash_index;
5728ae0d
PM		 4580 int err;
4581
fe1e0713 4582 hash_index = fdb_head_index(vxlan, fdb_info->eth_addr, fdb_info->vni);
4c59b7d1
PM		 4583 extack = switchdev_notifier_info_to_extack(&fdb_info->info);
4584
fe1e0713 4585 spin_lock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4586 err = vxlan_fdb_update(vxlan, fdb_info->eth_addr, &fdb_info->remote_ip,
4587 NUD_REACHABLE,
4588 NLM_F_CREATE | NLM_F_REPLACE,
4589 fdb_info->remote_port,
4590 fdb_info->vni,
4591 fdb_info->remote_vni,
4592 fdb_info->remote_ifindex,
4593 NTF_USE | NTF_SELF | NTF_EXT_LEARNED,
1274e1cc 4594 0, false, extack);
fe1e0713 4595 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4596
4597 return err;
4598}
4599
4600static int
4601vxlan_fdb_external_learn_del(struct net_device *dev,
4602 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
4603{
4604 struct vxlan_dev *vxlan = netdev_priv(dev);
4605 struct vxlan_fdb *f;
fe1e0713 4606 u32 hash_index;
5728ae0d
PM		 4607 int err = 0;
4608
fe1e0713
L		 4609 hash_index = fdb_head_index(vxlan, fdb_info->eth_addr, fdb_info->vni);
4610 spin_lock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4611
4612 f = vxlan_find_mac(vxlan, fdb_info->eth_addr, fdb_info->vni);
4613 if (!f)
4614 err = -ENOENT;
4615 else if (f->flags & NTF_EXT_LEARNED)
4616 err = __vxlan_fdb_delete(vxlan, fdb_info->eth_addr,
4617 fdb_info->remote_ip,
4618 fdb_info->remote_port,
4619 fdb_info->vni,
4620 fdb_info->remote_vni,
4621 fdb_info->remote_ifindex,
4622 false);
4623
fe1e0713 4624 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4625
4626 return err;
4627}
4628
0efe1173
PM		 4629static int vxlan_switchdev_event(struct notifier_block *unused,
4630 unsigned long event, void *ptr)
4631{
4632 struct net_device *dev = switchdev_notifier_info_to_dev(ptr);
5728ae0d
PM		 4633 struct switchdev_notifier_vxlan_fdb_info *fdb_info;
4634 int err = 0;
0efe1173
PM		 4635
4636 switch (event) {
4637 case SWITCHDEV_VXLAN_FDB_OFFLOADED:
4638 vxlan_fdb_offloaded_set(dev, ptr);
4639 break;
5728ae0d
PM		 4640 case SWITCHDEV_VXLAN_FDB_ADD_TO_BRIDGE:
4641 fdb_info = ptr;
4642 err = vxlan_fdb_external_learn_add(dev, fdb_info);
4643 if (err) {
4644 err = notifier_from_errno(err);
4645 break;
4646 }
4647 fdb_info->offloaded = true;
4648 vxlan_fdb_offloaded_set(dev, fdb_info);
4649 break;
4650 case SWITCHDEV_VXLAN_FDB_DEL_TO_BRIDGE:
4651 fdb_info = ptr;
4652 err = vxlan_fdb_external_learn_del(dev, fdb_info);
4653 if (err) {
4654 err = notifier_from_errno(err);
4655 break;
4656 }
4657 fdb_info->offloaded = false;
4658 vxlan_fdb_offloaded_set(dev, fdb_info);
4659 break;
0efe1173
PM		 4660 }
4661
5728ae0d 4662 return err;
0efe1173
PM		 4663}
4664
4665static struct notifier_block vxlan_switchdev_notifier_block __read_mostly = {
4666 .notifier_call = vxlan_switchdev_event,
4667};
4668
79472fe8
RP		 4669static void vxlan_fdb_nh_flush(struct nexthop *nh)
4670{
4671 struct vxlan_fdb *fdb;
4672 struct vxlan_dev *vxlan;
4673 u32 hash_index;
4674
4675 rcu_read_lock();
4676 list_for_each_entry_rcu(fdb, &nh->fdb_list, nh_list) {
4677 vxlan = rcu_dereference(fdb->vdev);
4678 WARN_ON(!vxlan);
4679 hash_index = fdb_head_index(vxlan, fdb->eth_addr,
4680 vxlan->default_dst.remote_vni);
4681 spin_lock_bh(&vxlan->hash_lock[hash_index]);
4682 if (!hlist_unhashed(&fdb->hlist))
4683 vxlan_fdb_destroy(vxlan, fdb, false, false);
4684 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
4685 }
4686 rcu_read_unlock();
4687}
4688
c7cdbe2e
RP		 4689static int vxlan_nexthop_event(struct notifier_block *nb,
4690 unsigned long event, void *ptr)
4691{
1ec69d18
IS		 4692 struct nh_notifier_info *info = ptr;
4693 struct nexthop *nh;
4694
4695 if (event != NEXTHOP_EVENT_DEL)
4696 return NOTIFY_DONE;
c7cdbe2e 4697
1ec69d18
IS		 4698 nh = nexthop_find_by_id(info->net, info->id);
4699 if (!nh)
c7cdbe2e
RP		 4700 return NOTIFY_DONE;
4701
79472fe8 4702 vxlan_fdb_nh_flush(nh);
c7cdbe2e
RP		 4703
4704 return NOTIFY_DONE;
4705}
4706
d342894c 4707static __net_init int vxlan_init_net(struct net *net)
4708{
4709 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
31fec5aa 4710 unsigned int h;
d342894c 4711
553675fb 4712 INIT_LIST_HEAD(&vn->vxlan_list);
1c51a915 4713 spin_lock_init(&vn->sock_lock);
626d667b 4714 vn->nexthop_notifier_block.notifier_call = vxlan_nexthop_event;
d342894c 4715
553675fb 4716 for (h = 0; h < PORT_HASH_SIZE; ++h)
4717 INIT_HLIST_HEAD(&vn->sock_list[h]);
d342894c 4718
ce7e9c8a
IS		 4719 return register_nexthop_notifier(net, &vn->nexthop_notifier_block,
4720 NULL);
d342894c 4721}
4722
57b61127 4723static void vxlan_destroy_tunnels(struct net *net, struct list_head *head)
f01ec1c0
ND		 4724{
4725 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
4726 struct vxlan_dev *vxlan, *next;
4727 struct net_device *dev, *aux;
f01ec1c0 4728
f01ec1c0
ND		 4729 for_each_netdev_safe(net, dev, aux)
4730 if (dev->rtnl_link_ops == &vxlan_link_ops)
57b61127 4731 unregister_netdevice_queue(dev, head);
f01ec1c0
ND		 4732
4733 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
4734 /* If vxlan->dev is in the same netns, it has already been added
4735 * to the list by the previous loop.
4736 */
cc4807bb 4737 if (!net_eq(dev_net(vxlan->dev), net))
57b61127 4738 unregister_netdevice_queue(vxlan->dev, head);
f01ec1c0
ND		 4739 }
4740
f01ec1c0
ND		 4741}
4742
57b61127
HY		 4743static void __net_exit vxlan_exit_batch_net(struct list_head *net_list)
4744{
4745 struct net *net;
4746 LIST_HEAD(list);
92584ddf 4747 unsigned int h;
57b61127 4748
626d667b
IS		 4749 list_for_each_entry(net, net_list, exit_list) {
4750 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
4751
4752 unregister_nexthop_notifier(net, &vn->nexthop_notifier_block);
4753 }
3106a084 4754 rtnl_lock();
57b61127
HY		 4755 list_for_each_entry(net, net_list, exit_list)
4756 vxlan_destroy_tunnels(net, &list);
4757
4758 unregister_netdevice_many(&list);
4759 rtnl_unlock();
92584ddf
TY		 4760
4761 list_for_each_entry(net, net_list, exit_list) {
4762 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
4763
4764 for (h = 0; h < PORT_HASH_SIZE; ++h)
4765 WARN_ON_ONCE(!hlist_empty(&vn->sock_list[h]));
4766 }
57b61127
HY		 4767}
4768
d342894c 4769static struct pernet_operations vxlan_net_ops = {
4770 .init = vxlan_init_net,
57b61127 4771 .exit_batch = vxlan_exit_batch_net,
d342894c 4772 .id = &vxlan_net_id,
4773 .size = sizeof(struct vxlan_net),
4774};
4775
4776static int __init vxlan_init_module(void)
4777{
4778 int rc;
4779
4780 get_random_bytes(&vxlan_salt, sizeof(vxlan_salt));
4781
783c1463 4782 rc = register_pernet_subsys(&vxlan_net_ops);
d342894c 4783 if (rc)
4784 goto out1;
4785
acaf4e70 4786 rc = register_netdevice_notifier(&vxlan_notifier_block);
d342894c 4787 if (rc)
4788 goto out2;
4789
0efe1173 4790 rc = register_switchdev_notifier(&vxlan_switchdev_notifier_block);
acaf4e70
DB		 4791 if (rc)
4792 goto out3;
d342894c 4793
0efe1173
PM		 4794 rc = rtnl_link_register(&vxlan_link_ops);
4795 if (rc)
4796 goto out4;
4797
f9c4bb0b
RP		 4798 vxlan_vnifilter_init();
4799
acaf4e70 4800 return 0;
0efe1173
PM		 4801out4:
4802 unregister_switchdev_notifier(&vxlan_switchdev_notifier_block);
acaf4e70
DB		 4803out3:
4804 unregister_netdevice_notifier(&vxlan_notifier_block);
d342894c 4805out2:
783c1463 4806 unregister_pernet_subsys(&vxlan_net_ops);
d342894c 4807out1:
4808 return rc;
4809}
7332a13b 4810late_initcall(vxlan_init_module);
d342894c 4811
4812static void __exit vxlan_cleanup_module(void)
4813{
f9c4bb0b 4814 vxlan_vnifilter_uninit();
b7153984 4815 rtnl_link_unregister(&vxlan_link_ops);
0efe1173 4816 unregister_switchdev_notifier(&vxlan_switchdev_notifier_block);
acaf4e70 4817 unregister_netdevice_notifier(&vxlan_notifier_block);
783c1463
DB		 4818 unregister_pernet_subsys(&vxlan_net_ops);
4819 /* rcu_barrier() is called by netns */
d342894c 4820}
4821module_exit(vxlan_cleanup_module);
4822
4823MODULE_LICENSE("GPL");
4824MODULE_VERSION(VXLAN_VERSION);
3b8df3c6 4825MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>");
ead5139a 4826MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic");
d342894c 4827MODULE_ALIAS_RTNL_LINK("vxlan");
Linux 6.x block layer and io_uring tree(s)