]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame - drivers/net/vxlan.c
projects / mirror_ubuntu-jammy-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge branch 'parisc-5.9-2' of git://git.kernel.org/pub/scm/linux/kernel/git/deller...
[mirror_ubuntu-jammy-kernel.git] / drivers / net / vxlan.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>
3616d08b 20#include <net/ipv6_stubs.h>
d342894c 21#include <net/ip.h>
22#include <net/icmp.h>
d342894c 23#include <net/rtnetlink.h>
d342894c 24#include <net/inet_ecn.h>
25#include <net/net_namespace.h>
26#include <net/netns/generic.h>
fa20e0e3 27#include <net/tun_proto.h>
012a5729 28#include <net/vxlan.h>
1274e1cc 29#include <net/nexthop.h>
40d29af0 30
e4c7ed41 31#if IS_ENABLED(CONFIG_IPV6)
e4c7ed41 32#include <net/ip6_tunnel.h>
660d98ca 33#include <net/ip6_checksum.h>
e4c7ed41 34#endif
d342894c 35
36#define VXLAN_VERSION "0.1"
37
553675fb 38#define PORT_HASH_BITS 8
39#define PORT_HASH_SIZE (1<<PORT_HASH_BITS)
d342894c 40#define FDB_AGE_DEFAULT 300 /* 5 min */
41#define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */
42
23c578bf 43/* UDP port for VXLAN traffic.
44 * The IANA assigned port is 4789, but the Linux default is 8472
234f5b73 45 * for compatibility with early adopters.
23c578bf 46 */
9daaa397
SH		 47static unsigned short vxlan_port __read_mostly = 8472;
48module_param_named(udp_port, vxlan_port, ushort, 0444);
d342894c 49MODULE_PARM_DESC(udp_port, "Destination UDP port");
50
51static bool log_ecn_error = true;
52module_param(log_ecn_error, bool, 0644);
53MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
54
c7d03a00 55static unsigned int vxlan_net_id;
0dfbdf41 56static struct rtnl_link_ops vxlan_link_ops;
553675fb 57
7256eac1 58static const u8 all_zeros_mac[ETH_ALEN + 2];
afbd8bae 59
205f356d 60static int vxlan_sock_add(struct vxlan_dev *vxlan);
614732ea 61
a53cb29b
MB		 62static void vxlan_vs_del_dev(struct vxlan_dev *vxlan);
63
553675fb 64/* per-network namespace private data for this module */
65struct vxlan_net {
66 struct list_head vxlan_list;
67 struct hlist_head sock_list[PORT_HASH_SIZE];
1c51a915 68 spinlock_t sock_lock;
553675fb 69};
70
d342894c 71/* Forwarding table entry */
72struct vxlan_fdb {
73 struct hlist_node hlist; /* linked list of entries */
74 struct rcu_head rcu;
75 unsigned long updated; /* jiffies */
76 unsigned long used;
3e61aa8f 77 struct list_head remotes;
7177a3b0 78 u8 eth_addr[ETH_ALEN];
d342894c 79 u16 state; /* see ndm_state */
3ad7a4b1 80 __be32 vni;
45598c1c 81 u16 flags; /* see ndm_flags and below */
1274e1cc
RP		 82 struct list_head nh_list;
83 struct nexthop __rcu *nh;
79472fe8 84 struct vxlan_dev __rcu *vdev;
d342894c 85};
86
45598c1c
PM		 87#define NTF_VXLAN_ADDED_BY_USER 0x100
88
d342894c 89/* salt for hash table */
90static u32 vxlan_salt __read_mostly;
91
ee122c79
TG		 92static inline bool vxlan_collect_metadata(struct vxlan_sock *vs)
93{
e7030878
TG		 94 return vs->flags & VXLAN_F_COLLECT_METADATA ||
95 ip_tunnel_collect_metadata();
ee122c79
TG		 96}
97
e4c7ed41
CW		 98#if IS_ENABLED(CONFIG_IPV6)
99static inline
100bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
101{
f0ef3126
JB		 102 if (a->sa.sa_family != b->sa.sa_family)
103 return false;
104 if (a->sa.sa_family == AF_INET6)
105 return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr);
106 else
107 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
e4c7ed41
CW		 108}
109
e4c7ed41
CW		 110static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
111{
f0ef3126 112 if (nla_len(nla) >= sizeof(struct in6_addr)) {
67b61f6c 113 ip->sin6.sin6_addr = nla_get_in6_addr(nla);
f0ef3126
JB		 114 ip->sa.sa_family = AF_INET6;
115 return 0;
116 } else if (nla_len(nla) >= sizeof(__be32)) {
67b61f6c 117 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
f0ef3126
JB		 118 ip->sa.sa_family = AF_INET;
119 return 0;
120 } else {
121 return -EAFNOSUPPORT;
122 }
e4c7ed41
CW		 123}
124
125static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
f0ef3126 126 const union vxlan_addr *ip)
e4c7ed41 127{
f0ef3126 128 if (ip->sa.sa_family == AF_INET6)
930345ea 129 return nla_put_in6_addr(skb, attr, &ip->sin6.sin6_addr);
f0ef3126 130 else
930345ea 131 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
e4c7ed41
CW		 132}
133
134#else /* !CONFIG_IPV6 */
135
136static inline
137bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
138{
f0ef3126 139 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
e4c7ed41
CW		 140}
141
e4c7ed41
CW		 142static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
143{
f0ef3126
JB		 144 if (nla_len(nla) >= sizeof(struct in6_addr)) {
145 return -EAFNOSUPPORT;
146 } else if (nla_len(nla) >= sizeof(__be32)) {
67b61f6c 147 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
f0ef3126
JB		 148 ip->sa.sa_family = AF_INET;
149 return 0;
150 } else {
151 return -EAFNOSUPPORT;
152 }
e4c7ed41
CW		 153}
154
155static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
f0ef3126 156 const union vxlan_addr *ip)
e4c7ed41 157{
930345ea 158 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
e4c7ed41
CW		 159}
160#endif
161
553675fb 162/* Virtual Network hash table head */
54bfd872 163static inline struct hlist_head *vni_head(struct vxlan_sock *vs, __be32 vni)
553675fb 164{
54bfd872 165 return &vs->vni_list[hash_32((__force u32)vni, VNI_HASH_BITS)];
553675fb 166}
167
168/* Socket hash table head */
169static inline struct hlist_head *vs_head(struct net *net, __be16 port)
d342894c 170{
171 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
172
553675fb 173 return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)];
174}
175
3e61aa8f
SH		 176/* First remote destination for a forwarding entry.
177 * Guaranteed to be non-NULL because remotes are never deleted.
178 */
5ca5461c 179static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb)
3e61aa8f 180{
1274e1cc
RP		 181 if (rcu_access_pointer(fdb->nh))
182 return NULL;
5ca5461c 183 return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list);
184}
185
186static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb)
187{
1274e1cc
RP		 188 if (rcu_access_pointer(fdb->nh))
189 return NULL;
5ca5461c 190 return list_first_entry(&fdb->remotes, struct vxlan_rdst, list);
3e61aa8f
SH		 191}
192
ac5132d1
TG		 193/* Find VXLAN socket based on network namespace, address family and UDP port
194 * and enabled unshareable flags.
195 */
196static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family,
aab8cc36 197 __be16 port, u32 flags, int ifindex)
553675fb 198{
199 struct vxlan_sock *vs;
af33c1ad
TH		 200
201 flags &= VXLAN_F_RCV_FLAGS;
553675fb 202
203 hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) {
19ca9fc1 204 if (inet_sk(vs->sock->sk)->inet_sport == port &&
705cc62f 205 vxlan_get_sk_family(vs) == family &&
aab8cc36
AB		 206 vs->flags == flags &&
207 vs->sock->sk->sk_bound_dev_if == ifindex)
553675fb 208 return vs;
209 }
210 return NULL;
d342894c 211}
212
49f810f0
MS		 213static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, int ifindex,
214 __be32 vni)
d342894c 215{
69e76661 216 struct vxlan_dev_node *node;
d342894c 217
b9167b2e
JB		 218 /* For flow based devices, map all packets to VNI 0 */
219 if (vs->flags & VXLAN_F_COLLECT_METADATA)
220 vni = 0;
221
69e76661
JB		 222 hlist_for_each_entry_rcu(node, vni_head(vs, vni), hlist) {
223 if (node->vxlan->default_dst.remote_vni != vni)
49f810f0
MS		 224 continue;
225
226 if (IS_ENABLED(CONFIG_IPV6)) {
69e76661 227 const struct vxlan_config *cfg = &node->vxlan->cfg;
49f810f0
MS		 228
229 if ((cfg->flags & VXLAN_F_IPV6_LINKLOCAL) &&
230 cfg->remote_ifindex != ifindex)
231 continue;
232 }
233
69e76661 234 return node->vxlan;
d342894c 235 }
236
237 return NULL;
238}
239
5cfccc5a 240/* Look up VNI in a per net namespace table */
49f810f0
MS		 241static struct vxlan_dev *vxlan_find_vni(struct net *net, int ifindex,
242 __be32 vni, sa_family_t family,
243 __be16 port, u32 flags)
5cfccc5a
PS		 244{
245 struct vxlan_sock *vs;
246
aab8cc36 247 vs = vxlan_find_sock(net, family, port, flags, ifindex);
5cfccc5a
PS		 248 if (!vs)
249 return NULL;
250
49f810f0 251 return vxlan_vs_find_vni(vs, ifindex, vni);
5cfccc5a
PS		 252}
253
d342894c 254/* Fill in neighbour message in skbuff. */
255static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan,
234f5b73
SH		 256 const struct vxlan_fdb *fdb,
257 u32 portid, u32 seq, int type, unsigned int flags,
258 const struct vxlan_rdst *rdst)
d342894c 259{
260 unsigned long now = jiffies;
261 struct nda_cacheinfo ci;
1274e1cc 262 bool send_ip, send_eth;
d342894c 263 struct nlmsghdr *nlh;
1274e1cc 264 struct nexthop *nh;
d342894c 265 struct ndmsg *ndm;
06ec313e
IS		 266 int nh_family;
267 u32 nh_id;
d342894c 268
269 nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags);
270 if (nlh == NULL)
271 return -EMSGSIZE;
272
273 ndm = nlmsg_data(nlh);
274 memset(ndm, 0, sizeof(*ndm));
e4f67add
DS		 275
276 send_eth = send_ip = true;
277
06ec313e
IS		 278 rcu_read_lock();
279 nh = rcu_dereference(fdb->nh);
280 if (nh) {
281 nh_family = nexthop_get_family(nh);
282 nh_id = nh->id;
283 }
284 rcu_read_unlock();
285
e4f67add 286 if (type == RTM_GETNEIGH) {
1274e1cc
RP		 287 if (rdst) {
288 send_ip = !vxlan_addr_any(&rdst->remote_ip);
289 ndm->ndm_family = send_ip ? rdst->remote_ip.sa.sa_family : AF_INET;
290 } else if (nh) {
06ec313e 291 ndm->ndm_family = nh_family;
1274e1cc 292 }
e4f67add
DS		 293 send_eth = !is_zero_ether_addr(fdb->eth_addr);
294 } else
295 ndm->ndm_family = AF_BRIDGE;
d342894c 296 ndm->ndm_state = fdb->state;
297 ndm->ndm_ifindex = vxlan->dev->ifindex;
ae884082 298 ndm->ndm_flags = fdb->flags;
1274e1cc 299 if (rdst && rdst->offloaded)
0efe1173 300 ndm->ndm_flags |= NTF_OFFLOADED;
545469f7 301 ndm->ndm_type = RTN_UNICAST;
d342894c 302
193523bf 303 if (!net_eq(dev_net(vxlan->dev), vxlan->net) &&
4967082b 304 nla_put_s32(skb, NDA_LINK_NETNSID,
38f507f1 305 peernet2id(dev_net(vxlan->dev), vxlan->net)))
193523bf
ND		 306 goto nla_put_failure;
307
e4f67add 308 if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr))
d342894c 309 goto nla_put_failure;
1274e1cc 310 if (nh) {
06ec313e 311 if (nla_put_u32(skb, NDA_NH_ID, nh_id))
1274e1cc
RP		 312 goto nla_put_failure;
313 } else if (rdst) {
314 if (send_ip && vxlan_nla_put_addr(skb, NDA_DST,
315 &rdst->remote_ip))
316 goto nla_put_failure;
317
318 if (rdst->remote_port &&
319 rdst->remote_port != vxlan->cfg.dst_port &&
320 nla_put_be16(skb, NDA_PORT, rdst->remote_port))
321 goto nla_put_failure;
322 if (rdst->remote_vni != vxlan->default_dst.remote_vni &&
323 nla_put_u32(skb, NDA_VNI, be32_to_cpu(rdst->remote_vni)))
324 goto nla_put_failure;
325 if (rdst->remote_ifindex &&
326 nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex))
327 goto nla_put_failure;
328 }
d342894c 329
dc5321d7 330 if ((vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) && fdb->vni &&
3ad7a4b1
RP		 331 nla_put_u32(skb, NDA_SRC_VNI,
332 be32_to_cpu(fdb->vni)))
333 goto nla_put_failure;
d342894c 334
335 ci.ndm_used = jiffies_to_clock_t(now - fdb->used);
336 ci.ndm_confirmed = 0;
337 ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated);
338 ci.ndm_refcnt = 0;
339
340 if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci))
341 goto nla_put_failure;
342
053c095a
JB		 343 nlmsg_end(skb, nlh);
344 return 0;
d342894c 345
346nla_put_failure:
347 nlmsg_cancel(skb, nlh);
348 return -EMSGSIZE;
349}
350
351static inline size_t vxlan_nlmsg_size(void)
352{
353 return NLMSG_ALIGN(sizeof(struct ndmsg))
354 + nla_total_size(ETH_ALEN) /* NDA_LLADDR */
e4c7ed41 355 + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */
73cf3317 356 + nla_total_size(sizeof(__be16)) /* NDA_PORT */
6681712d
DS		 357 + nla_total_size(sizeof(__be32)) /* NDA_VNI */
358 + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */
4967082b 359 + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */
d342894c 360 + nla_total_size(sizeof(struct nda_cacheinfo));
361}
362
9a997353
PM		 363static void __vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
364 struct vxlan_rdst *rd, int type)
d342894c 365{
366 struct net *net = dev_net(vxlan->dev);
367 struct sk_buff *skb;
368 int err = -ENOBUFS;
369
370 skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC);
371 if (skb == NULL)
372 goto errout;
373
9e4b93f9 374 err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd);
d342894c 375 if (err < 0) {
376 /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */
377 WARN_ON(err == -EMSGSIZE);
378 kfree_skb(skb);
379 goto errout;
380 }
381
382 rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC);
383 return;
384errout:
385 if (err < 0)
386 rtnl_set_sk_err(net, RTNLGRP_NEIGH, err);
387}
388
ff23b91c
PM		 389static void vxlan_fdb_switchdev_notifier_info(const struct vxlan_dev *vxlan,
390 const struct vxlan_fdb *fdb,
391 const struct vxlan_rdst *rd,
4c59b7d1 392 struct netlink_ext_ack *extack,
ff23b91c
PM		 393 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
394{
395 fdb_info->info.dev = vxlan->dev;
4c59b7d1 396 fdb_info->info.extack = extack;
ff23b91c
PM		 397 fdb_info->remote_ip = rd->remote_ip;
398 fdb_info->remote_port = rd->remote_port;
399 fdb_info->remote_vni = rd->remote_vni;
400 fdb_info->remote_ifindex = rd->remote_ifindex;
401 memcpy(fdb_info->eth_addr, fdb->eth_addr, ETH_ALEN);
402 fdb_info->vni = fdb->vni;
403 fdb_info->offloaded = rd->offloaded;
404 fdb_info->added_by_user = fdb->flags & NTF_VXLAN_ADDED_BY_USER;
405}
406
61f46fe8
PM		 407static int vxlan_fdb_switchdev_call_notifiers(struct vxlan_dev *vxlan,
408 struct vxlan_fdb *fdb,
409 struct vxlan_rdst *rd,
4c59b7d1
PM		 410 bool adding,
411 struct netlink_ext_ack *extack)
9a997353
PM		 412{
413 struct switchdev_notifier_vxlan_fdb_info info;
414 enum switchdev_notifier_type notifier_type;
61f46fe8 415 int ret;
9a997353
PM		 416
417 if (WARN_ON(!rd))
61f46fe8 418 return 0;
9a997353
PM		 419
420 notifier_type = adding ? SWITCHDEV_VXLAN_FDB_ADD_TO_DEVICE
421 : SWITCHDEV_VXLAN_FDB_DEL_TO_DEVICE;
4c59b7d1 422 vxlan_fdb_switchdev_notifier_info(vxlan, fdb, rd, NULL, &info);
61f46fe8 423 ret = call_switchdev_notifiers(notifier_type, vxlan->dev,
6685987c 424 &info.info, extack);
61f46fe8 425 return notifier_to_errno(ret);
9a997353
PM		 426}
427
61f46fe8 428static int vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
4c59b7d1
PM		 429 struct vxlan_rdst *rd, int type, bool swdev_notify,
430 struct netlink_ext_ack *extack)
9a997353 431{
61f46fe8
PM		 432 int err;
433
1274e1cc 434 if (swdev_notify && rd) {
0e6160f3
PM		 435 switch (type) {
436 case RTM_NEWNEIGH:
61f46fe8 437 err = vxlan_fdb_switchdev_call_notifiers(vxlan, fdb, rd,
4c59b7d1 438 true, extack);
61f46fe8
PM		 439 if (err)
440 return err;
0e6160f3
PM		 441 break;
442 case RTM_DELNEIGH:
443 vxlan_fdb_switchdev_call_notifiers(vxlan, fdb, rd,
4c59b7d1 444 false, extack);
0e6160f3
PM		 445 break;
446 }
9a997353
PM		 447 }
448
449 __vxlan_fdb_notify(vxlan, fdb, rd, type);
61f46fe8 450 return 0;
9a997353
PM		 451}
452
e4c7ed41 453static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa)
e4f67add
DS		 454{
455 struct vxlan_dev *vxlan = netdev_priv(dev);
bb3fd687
SH		 456 struct vxlan_fdb f = {
457 .state = NUD_STALE,
458 };
459 struct vxlan_rdst remote = {
e4c7ed41 460 .remote_ip = *ipa, /* goes to NDA_DST */
54bfd872 461 .remote_vni = cpu_to_be32(VXLAN_N_VID),
bb3fd687 462 };
3e61aa8f 463
4c59b7d1 464 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH, true, NULL);
e4f67add
DS		 465}
466
467static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN])
468{
bb3fd687
SH		 469 struct vxlan_fdb f = {
470 .state = NUD_STALE,
471 };
9e4b93f9 472 struct vxlan_rdst remote = { };
e4f67add 473
e4f67add
DS		 474 memcpy(f.eth_addr, eth_addr, ETH_ALEN);
475
4c59b7d1 476 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH, true, NULL);
e4f67add
DS		 477}
478
d342894c 479/* Hash Ethernet address */
480static u32 eth_hash(const unsigned char *addr)
481{
482 u64 value = get_unaligned((u64 *)addr);
483
484 /* only want 6 bytes */
485#ifdef __BIG_ENDIAN
d342894c 486 value >>= 16;
321fb991 487#else
488 value <<= 16;
d342894c 489#endif
490 return hash_64(value, FDB_HASH_BITS);
491}
492
3ad7a4b1
RP		 493static u32 eth_vni_hash(const unsigned char *addr, __be32 vni)
494{
495 /* use 1 byte of OUI and 3 bytes of NIC */
496 u32 key = get_unaligned((u32 *)(addr + 2));
497
498 return jhash_2words(key, vni, vxlan_salt) & (FDB_HASH_SIZE - 1);
499}
500
fe1e0713
L		 501static u32 fdb_head_index(struct vxlan_dev *vxlan, const u8 *mac, __be32 vni)
502{
503 if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA)
504 return eth_vni_hash(mac, vni);
505 else
506 return eth_hash(mac);
507}
508
d342894c 509/* Hash chain to use given mac address */
510static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan,
3ad7a4b1 511 const u8 *mac, __be32 vni)
d342894c 512{
fe1e0713 513 return &vxlan->fdb_head[fdb_head_index(vxlan, mac, vni)];
d342894c 514}
515
516/* Look up Ethernet address in forwarding table */
014be2c8 517static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan,
3ad7a4b1 518 const u8 *mac, __be32 vni)
d342894c 519{
3ad7a4b1 520 struct hlist_head *head = vxlan_fdb_head(vxlan, mac, vni);
d342894c 521 struct vxlan_fdb *f;
d342894c 522
b67bfe0d 523 hlist_for_each_entry_rcu(f, head, hlist) {
3ad7a4b1 524 if (ether_addr_equal(mac, f->eth_addr)) {
dc5321d7 525 if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) {
3ad7a4b1
RP		 526 if (vni == f->vni)
527 return f;
528 } else {
529 return f;
530 }
531 }
d342894c 532 }
533
534 return NULL;
535}
536
014be2c8 537static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan,
3ad7a4b1 538 const u8 *mac, __be32 vni)
014be2c8
SS		 539{
540 struct vxlan_fdb *f;
541
3ad7a4b1 542 f = __vxlan_find_mac(vxlan, mac, vni);
016f3d18 543 if (f && f->used != jiffies)
014be2c8
SS		 544 f->used = jiffies;
545
546 return f;
547}
548
a5e7c10a
MR		 549/* caller should hold vxlan->hash_lock */
550static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f,
e4c7ed41 551 union vxlan_addr *ip, __be16 port,
54bfd872 552 __be32 vni, __u32 ifindex)
6681712d 553{
3e61aa8f 554 struct vxlan_rdst *rd;
6681712d 555
3e61aa8f 556 list_for_each_entry(rd, &f->remotes, list) {
e4c7ed41 557 if (vxlan_addr_equal(&rd->remote_ip, ip) &&
6681712d
DS		 558 rd->remote_port == port &&
559 rd->remote_vni == vni &&
560 rd->remote_ifindex == ifindex)
a5e7c10a 561 return rd;
6681712d 562 }
3e61aa8f 563
a5e7c10a
MR		 564 return NULL;
565}
566
1941f1d6
PM		 567int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
568 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
569{
570 struct vxlan_dev *vxlan = netdev_priv(dev);
571 u8 eth_addr[ETH_ALEN + 2] = { 0 };
572 struct vxlan_rdst *rdst;
573 struct vxlan_fdb *f;
574 int rc = 0;
575
576 if (is_multicast_ether_addr(mac) ||
577 is_zero_ether_addr(mac))
578 return -EINVAL;
579
580 ether_addr_copy(eth_addr, mac);
581
582 rcu_read_lock();
583
584 f = __vxlan_find_mac(vxlan, eth_addr, vni);
585 if (!f) {
586 rc = -ENOENT;
587 goto out;
588 }
589
590 rdst = first_remote_rcu(f);
4c59b7d1 591 vxlan_fdb_switchdev_notifier_info(vxlan, f, rdst, NULL, fdb_info);
1941f1d6
PM		 592
593out:
594 rcu_read_unlock();
595 return rc;
596}
597EXPORT_SYMBOL_GPL(vxlan_fdb_find_uc);
598
4f89f5b5
PM		 599static int vxlan_fdb_notify_one(struct notifier_block *nb,
600 const struct vxlan_dev *vxlan,
601 const struct vxlan_fdb *f,
4c59b7d1
PM		 602 const struct vxlan_rdst *rdst,
603 struct netlink_ext_ack *extack)
4f89f5b5
PM		 604{
605 struct switchdev_notifier_vxlan_fdb_info fdb_info;
606 int rc;
607
4c59b7d1 608 vxlan_fdb_switchdev_notifier_info(vxlan, f, rdst, extack, &fdb_info);
4f89f5b5
PM		 609 rc = nb->notifier_call(nb, SWITCHDEV_VXLAN_FDB_ADD_TO_DEVICE,
610 &fdb_info);
611 return notifier_to_errno(rc);
612}
613
614int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
4c59b7d1
PM		 615 struct notifier_block *nb,
616 struct netlink_ext_ack *extack)
4f89f5b5
PM		 617{
618 struct vxlan_dev *vxlan;
619 struct vxlan_rdst *rdst;
620 struct vxlan_fdb *f;
621 unsigned int h;
622 int rc = 0;
623
624 if (!netif_is_vxlan(dev))
625 return -EINVAL;
626 vxlan = netdev_priv(dev);
627
4f89f5b5 628 for (h = 0; h < FDB_HASH_SIZE; ++h) {
fe1e0713 629 spin_lock_bh(&vxlan->hash_lock[h]);
4f89f5b5
PM		 630 hlist_for_each_entry(f, &vxlan->fdb_head[h], hlist) {
631 if (f->vni == vni) {
632 list_for_each_entry(rdst, &f->remotes, list) {
633 rc = vxlan_fdb_notify_one(nb, vxlan,
4c59b7d1
PM		 634 f, rdst,
635 extack);
4f89f5b5 636 if (rc)
fe1e0713 637 goto unlock;
4f89f5b5
PM		 638 }
639 }
640 }
fe1e0713 641 spin_unlock_bh(&vxlan->hash_lock[h]);
4f89f5b5 642 }
fe1e0713 643 return 0;
4f89f5b5 644
fe1e0713
L		 645unlock:
646 spin_unlock_bh(&vxlan->hash_lock[h]);
4f89f5b5
PM		 647 return rc;
648}
649EXPORT_SYMBOL_GPL(vxlan_fdb_replay);
650
e5ff4b19
PM		 651void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni)
652{
653 struct vxlan_dev *vxlan;
654 struct vxlan_rdst *rdst;
655 struct vxlan_fdb *f;
656 unsigned int h;
657
658 if (!netif_is_vxlan(dev))
659 return;
660 vxlan = netdev_priv(dev);
661
e5ff4b19 662 for (h = 0; h < FDB_HASH_SIZE; ++h) {
fe1e0713 663 spin_lock_bh(&vxlan->hash_lock[h]);
e5ff4b19
PM		 664 hlist_for_each_entry(f, &vxlan->fdb_head[h], hlist)
665 if (f->vni == vni)
666 list_for_each_entry(rdst, &f->remotes, list)
667 rdst->offloaded = false;
fe1e0713 668 spin_unlock_bh(&vxlan->hash_lock[h]);
e5ff4b19 669 }
fe1e0713 670
e5ff4b19
PM		 671}
672EXPORT_SYMBOL_GPL(vxlan_fdb_clear_offload);
673
906dc186
TR		 674/* Replace destination of unicast mac */
675static int vxlan_fdb_replace(struct vxlan_fdb *f,
54bfd872 676 union vxlan_addr *ip, __be16 port, __be32 vni,
ccdfd4f7 677 __u32 ifindex, struct vxlan_rdst *oldrd)
906dc186
TR		 678{
679 struct vxlan_rdst *rd;
680
681 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
682 if (rd)
683 return 0;
684
685 rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list);
686 if (!rd)
687 return 0;
0c1d70af 688
ccdfd4f7 689 *oldrd = *rd;
0c1d70af 690 dst_cache_reset(&rd->dst_cache);
e4c7ed41 691 rd->remote_ip = *ip;
906dc186
TR		 692 rd->remote_port = port;
693 rd->remote_vni = vni;
694 rd->remote_ifindex = ifindex;
6ad0b5a4 695 rd->offloaded = false;
906dc186
TR		 696 return 1;
697}
698
a5e7c10a
MR		 699/* Add/update destinations for multicast */
700static int vxlan_fdb_append(struct vxlan_fdb *f,
54bfd872 701 union vxlan_addr *ip, __be16 port, __be32 vni,
9e4b93f9 702 __u32 ifindex, struct vxlan_rdst **rdp)
a5e7c10a
MR		 703{
704 struct vxlan_rdst *rd;
705
706 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
707 if (rd)
708 return 0;
709
6681712d
DS		 710 rd = kmalloc(sizeof(*rd), GFP_ATOMIC);
711 if (rd == NULL)
712 return -ENOBUFS;
0c1d70af
PA		 713
714 if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) {
715 kfree(rd);
716 return -ENOBUFS;
717 }
718
e4c7ed41 719 rd->remote_ip = *ip;
6681712d 720 rd->remote_port = port;
0efe1173 721 rd->offloaded = false;
6681712d
DS		 722 rd->remote_vni = vni;
723 rd->remote_ifindex = ifindex;
3e61aa8f
SH		 724
725 list_add_tail_rcu(&rd->list, &f->remotes);
726
9e4b93f9 727 *rdp = rd;
6681712d
DS		 728 return 1;
729}
730
dfd8645e
TH		 731static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb,
732 unsigned int off,
733 struct vxlanhdr *vh, size_t hdrlen,
54bfd872
JB		 734 __be32 vni_field,
735 struct gro_remcsum *grc,
0ace2ca8 736 bool nopartial)
dfd8645e 737{
b7fe10e5 738 size_t start, offset;
dfd8645e
TH		 739
740 if (skb->remcsum_offload)
b7fe10e5 741 return vh;
dfd8645e
TH		 742
743 if (!NAPI_GRO_CB(skb)->csum_valid)
744 return NULL;
745
54bfd872
JB		 746 start = vxlan_rco_start(vni_field);
747 offset = start + vxlan_rco_offset(vni_field);
dfd8645e 748
b7fe10e5
TH		 749 vh = skb_gro_remcsum_process(skb, (void *)vh, off, hdrlen,
750 start, offset, grc, nopartial);
dfd8645e
TH		 751
752 skb->remcsum_offload = 1;
753
754 return vh;
755}
756
d4546c25
DM		 757static struct sk_buff *vxlan_gro_receive(struct sock *sk,
758 struct list_head *head,
759 struct sk_buff *skb)
dc01e7d3 760{
d4546c25
DM		 761 struct sk_buff *pp = NULL;
762 struct sk_buff *p;
dc01e7d3 763 struct vxlanhdr *vh, *vh2;
9b174d88 764 unsigned int hlen, off_vx;
dc01e7d3 765 int flush = 1;
5602c48c 766 struct vxlan_sock *vs = rcu_dereference_sk_user_data(sk);
54bfd872 767 __be32 flags;
26c4f7da
TH		 768 struct gro_remcsum grc;
769
770 skb_gro_remcsum_init(&grc);
dc01e7d3
OG		 771
772 off_vx = skb_gro_offset(skb);
773 hlen = off_vx + sizeof(*vh);
774 vh = skb_gro_header_fast(skb, off_vx);
775 if (skb_gro_header_hard(skb, hlen)) {
776 vh = skb_gro_header_slow(skb, hlen, off_vx);
777 if (unlikely(!vh))
778 goto out;
779 }
dc01e7d3 780
dfd8645e
TH		 781 skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr));
782
54bfd872 783 flags = vh->vx_flags;
dfd8645e
TH		 784
785 if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) {
786 vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr),
54bfd872 787 vh->vx_vni, &grc,
0ace2ca8
TH		 788 !!(vs->flags &
789 VXLAN_F_REMCSUM_NOPARTIAL));
dfd8645e
TH		 790
791 if (!vh)
792 goto out;
793 }
794
b7fe10e5
TH		 795 skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */
796
d4546c25 797 list_for_each_entry(p, head, list) {
dc01e7d3
OG		 798 if (!NAPI_GRO_CB(p)->same_flow)
799 continue;
800
801 vh2 = (struct vxlanhdr *)(p->data + off_vx);
3511494c
TG		 802 if (vh->vx_flags != vh2->vx_flags ||
803 vh->vx_vni != vh2->vx_vni) {
dc01e7d3
OG		 804 NAPI_GRO_CB(p)->same_flow = 0;
805 continue;
806 }
dc01e7d3
OG		 807 }
808
fcd91dd4 809 pp = call_gro_receive(eth_gro_receive, head, skb);
c194cf93 810 flush = 0;
dc01e7d3 811
dc01e7d3 812out:
603d4cf8 813 skb_gro_flush_final_remcsum(skb, pp, flush, &grc);
dc01e7d3
OG		 814
815 return pp;
816}
817
5602c48c 818static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
dc01e7d3 819{
229740c6
JR		 820 /* Sets 'skb->inner_mac_header' since we are always called with
821 * 'skb->encapsulation' set.
822 */
9b174d88 823 return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr));
dc01e7d3
OG		 824}
825
c7cdbe2e
RP		 826static struct vxlan_fdb *vxlan_fdb_alloc(struct vxlan_dev *vxlan, const u8 *mac,
827 __u16 state, __be32 src_vni,
828 __u16 ndm_flags)
25e20e73
RP		 829{
830 struct vxlan_fdb *f;
831
832 f = kmalloc(sizeof(*f), GFP_ATOMIC);
833 if (!f)
834 return NULL;
835 f->state = state;
836 f->flags = ndm_flags;
837 f->updated = f->used = jiffies;
838 f->vni = src_vni;
1274e1cc 839 f->nh = NULL;
79472fe8 840 RCU_INIT_POINTER(f->vdev, vxlan);
1274e1cc 841 INIT_LIST_HEAD(&f->nh_list);
25e20e73
RP		 842 INIT_LIST_HEAD(&f->remotes);
843 memcpy(f->eth_addr, mac, ETH_ALEN);
844
845 return f;
846}
847
7c31e54a
TY		 848static void vxlan_fdb_insert(struct vxlan_dev *vxlan, const u8 *mac,
849 __be32 src_vni, struct vxlan_fdb *f)
850{
851 ++vxlan->addrcnt;
852 hlist_add_head_rcu(&f->hlist,
853 vxlan_fdb_head(vxlan, mac, src_vni));
854}
855
1274e1cc
RP		 856static int vxlan_fdb_nh_update(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
857 u32 nhid, struct netlink_ext_ack *extack)
858{
859 struct nexthop *old_nh = rtnl_dereference(fdb->nh);
1274e1cc
RP		 860 struct nexthop *nh;
861 int err = -EINVAL;
862
863 if (old_nh && old_nh->id == nhid)
864 return 0;
865
866 nh = nexthop_find_by_id(vxlan->net, nhid);
867 if (!nh) {
868 NL_SET_ERR_MSG(extack, "Nexthop id does not exist");
869 goto err_inval;
870 }
871
872 if (nh) {
873 if (!nexthop_get(nh)) {
874 NL_SET_ERR_MSG(extack, "Nexthop has been deleted");
875 nh = NULL;
876 goto err_inval;
877 }
ce9ac056 878 if (!nexthop_is_fdb(nh)) {
1274e1cc
RP		 879 NL_SET_ERR_MSG(extack, "Nexthop is not a fdb nexthop");
880 goto err_inval;
881 }
882
50cb8769 883 if (!nexthop_is_multipath(nh)) {
1274e1cc
RP		 884 NL_SET_ERR_MSG(extack, "Nexthop is not a multipath group");
885 goto err_inval;
886 }
887
888 /* check nexthop group family */
1274e1cc
RP		 889 switch (vxlan->default_dst.remote_ip.sa.sa_family) {
890 case AF_INET:
50cb8769 891 if (!nexthop_has_v4(nh)) {
1274e1cc
RP		 892 err = -EAFNOSUPPORT;
893 NL_SET_ERR_MSG(extack, "Nexthop group family not supported");
894 goto err_inval;
895 }
896 break;
897 case AF_INET6:
50cb8769 898 if (nexthop_has_v4(nh)) {
1274e1cc
RP		 899 err = -EAFNOSUPPORT;
900 NL_SET_ERR_MSG(extack, "Nexthop group family not supported");
901 goto err_inval;
902 }
903 }
904 }
905
906 if (old_nh) {
907 list_del_rcu(&fdb->nh_list);
908 nexthop_put(old_nh);
909 }
910 rcu_assign_pointer(fdb->nh, nh);
911 list_add_tail_rcu(&fdb->nh_list, &nh->fdb_list);
912 return 1;
913
914err_inval:
915 if (nh)
916 nexthop_put(nh);
917 return err;
918}
919
d342894c 920static int vxlan_fdb_create(struct vxlan_dev *vxlan,
25e20e73
RP		 921 const u8 *mac, union vxlan_addr *ip,
922 __u16 state, __be16 port, __be32 src_vni,
45598c1c 923 __be32 vni, __u32 ifindex, __u16 ndm_flags,
1274e1cc
RP		 924 u32 nhid, struct vxlan_fdb **fdb,
925 struct netlink_ext_ack *extack)
25e20e73
RP		 926{
927 struct vxlan_rdst *rd = NULL;
928 struct vxlan_fdb *f;
929 int rc;
930
931 if (vxlan->cfg.addrmax &&
932 vxlan->addrcnt >= vxlan->cfg.addrmax)
933 return -ENOSPC;
934
935 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
c7cdbe2e 936 f = vxlan_fdb_alloc(vxlan, mac, state, src_vni, ndm_flags);
25e20e73
RP		 937 if (!f)
938 return -ENOMEM;
939
1274e1cc
RP		 940 if (nhid)
941 rc = vxlan_fdb_nh_update(vxlan, f, nhid, extack);
942 else
943 rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
944 if (rc < 0)
945 goto errout;
25e20e73 946
25e20e73
RP		 947 *fdb = f;
948
949 return 0;
1274e1cc
RP		 950
951errout:
952 kfree(f);
953 return rc;
25e20e73
RP		 954}
955
7c31e54a 956static void __vxlan_fdb_free(struct vxlan_fdb *f)
c2b200e0 957{
c2b200e0 958 struct vxlan_rdst *rd, *nd;
1274e1cc
RP		 959 struct nexthop *nh;
960
961 nh = rcu_dereference_raw(f->nh);
962 if (nh) {
963 rcu_assign_pointer(f->nh, NULL);
79472fe8 964 rcu_assign_pointer(f->vdev, NULL);
1274e1cc
RP		 965 nexthop_put(nh);
966 }
c2b200e0
PM		 967
968 list_for_each_entry_safe(rd, nd, &f->remotes, list) {
969 dst_cache_destroy(&rd->dst_cache);
970 kfree(rd);
971 }
972 kfree(f);
973}
974
7c31e54a
TY		 975static void vxlan_fdb_free(struct rcu_head *head)
976{
977 struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu);
978
979 __vxlan_fdb_free(f);
980}
981
c2b200e0
PM		 982static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f,
983 bool do_notify, bool swdev_notify)
984{
985 struct vxlan_rdst *rd;
986
987 netdev_dbg(vxlan->dev, "delete %pM\n", f->eth_addr);
988
989 --vxlan->addrcnt;
1274e1cc
RP		 990 if (do_notify) {
991 if (rcu_access_pointer(f->nh))
992 vxlan_fdb_notify(vxlan, f, NULL, RTM_DELNEIGH,
4c59b7d1 993 swdev_notify, NULL);
1274e1cc
RP		 994 else
995 list_for_each_entry(rd, &f->remotes, list)
996 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH,
997 swdev_notify, NULL);
998 }
c2b200e0
PM		 999
1000 hlist_del_rcu(&f->hlist);
79472fe8 1001 list_del_rcu(&f->nh_list);
c2b200e0
PM		 1002 call_rcu(&f->rcu, vxlan_fdb_free);
1003}
1004
fc4aa1ca
PM		 1005static void vxlan_dst_free(struct rcu_head *head)
1006{
1007 struct vxlan_rdst *rd = container_of(head, struct vxlan_rdst, rcu);
1008
1009 dst_cache_destroy(&rd->dst_cache);
1010 kfree(rd);
1011}
1012
a76d1ca2
PM		 1013static int vxlan_fdb_update_existing(struct vxlan_dev *vxlan,
1014 union vxlan_addr *ip,
1015 __u16 state, __u16 flags,
1016 __be16 port, __be32 vni,
1017 __u32 ifindex, __u16 ndm_flags,
1274e1cc 1018 struct vxlan_fdb *f, u32 nhid,
4c59b7d1
PM		 1019 bool swdev_notify,
1020 struct netlink_ext_ack *extack)
d342894c 1021{
45598c1c 1022 __u16 fdb_flags = (ndm_flags & ~NTF_USE);
9e4b93f9 1023 struct vxlan_rdst *rd = NULL;
ccdfd4f7 1024 struct vxlan_rdst oldrd;
d342894c 1025 int notify = 0;
61f46fe8
PM		 1026 int rc = 0;
1027 int err;
d342894c 1028
1274e1cc
RP		 1029 if (nhid && !rcu_access_pointer(f->nh)) {
1030 NL_SET_ERR_MSG(extack,
1031 "Cannot replace an existing non nexthop fdb with a nexthop");
1032 return -EOPNOTSUPP;
1033 }
1034
1035 if (nhid && (flags & NLM_F_APPEND)) {
1036 NL_SET_ERR_MSG(extack,
1037 "Cannot append to a nexthop fdb");
1038 return -EOPNOTSUPP;
1039 }
1040
a76d1ca2
PM		 1041 /* Do not allow an externally learned entry to take over an entry added
1042 * by the user.
1043 */
1044 if (!(fdb_flags & NTF_EXT_LEARNED) ||
1045 !(f->flags & NTF_VXLAN_ADDED_BY_USER)) {
1046 if (f->state != state) {
1047 f->state = state;
1048 f->updated = jiffies;
1049 notify = 1;
ae884082 1050 }
a76d1ca2
PM		 1051 if (f->flags != fdb_flags) {
1052 f->flags = fdb_flags;
1053 f->updated = jiffies;
1054 notify = 1;
906dc186 1055 }
a76d1ca2 1056 }
6681712d 1057
a76d1ca2
PM		 1058 if ((flags & NLM_F_REPLACE)) {
1059 /* Only change unicasts */
1060 if (!(is_multicast_ether_addr(f->eth_addr) ||
1061 is_zero_ether_addr(f->eth_addr))) {
1274e1cc
RP		 1062 if (nhid) {
1063 rc = vxlan_fdb_nh_update(vxlan, f, nhid, extack);
1064 if (rc < 0)
1065 return rc;
1066 } else {
1067 rc = vxlan_fdb_replace(f, ip, port, vni,
1068 ifindex, &oldrd);
1069 }
6681712d 1070 notify |= rc;
a76d1ca2 1071 } else {
1274e1cc 1072 NL_SET_ERR_MSG(extack, "Cannot replace non-unicast fdb entries");
906dc186 1073 return -EOPNOTSUPP;
a76d1ca2
PM		 1074 }
1075 }
1076 if ((flags & NLM_F_APPEND) &&
1077 (is_multicast_ether_addr(f->eth_addr) ||
1078 is_zero_ether_addr(f->eth_addr))) {
1079 rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
906dc186 1080
25e20e73 1081 if (rc < 0)
17b46365 1082 return rc;
a76d1ca2 1083 notify |= rc;
d342894c 1084 }
1085
a76d1ca2
PM		 1086 if (ndm_flags & NTF_USE)
1087 f->used = jiffies;
1088
9e4b93f9
ND		 1089 if (notify) {
1090 if (rd == NULL)
1091 rd = first_remote_rtnl(f);
a76d1ca2 1092
61f46fe8 1093 err = vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH,
4c59b7d1 1094 swdev_notify, extack);
61f46fe8
PM		 1095 if (err)
1096 goto err_notify;
9e4b93f9 1097 }
d342894c 1098
1099 return 0;
61f46fe8
PM		 1100
1101err_notify:
1274e1cc
RP		 1102 if (nhid)
1103 return err;
61f46fe8
PM		 1104 if ((flags & NLM_F_REPLACE) && rc)
1105 *rd = oldrd;
fc4aa1ca 1106 else if ((flags & NLM_F_APPEND) && rc) {
61f46fe8 1107 list_del_rcu(&rd->list);
fc4aa1ca
PM		 1108 call_rcu(&rd->rcu, vxlan_dst_free);
1109 }
61f46fe8 1110 return err;
d342894c 1111}
1112
a76d1ca2
PM		 1113static int vxlan_fdb_update_create(struct vxlan_dev *vxlan,
1114 const u8 *mac, union vxlan_addr *ip,
1115 __u16 state, __u16 flags,
1116 __be16 port, __be32 src_vni, __be32 vni,
1274e1cc 1117 __u32 ifindex, __u16 ndm_flags, u32 nhid,
4c59b7d1
PM		 1118 bool swdev_notify,
1119 struct netlink_ext_ack *extack)
a76d1ca2
PM		 1120{
1121 __u16 fdb_flags = (ndm_flags & ~NTF_USE);
1122 struct vxlan_fdb *f;
1123 int rc;
1124
1125 /* Disallow replace to add a multicast entry */
1126 if ((flags & NLM_F_REPLACE) &&
1127 (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac)))
1128 return -EOPNOTSUPP;
1129
1130 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
1131 rc = vxlan_fdb_create(vxlan, mac, ip, state, port, src_vni,
1274e1cc 1132 vni, ifindex, fdb_flags, nhid, &f, extack);
a76d1ca2
PM		 1133 if (rc < 0)
1134 return rc;
1135
7c31e54a 1136 vxlan_fdb_insert(vxlan, mac, src_vni, f);
61f46fe8 1137 rc = vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_NEWNEIGH,
4c59b7d1 1138 swdev_notify, extack);
61f46fe8
PM		 1139 if (rc)
1140 goto err_notify;
1141
a76d1ca2 1142 return 0;
61f46fe8
PM		 1143
1144err_notify:
1145 vxlan_fdb_destroy(vxlan, f, false, false);
1146 return rc;
a76d1ca2
PM		 1147}
1148
1149/* Add new entry to forwarding table -- assumes lock held */
1150static int vxlan_fdb_update(struct vxlan_dev *vxlan,
1151 const u8 *mac, union vxlan_addr *ip,
1152 __u16 state, __u16 flags,
1153 __be16 port, __be32 src_vni, __be32 vni,
1274e1cc 1154 __u32 ifindex, __u16 ndm_flags, u32 nhid,
4c59b7d1
PM		 1155 bool swdev_notify,
1156 struct netlink_ext_ack *extack)
a76d1ca2
PM		 1157{
1158 struct vxlan_fdb *f;
1159
1160 f = __vxlan_find_mac(vxlan, mac, src_vni);
1161 if (f) {
1162 if (flags & NLM_F_EXCL) {
1163 netdev_dbg(vxlan->dev,
1164 "lost race to create %pM\n", mac);
1165 return -EEXIST;
1166 }
1167
1168 return vxlan_fdb_update_existing(vxlan, ip, state, flags, port,
1169 vni, ifindex, ndm_flags, f,
1274e1cc 1170 nhid, swdev_notify, extack);
a76d1ca2
PM		 1171 } else {
1172 if (!(flags & NLM_F_CREATE))
1173 return -ENOENT;
1174
1175 return vxlan_fdb_update_create(vxlan, mac, ip, state, flags,
1176 port, src_vni, vni, ifindex,
1274e1cc
RP		 1177 ndm_flags, nhid, swdev_notify,
1178 extack);
a76d1ca2
PM		 1179 }
1180}
1181
35cf2845 1182static void vxlan_fdb_dst_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f,
0e6160f3 1183 struct vxlan_rdst *rd, bool swdev_notify)
35cf2845
LR		 1184{
1185 list_del_rcu(&rd->list);
4c59b7d1 1186 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH, swdev_notify, NULL);
35cf2845
LR		 1187 call_rcu(&rd->rcu, vxlan_dst_free);
1188}
1189
f0b074be 1190static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan,
3ad7a4b1 1191 union vxlan_addr *ip, __be16 *port, __be32 *src_vni,
1274e1cc 1192 __be32 *vni, u32 *ifindex, u32 *nhid)
d342894c 1193{
6681712d 1194 struct net *net = dev_net(vxlan->dev);
e4c7ed41 1195 int err;
d342894c 1196
72b48682
RP		 1197 if (tb[NDA_NH_ID] && (tb[NDA_DST] || tb[NDA_VNI] || tb[NDA_IFINDEX] ||
1198 tb[NDA_PORT]))
1199 return -EINVAL;
1200
f0b074be 1201 if (tb[NDA_DST]) {
e4c7ed41
CW		 1202 err = vxlan_nla_get_addr(ip, tb[NDA_DST]);
1203 if (err)
1204 return err;
f0b074be 1205 } else {
e4c7ed41 1206 union vxlan_addr *remote = &vxlan->default_dst.remote_ip;
98c81476 1207
e4c7ed41
CW		 1208 if (remote->sa.sa_family == AF_INET) {
1209 ip->sin.sin_addr.s_addr = htonl(INADDR_ANY);
1210 ip->sa.sa_family = AF_INET;
1211#if IS_ENABLED(CONFIG_IPV6)
1212 } else {
1213 ip->sin6.sin6_addr = in6addr_any;
1214 ip->sa.sa_family = AF_INET6;
1215#endif
1216 }
f0b074be 1217 }
d342894c 1218
6681712d 1219 if (tb[NDA_PORT]) {
73cf3317 1220 if (nla_len(tb[NDA_PORT]) != sizeof(__be16))
6681712d 1221 return -EINVAL;
f0b074be
MR		 1222 *port = nla_get_be16(tb[NDA_PORT]);
1223 } else {
0dfbdf41 1224 *port = vxlan->cfg.dst_port;
f0b074be 1225 }
6681712d
DS		 1226
1227 if (tb[NDA_VNI]) {
1228 if (nla_len(tb[NDA_VNI]) != sizeof(u32))
1229 return -EINVAL;
54bfd872 1230 *vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI]));
f0b074be
MR		 1231 } else {
1232 *vni = vxlan->default_dst.remote_vni;
1233 }
6681712d 1234
3ad7a4b1
RP		 1235 if (tb[NDA_SRC_VNI]) {
1236 if (nla_len(tb[NDA_SRC_VNI]) != sizeof(u32))
1237 return -EINVAL;
1238 *src_vni = cpu_to_be32(nla_get_u32(tb[NDA_SRC_VNI]));
1239 } else {
1240 *src_vni = vxlan->default_dst.remote_vni;
1241 }
1242
6681712d 1243 if (tb[NDA_IFINDEX]) {
5abb0029 1244 struct net_device *tdev;
6681712d
DS		 1245
1246 if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32))
1247 return -EINVAL;
f0b074be 1248 *ifindex = nla_get_u32(tb[NDA_IFINDEX]);
73763949 1249 tdev = __dev_get_by_index(net, *ifindex);
5abb0029 1250 if (!tdev)
6681712d 1251 return -EADDRNOTAVAIL;
f0b074be
MR		 1252 } else {
1253 *ifindex = 0;
1254 }
1255
1274e1cc
RP		 1256 if (tb[NDA_NH_ID])
1257 *nhid = nla_get_u32(tb[NDA_NH_ID]);
1258 else
1259 *nhid = 0;
1260
f0b074be
MR		 1261 return 0;
1262}
1263
1264/* Add static entry (via netlink) */
1265static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[],
1266 struct net_device *dev,
87b0984e
PM		 1267 const unsigned char *addr, u16 vid, u16 flags,
1268 struct netlink_ext_ack *extack)
f0b074be
MR		 1269{
1270 struct vxlan_dev *vxlan = netdev_priv(dev);
1271 /* struct net *net = dev_net(vxlan->dev); */
e4c7ed41 1272 union vxlan_addr ip;
f0b074be 1273 __be16 port;
3ad7a4b1 1274 __be32 src_vni, vni;
1274e1cc 1275 u32 ifindex, nhid;
fe1e0713 1276 u32 hash_index;
f0b074be
MR		 1277 int err;
1278
1279 if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) {
1280 pr_info("RTM_NEWNEIGH with invalid state %#x\n",
1281 ndm->ndm_state);
1282 return -EINVAL;
1283 }
1284
1274e1cc 1285 if (!tb || (!tb[NDA_DST] && !tb[NDA_NH_ID]))
f0b074be
MR		 1286 return -EINVAL;
1287
1274e1cc
RP		 1288 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &src_vni, &vni, &ifindex,
1289 &nhid);
f0b074be
MR		 1290 if (err)
1291 return err;
6681712d 1292
5933a7bb
MR		 1293 if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family)
1294 return -EAFNOSUPPORT;
1295
fe1e0713
L		 1296 hash_index = fdb_head_index(vxlan, addr, src_vni);
1297 spin_lock_bh(&vxlan->hash_lock[hash_index]);
25e20e73 1298 err = vxlan_fdb_update(vxlan, addr, &ip, ndm->ndm_state, flags,
45598c1c
PM		 1299 port, src_vni, vni, ifindex,
1300 ndm->ndm_flags | NTF_VXLAN_ADDED_BY_USER,
1274e1cc 1301 nhid, true, extack);
fe1e0713 1302 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
d342894c 1303
1304 return err;
1305}
1306
3ad7a4b1
RP		 1307static int __vxlan_fdb_delete(struct vxlan_dev *vxlan,
1308 const unsigned char *addr, union vxlan_addr ip,
fc39c38b 1309 __be16 port, __be32 src_vni, __be32 vni,
0e6160f3 1310 u32 ifindex, bool swdev_notify)
d342894c 1311{
bc7892ba 1312 struct vxlan_rdst *rd = NULL;
1274e1cc 1313 struct vxlan_fdb *f;
3ad7a4b1 1314 int err = -ENOENT;
bc7892ba 1315
3ad7a4b1 1316 f = vxlan_find_mac(vxlan, addr, src_vni);
bc7892ba 1317 if (!f)
3ad7a4b1 1318 return err;
bc7892ba 1319
e4c7ed41
CW		 1320 if (!vxlan_addr_any(&ip)) {
1321 rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex);
bc7892ba
MR		 1322 if (!rd)
1323 goto out;
1324 }
1325
bc7892ba
MR		 1326 /* remove a destination if it's not the only one on the list,
1327 * otherwise destroy the fdb entry
1328 */
1329 if (rd && !list_is_singular(&f->remotes)) {
0e6160f3 1330 vxlan_fdb_dst_destroy(vxlan, f, rd, swdev_notify);
bc7892ba 1331 goto out;
d342894c 1332 }
bc7892ba 1333
0e6160f3 1334 vxlan_fdb_destroy(vxlan, f, true, swdev_notify);
bc7892ba
MR		 1335
1336out:
3ad7a4b1
RP		 1337 return 0;
1338}
1339
1340/* Delete entry (via netlink) */
1341static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[],
1342 struct net_device *dev,
1343 const unsigned char *addr, u16 vid)
1344{
1345 struct vxlan_dev *vxlan = netdev_priv(dev);
1346 union vxlan_addr ip;
1347 __be32 src_vni, vni;
1274e1cc 1348 u32 ifindex, nhid;
fe1e0713 1349 u32 hash_index;
1274e1cc 1350 __be16 port;
3ad7a4b1
RP		 1351 int err;
1352
1274e1cc
RP		 1353 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &src_vni, &vni, &ifindex,
1354 &nhid);
3ad7a4b1
RP		 1355 if (err)
1356 return err;
1357
fe1e0713
L		 1358 hash_index = fdb_head_index(vxlan, addr, src_vni);
1359 spin_lock_bh(&vxlan->hash_lock[hash_index]);
0e6160f3
PM		 1360 err = __vxlan_fdb_delete(vxlan, addr, ip, port, src_vni, vni, ifindex,
1361 true);
fe1e0713 1362 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
d342894c 1363
1364 return err;
1365}
1366
1367/* Dump forwarding table */
1368static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb,
5d5eacb3 1369 struct net_device *dev,
d297653d 1370 struct net_device *filter_dev, int *idx)
d342894c 1371{
1372 struct vxlan_dev *vxlan = netdev_priv(dev);
1373 unsigned int h;
d297653d 1374 int err = 0;
d342894c 1375
1376 for (h = 0; h < FDB_HASH_SIZE; ++h) {
1377 struct vxlan_fdb *f;
d342894c 1378
b5141915 1379 rcu_read_lock();
b67bfe0d 1380 hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) {
6681712d 1381 struct vxlan_rdst *rd;
6681712d 1382
1274e1cc 1383 if (rcu_access_pointer(f->nh)) {
b18e9834
RP		 1384 if (*idx < cb->args[2])
1385 goto skip_nh;
1274e1cc
RP		 1386 err = vxlan_fdb_info(skb, vxlan, f,
1387 NETLINK_CB(cb->skb).portid,
1388 cb->nlh->nlmsg_seq,
1389 RTM_NEWNEIGH,
1390 NLM_F_MULTI, NULL);
b5141915
IS		 1391 if (err < 0) {
1392 rcu_read_unlock();
1274e1cc 1393 goto out;
b5141915 1394 }
b18e9834
RP		 1395skip_nh:
1396 *idx += 1;
1274e1cc
RP		 1397 continue;
1398 }
1399
3e61aa8f 1400 list_for_each_entry_rcu(rd, &f->remotes, list) {
d297653d 1401 if (*idx < cb->args[2])
07a51cd3
AW		 1402 goto skip;
1403
6681712d
DS		 1404 err = vxlan_fdb_info(skb, vxlan, f,
1405 NETLINK_CB(cb->skb).portid,
1406 cb->nlh->nlmsg_seq,
1407 RTM_NEWNEIGH,
1408 NLM_F_MULTI, rd);
b5141915
IS		 1409 if (err < 0) {
1410 rcu_read_unlock();
3e61aa8f 1411 goto out;
b5141915 1412 }
3e61aa8f 1413skip:
d297653d 1414 *idx += 1;
07a51cd3 1415 }
d342894c 1416 }
b5141915 1417 rcu_read_unlock();
d342894c 1418 }
3e61aa8f 1419out:
d297653d 1420 return err;
d342894c 1421}
1422
474c3c89
RP		 1423static int vxlan_fdb_get(struct sk_buff *skb,
1424 struct nlattr *tb[],
1425 struct net_device *dev,
1426 const unsigned char *addr,
1427 u16 vid, u32 portid, u32 seq,
1428 struct netlink_ext_ack *extack)
1429{
1430 struct vxlan_dev *vxlan = netdev_priv(dev);
1431 struct vxlan_fdb *f;
1432 __be32 vni;
1433 int err;
1434
1435 if (tb[NDA_VNI])
1436 vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI]));
1437 else
1438 vni = vxlan->default_dst.remote_vni;
1439
1440 rcu_read_lock();
1441
1442 f = __vxlan_find_mac(vxlan, addr, vni);
1443 if (!f) {
1444 NL_SET_ERR_MSG(extack, "Fdb entry not found");
1445 err = -ENOENT;
1446 goto errout;
1447 }
1448
1449 err = vxlan_fdb_info(skb, vxlan, f, portid, seq,
1450 RTM_NEWNEIGH, 0, first_remote_rcu(f));
1451errout:
1452 rcu_read_unlock();
1453 return err;
1454}
1455
d342894c 1456/* Watch incoming packets to learn mapping between Ethernet address
1457 * and Tunnel endpoint.
c4b49512 1458 * Return true if packet is bogus and should be dropped.
d342894c 1459 */
26a41ae6 1460static bool vxlan_snoop(struct net_device *dev,
3ad7a4b1 1461 union vxlan_addr *src_ip, const u8 *src_mac,
87613de9 1462 u32 src_ifindex, __be32 vni)
d342894c 1463{
1464 struct vxlan_dev *vxlan = netdev_priv(dev);
1465 struct vxlan_fdb *f;
87613de9
MS		 1466 u32 ifindex = 0;
1467
1468#if IS_ENABLED(CONFIG_IPV6)
1469 if (src_ip->sa.sa_family == AF_INET6 &&
1470 (ipv6_addr_type(&src_ip->sin6.sin6_addr) & IPV6_ADDR_LINKLOCAL))
1471 ifindex = src_ifindex;
1472#endif
d342894c 1473
3ad7a4b1 1474 f = vxlan_find_mac(vxlan, src_mac, vni);
d342894c 1475 if (likely(f)) {
5ca5461c 1476 struct vxlan_rdst *rdst = first_remote_rcu(f);
3e61aa8f 1477
87613de9
MS		 1478 if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip) &&
1479 rdst->remote_ifindex == ifindex))
26a41ae6 1480 return false;
1481
1482 /* Don't migrate static entries, drop packets */
e0090a9e 1483 if (f->state & (NUD_PERMANENT | NUD_NOARP))
26a41ae6 1484 return true;
d342894c 1485
1274e1cc
RP		 1486 /* Don't override an fdb with nexthop with a learnt entry */
1487 if (rcu_access_pointer(f->nh))
1488 return true;
1489
d342894c 1490 if (net_ratelimit())
1491 netdev_info(dev,
e4c7ed41 1492 "%pM migrated from %pIS to %pIS\n",
a4870f79 1493 src_mac, &rdst->remote_ip.sa, &src_ip->sa);
d342894c 1494
e4c7ed41 1495 rdst->remote_ip = *src_ip;
d342894c 1496 f->updated = jiffies;
4c59b7d1 1497 vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH, true, NULL);
d342894c 1498 } else {
fe1e0713
L		 1499 u32 hash_index = fdb_head_index(vxlan, src_mac, vni);
1500
d342894c 1501 /* learned new entry */
fe1e0713 1502 spin_lock(&vxlan->hash_lock[hash_index]);
3bf74b1a 1503
1504 /* close off race between vxlan_flush and incoming packets */
1505 if (netif_running(dev))
25e20e73 1506 vxlan_fdb_update(vxlan, src_mac, src_ip,
3bf74b1a 1507 NUD_REACHABLE,
1508 NLM_F_EXCL|NLM_F_CREATE,
0dfbdf41 1509 vxlan->cfg.dst_port,
3ad7a4b1 1510 vni,
3bf74b1a 1511 vxlan->default_dst.remote_vni,
1274e1cc 1512 ifindex, NTF_SELF, 0, true, NULL);
fe1e0713 1513 spin_unlock(&vxlan->hash_lock[hash_index]);
d342894c 1514 }
26a41ae6 1515
1516 return false;
d342894c 1517}
1518
d342894c 1519/* See if multicast group is already in use by other ID */
95ab0991 1520static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev)
d342894c 1521{
553675fb 1522 struct vxlan_dev *vxlan;
c6fcc4fc 1523 struct vxlan_sock *sock4;
4053ab1b
AB		 1524#if IS_ENABLED(CONFIG_IPV6)
1525 struct vxlan_sock *sock6;
1526#endif
b1be00a6 1527 unsigned short family = dev->default_dst.remote_ip.sa.sa_family;
d342894c 1528
c6fcc4fc 1529 sock4 = rtnl_dereference(dev->vn4_sock);
1530
95ab0991
G		 1531 /* The vxlan_sock is only used by dev, leaving group has
1532 * no effect on other vxlan devices.
1533 */
66af846f 1534 if (family == AF_INET && sock4 && refcount_read(&sock4->refcnt) == 1)
95ab0991 1535 return false;
b1be00a6 1536#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 1537 sock6 = rtnl_dereference(dev->vn6_sock);
66af846f 1538 if (family == AF_INET6 && sock6 && refcount_read(&sock6->refcnt) == 1)
b1be00a6
JB		 1539 return false;
1540#endif
95ab0991 1541
553675fb 1542 list_for_each_entry(vxlan, &vn->vxlan_list, next) {
95ab0991 1543 if (!netif_running(vxlan->dev) || vxlan == dev)
553675fb 1544 continue;
d342894c 1545
c6fcc4fc 1546 if (family == AF_INET &&
1547 rtnl_dereference(vxlan->vn4_sock) != sock4)
95ab0991 1548 continue;
b1be00a6 1549#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 1550 if (family == AF_INET6 &&
1551 rtnl_dereference(vxlan->vn6_sock) != sock6)
b1be00a6
JB		 1552 continue;
1553#endif
95ab0991
G		 1554
1555 if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip,
1556 &dev->default_dst.remote_ip))
1557 continue;
1558
1559 if (vxlan->default_dst.remote_ifindex !=
1560 dev->default_dst.remote_ifindex)
1561 continue;
1562
1563 return true;
553675fb 1564 }
d342894c 1565
1566 return false;
1567}
1568
544a773a 1569static bool __vxlan_sock_release_prep(struct vxlan_sock *vs)
7c47cedf 1570{
b1be00a6 1571 struct vxlan_net *vn;
012a5729 1572
b1be00a6 1573 if (!vs)
544a773a 1574 return false;
66af846f 1575 if (!refcount_dec_and_test(&vs->refcnt))
544a773a 1576 return false;
d342894c 1577
b1be00a6 1578 vn = net_generic(sock_net(vs->sock->sk), vxlan_net_id);
1c51a915 1579 spin_lock(&vn->sock_lock);
7c47cedf 1580 hlist_del_rcu(&vs->hlist);
e7b3db5e 1581 udp_tunnel_notify_del_rx_port(vs->sock,
b9adcd69
AD		 1582 (vs->flags & VXLAN_F_GPE) ?
1583 UDP_TUNNEL_TYPE_VXLAN_GPE :
e7b3db5e 1584 UDP_TUNNEL_TYPE_VXLAN);
1c51a915
SH		 1585 spin_unlock(&vn->sock_lock);
1586
544a773a 1587 return true;
d342894c 1588}
1589
b1be00a6
JB		 1590static void vxlan_sock_release(struct vxlan_dev *vxlan)
1591{
c6fcc4fc 1592 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
b1be00a6 1593#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 1594 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1595
57d88182 1596 RCU_INIT_POINTER(vxlan->vn6_sock, NULL);
544a773a
HFS		 1597#endif
1598
57d88182 1599 RCU_INIT_POINTER(vxlan->vn4_sock, NULL);
544a773a
HFS		 1600 synchronize_net();
1601
a53cb29b
MB		 1602 vxlan_vs_del_dev(vxlan);
1603
c6fcc4fc 1604 if (__vxlan_sock_release_prep(sock4)) {
1605 udp_tunnel_sock_release(sock4->sock);
1606 kfree(sock4);
544a773a
HFS		 1607 }
1608
1609#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 1610 if (__vxlan_sock_release_prep(sock6)) {
1611 udp_tunnel_sock_release(sock6->sock);
1612 kfree(sock6);
544a773a 1613 }
b1be00a6
JB		 1614#endif
1615}
1616
56ef9c90 1617/* Update multicast group membership when first VNI on
c4b49512 1618 * multicast address is brought up
7c47cedf 1619 */
56ef9c90 1620static int vxlan_igmp_join(struct vxlan_dev *vxlan)
d342894c 1621{
b1be00a6 1622 struct sock *sk;
e4c7ed41
CW		 1623 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1624 int ifindex = vxlan->default_dst.remote_ifindex;
149d7549 1625 int ret = -EINVAL;
d342894c 1626
e4c7ed41 1627 if (ip->sa.sa_family == AF_INET) {
c6fcc4fc 1628 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
e4c7ed41
CW		 1629 struct ip_mreqn mreq = {
1630 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1631 .imr_ifindex = ifindex,
1632 };
1633
c6fcc4fc 1634 sk = sock4->sock->sk;
b1be00a6 1635 lock_sock(sk);
56ef9c90 1636 ret = ip_mc_join_group(sk, &mreq);
b1be00a6 1637 release_sock(sk);
e4c7ed41
CW		 1638#if IS_ENABLED(CONFIG_IPV6)
1639 } else {
c6fcc4fc 1640 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1641
1642 sk = sock6->sock->sk;
b1be00a6 1643 lock_sock(sk);
56ef9c90
MRL		 1644 ret = ipv6_stub->ipv6_sock_mc_join(sk, ifindex,
1645 &ip->sin6.sin6_addr);
b1be00a6 1646 release_sock(sk);
e4c7ed41
CW		 1647#endif
1648 }
3fc2de2f 1649
56ef9c90 1650 return ret;
3fc2de2f 1651}
1652
1653/* Inverse of vxlan_igmp_join when last VNI is brought down */
56ef9c90 1654static int vxlan_igmp_leave(struct vxlan_dev *vxlan)
3fc2de2f 1655{
b1be00a6 1656 struct sock *sk;
e4c7ed41
CW		 1657 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1658 int ifindex = vxlan->default_dst.remote_ifindex;
149d7549 1659 int ret = -EINVAL;
3fc2de2f 1660
e4c7ed41 1661 if (ip->sa.sa_family == AF_INET) {
c6fcc4fc 1662 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
e4c7ed41
CW		 1663 struct ip_mreqn mreq = {
1664 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1665 .imr_ifindex = ifindex,
1666 };
1667
c6fcc4fc 1668 sk = sock4->sock->sk;
b1be00a6 1669 lock_sock(sk);
56ef9c90 1670 ret = ip_mc_leave_group(sk, &mreq);
b1be00a6 1671 release_sock(sk);
e4c7ed41
CW		 1672#if IS_ENABLED(CONFIG_IPV6)
1673 } else {
c6fcc4fc 1674 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1675
1676 sk = sock6->sock->sk;
b1be00a6 1677 lock_sock(sk);
56ef9c90
MRL		 1678 ret = ipv6_stub->ipv6_sock_mc_drop(sk, ifindex,
1679 &ip->sin6.sin6_addr);
b1be00a6 1680 release_sock(sk);
e4c7ed41
CW		 1681#endif
1682 }
d342894c 1683
56ef9c90 1684 return ret;
d342894c 1685}
1686
f14ecebb
JB		 1687static bool vxlan_remcsum(struct vxlanhdr *unparsed,
1688 struct sk_buff *skb, u32 vxflags)
dfd8645e 1689{
7d34fa75 1690 size_t start, offset;
dfd8645e 1691
f14ecebb
JB		 1692 if (!(unparsed->vx_flags & VXLAN_HF_RCO) || skb->remcsum_offload)
1693 goto out;
b7fe10e5 1694
f14ecebb
JB		 1695 start = vxlan_rco_start(unparsed->vx_vni);
1696 offset = start + vxlan_rco_offset(unparsed->vx_vni);
dfd8645e 1697
7d34fa75 1698 if (!pskb_may_pull(skb, offset + sizeof(u16)))
be5cfeab 1699 return false;
dfd8645e 1700
be5cfeab
JB		 1701 skb_remcsum_process(skb, (void *)(vxlan_hdr(skb) + 1), start, offset,
1702 !!(vxflags & VXLAN_F_REMCSUM_NOPARTIAL));
f14ecebb
JB		 1703out:
1704 unparsed->vx_flags &= ~VXLAN_HF_RCO;
1705 unparsed->vx_vni &= VXLAN_VNI_MASK;
be5cfeab 1706 return true;
dfd8645e
TH		 1707}
1708
f14ecebb 1709static void vxlan_parse_gbp_hdr(struct vxlanhdr *unparsed,
64f87d36 1710 struct sk_buff *skb, u32 vxflags,
10a5af23 1711 struct vxlan_metadata *md)
3288af08 1712{
f14ecebb 1713 struct vxlanhdr_gbp *gbp = (struct vxlanhdr_gbp *)unparsed;
10a5af23 1714 struct metadata_dst *tun_dst;
f14ecebb
JB		 1715
1716 if (!(unparsed->vx_flags & VXLAN_HF_GBP))
1717 goto out;
3288af08 1718
3288af08
JB		 1719 md->gbp = ntohs(gbp->policy_id);
1720
10a5af23 1721 tun_dst = (struct metadata_dst *)skb_dst(skb);
810813c4 1722 if (tun_dst) {
3288af08 1723 tun_dst->u.tun_info.key.tun_flags |= TUNNEL_VXLAN_OPT;
810813c4
DM		 1724 tun_dst->u.tun_info.options_len = sizeof(*md);
1725 }
3288af08
JB		 1726 if (gbp->dont_learn)
1727 md->gbp |= VXLAN_GBP_DONT_LEARN;
1728
1729 if (gbp->policy_applied)
1730 md->gbp |= VXLAN_GBP_POLICY_APPLIED;
f14ecebb 1731
64f87d36
JB		 1732 /* In flow-based mode, GBP is carried in dst_metadata */
1733 if (!(vxflags & VXLAN_F_COLLECT_METADATA))
1734 skb->mark = md->gbp;
f14ecebb
JB		 1735out:
1736 unparsed->vx_flags &= ~VXLAN_GBP_USED_BITS;
3288af08
JB		 1737}
1738
e1e5314d 1739static bool vxlan_parse_gpe_hdr(struct vxlanhdr *unparsed,
61618eea 1740 __be16 *protocol,
e1e5314d
JB		 1741 struct sk_buff *skb, u32 vxflags)
1742{
1743 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)unparsed;
1744
1745 /* Need to have Next Protocol set for interfaces in GPE mode. */
1746 if (!gpe->np_applied)
1747 return false;
1748 /* "The initial version is 0. If a receiver does not support the
1749 * version indicated it MUST drop the packet.
1750 */
1751 if (gpe->version != 0)
1752 return false;
1753 /* "When the O bit is set to 1, the packet is an OAM packet and OAM
1754 * processing MUST occur." However, we don't implement OAM
1755 * processing, thus drop the packet.
1756 */
1757 if (gpe->oam_flag)
1758 return false;
1759
fa20e0e3
JB		 1760 *protocol = tun_p_to_eth_p(gpe->next_protocol);
1761 if (!*protocol)
e1e5314d 1762 return false;
e1e5314d
JB		 1763
1764 unparsed->vx_flags &= ~VXLAN_GPE_USED_BITS;
1765 return true;
1766}
1767
1ab016e2
JB		 1768static bool vxlan_set_mac(struct vxlan_dev *vxlan,
1769 struct vxlan_sock *vs,
3ad7a4b1 1770 struct sk_buff *skb, __be32 vni)
614732ea 1771{
614732ea 1772 union vxlan_addr saddr;
87613de9 1773 u32 ifindex = skb->dev->ifindex;
614732ea 1774
614732ea 1775 skb_reset_mac_header(skb);
614732ea
TG		 1776 skb->protocol = eth_type_trans(skb, vxlan->dev);
1777 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1778
1779 /* Ignore packet loops (and multicast echo) */
1780 if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr))
1ab016e2 1781 return false;
614732ea 1782
760c6805 1783 /* Get address from the outer IP header */
ce212d0f 1784 if (vxlan_get_sk_family(vs) == AF_INET) {
1ab016e2 1785 saddr.sin.sin_addr.s_addr = ip_hdr(skb)->saddr;
614732ea
TG		 1786 saddr.sa.sa_family = AF_INET;
1787#if IS_ENABLED(CONFIG_IPV6)
1788 } else {
1ab016e2 1789 saddr.sin6.sin6_addr = ipv6_hdr(skb)->saddr;
614732ea
TG		 1790 saddr.sa.sa_family = AF_INET6;
1791#endif
1792 }
1793
dc5321d7 1794 if ((vxlan->cfg.flags & VXLAN_F_LEARN) &&
87613de9 1795 vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source, ifindex, vni))
1ab016e2
JB		 1796 return false;
1797
1798 return true;
1799}
1800
760c6805
JB		 1801static bool vxlan_ecn_decapsulate(struct vxlan_sock *vs, void *oiph,
1802 struct sk_buff *skb)
1803{
1804 int err = 0;
1805
1806 if (vxlan_get_sk_family(vs) == AF_INET)
1807 err = IP_ECN_decapsulate(oiph, skb);
1808#if IS_ENABLED(CONFIG_IPV6)
1809 else
1810 err = IP6_ECN_decapsulate(oiph, skb);
1811#endif
1812
1813 if (unlikely(err) && log_ecn_error) {
1814 if (vxlan_get_sk_family(vs) == AF_INET)
1815 net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n",
1816 &((struct iphdr *)oiph)->saddr,
1817 ((struct iphdr *)oiph)->tos);
1818 else
1819 net_info_ratelimited("non-ECT from %pI6\n",
1820 &((struct ipv6hdr *)oiph)->saddr);
1821 }
1822 return err <= 1;
1823}
1824
d342894c 1825/* Callback from net/ipv4/udp.c to receive packets */
f2d1968e 1826static int vxlan_rcv(struct sock *sk, struct sk_buff *skb)
d342894c 1827{
f2d1968e 1828 struct pcpu_sw_netstats *stats;
c9e78efb 1829 struct vxlan_dev *vxlan;
5cfccc5a 1830 struct vxlan_sock *vs;
f14ecebb 1831 struct vxlanhdr unparsed;
ee122c79
TG		 1832 struct vxlan_metadata _md;
1833 struct vxlan_metadata *md = &_md;
61618eea 1834 __be16 protocol = htons(ETH_P_TEB);
e1e5314d 1835 bool raw_proto = false;
f2d1968e 1836 void *oiph;
3ad7a4b1 1837 __be32 vni = 0;
d342894c 1838
e1e5314d 1839 /* Need UDP and VXLAN header to be present */
7ce04758 1840 if (!pskb_may_pull(skb, VXLAN_HLEN))
e5aed006 1841 goto drop;
d342894c 1842
f14ecebb 1843 unparsed = *vxlan_hdr(skb);
288b01c8
JB		 1844 /* VNI flag always required to be set */
1845 if (!(unparsed.vx_flags & VXLAN_HF_VNI)) {
1846 netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n",
1847 ntohl(vxlan_hdr(skb)->vx_flags),
1848 ntohl(vxlan_hdr(skb)->vx_vni));
1849 /* Return non vxlan pkt */
e5aed006 1850 goto drop;
d342894c 1851 }
288b01c8
JB		 1852 unparsed.vx_flags &= ~VXLAN_HF_VNI;
1853 unparsed.vx_vni &= ~VXLAN_VNI_MASK;
d342894c 1854
559835ea 1855 vs = rcu_dereference_sk_user_data(sk);
5cfccc5a 1856 if (!vs)
d342894c 1857 goto drop;
d342894c 1858
3ad7a4b1
RP		 1859 vni = vxlan_vni(vxlan_hdr(skb)->vx_vni);
1860
49f810f0 1861 vxlan = vxlan_vs_find_vni(vs, skb->dev->ifindex, vni);
c9e78efb
JB		 1862 if (!vxlan)
1863 goto drop;
1864
e1e5314d
JB		 1865 /* For backwards compatibility, only allow reserved fields to be
1866 * used by VXLAN extensions if explicitly requested.
1867 */
1868 if (vs->flags & VXLAN_F_GPE) {
1869 if (!vxlan_parse_gpe_hdr(&unparsed, &protocol, skb, vs->flags))
1870 goto drop;
1871 raw_proto = true;
1872 }
1873
1874 if (__iptunnel_pull_header(skb, VXLAN_HLEN, protocol, raw_proto,
1875 !net_eq(vxlan->net, dev_net(vxlan->dev))))
98c81476 1876 goto drop;
c9e78efb 1877
ee122c79 1878 if (vxlan_collect_metadata(vs)) {
10a5af23 1879 struct metadata_dst *tun_dst;
07dabf20 1880
c29a70d2 1881 tun_dst = udp_tun_rx_dst(skb, vxlan_get_sk_family(vs), TUNNEL_KEY,
d817f432 1882 key32_to_tunnel_id(vni), sizeof(*md));
c29a70d2 1883
ee122c79
TG		 1884 if (!tun_dst)
1885 goto drop;
1886
0f1b7354 1887 md = ip_tunnel_info_opts(&tun_dst->u.tun_info);
10a5af23
JB		 1888
1889 skb_dst_set(skb, (struct dst_entry *)tun_dst);
ee122c79
TG		 1890 } else {
1891 memset(md, 0, sizeof(*md));
1892 }
1893
f14ecebb
JB		 1894 if (vs->flags & VXLAN_F_REMCSUM_RX)
1895 if (!vxlan_remcsum(&unparsed, skb, vs->flags))
1896 goto drop;
1897 if (vs->flags & VXLAN_F_GBP)
10a5af23 1898 vxlan_parse_gbp_hdr(&unparsed, skb, vs->flags, md);
e1e5314d
JB		 1899 /* Note that GBP and GPE can never be active together. This is
1900 * ensured in vxlan_dev_configure.
1901 */
3511494c 1902
f14ecebb 1903 if (unparsed.vx_flags || unparsed.vx_vni) {
3bf39475
TH		 1904 /* If there are any unprocessed flags remaining treat
1905 * this as a malformed packet. This behavior diverges from
1906 * VXLAN RFC (RFC7348) which stipulates that bits in reserved
1907 * in reserved fields are to be ignored. The approach here
c4b49512 1908 * maintains compatibility with previous stack code, and also
3bf39475
TH		 1909 * is more robust and provides a little more security in
1910 * adding extensions to VXLAN.
1911 */
288b01c8 1912 goto drop;
3bf39475
TH		 1913 }
1914
e1e5314d 1915 if (!raw_proto) {
3ad7a4b1 1916 if (!vxlan_set_mac(vxlan, vs, skb, vni))
e1e5314d
JB		 1917 goto drop;
1918 } else {
8be0cfa4 1919 skb_reset_mac_header(skb);
e1e5314d
JB		 1920 skb->dev = vxlan->dev;
1921 skb->pkt_type = PACKET_HOST;
1922 }
f2d1968e 1923
f2d1968e
JB		 1924 oiph = skb_network_header(skb);
1925 skb_reset_network_header(skb);
1926
1927 if (!vxlan_ecn_decapsulate(vs, oiph, skb)) {
1928 ++vxlan->dev->stats.rx_frame_errors;
1929 ++vxlan->dev->stats.rx_errors;
1930 goto drop;
1931 }
1932
59cbf56f
ED		 1933 rcu_read_lock();
1934
1935 if (unlikely(!(vxlan->dev->flags & IFF_UP))) {
1936 rcu_read_unlock();
1937 atomic_long_inc(&vxlan->dev->rx_dropped);
1938 goto drop;
1939 }
1940
f2d1968e
JB		 1941 stats = this_cpu_ptr(vxlan->dev->tstats);
1942 u64_stats_update_begin(&stats->syncp);
1943 stats->rx_packets++;
1944 stats->rx_bytes += skb->len;
1945 u64_stats_update_end(&stats->syncp);
1946
1947 gro_cells_receive(&vxlan->gro_cells, skb);
59cbf56f
ED		 1948
1949 rcu_read_unlock();
1950
5cfccc5a
PS		 1951 return 0;
1952
1953drop:
288b01c8
JB		 1954 /* Consume bad packet */
1955 kfree_skb(skb);
1956 return 0;
5cfccc5a
PS		 1957}
1958
c3a43b9f
SB		 1959/* Callback from net/ipv{4,6}/udp.c to check that we have a VNI for errors */
1960static int vxlan_err_lookup(struct sock *sk, struct sk_buff *skb)
1961{
1962 struct vxlan_dev *vxlan;
1963 struct vxlan_sock *vs;
1964 struct vxlanhdr *hdr;
1965 __be32 vni;
1966
8399a693 1967 if (!pskb_may_pull(skb, skb_transport_offset(skb) + VXLAN_HLEN))
c3a43b9f
SB		 1968 return -EINVAL;
1969
1970 hdr = vxlan_hdr(skb);
1971
1972 if (!(hdr->vx_flags & VXLAN_HF_VNI))
1973 return -EINVAL;
1974
1975 vs = rcu_dereference_sk_user_data(sk);
1976 if (!vs)
1977 return -ENOENT;
1978
1979 vni = vxlan_vni(hdr->vx_vni);
1980 vxlan = vxlan_vs_find_vni(vs, skb->dev->ifindex, vni);
1981 if (!vxlan)
1982 return -ENOENT;
1983
1984 return 0;
1985}
1986
3ad7a4b1 1987static int arp_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni)
e4f67add
DS		 1988{
1989 struct vxlan_dev *vxlan = netdev_priv(dev);
1990 struct arphdr *parp;
1991 u8 *arpptr, *sha;
1992 __be32 sip, tip;
1993 struct neighbour *n;
1994
1995 if (dev->flags & IFF_NOARP)
1996 goto out;
1997
1998 if (!pskb_may_pull(skb, arp_hdr_len(dev))) {
1999 dev->stats.tx_dropped++;
2000 goto out;
2001 }
2002 parp = arp_hdr(skb);
2003
2004 if ((parp->ar_hrd != htons(ARPHRD_ETHER) &&
2005 parp->ar_hrd != htons(ARPHRD_IEEE802)) ||
2006 parp->ar_pro != htons(ETH_P_IP) ||
2007 parp->ar_op != htons(ARPOP_REQUEST) ||
2008 parp->ar_hln != dev->addr_len ||
2009 parp->ar_pln != 4)
2010 goto out;
2011 arpptr = (u8 *)parp + sizeof(struct arphdr);
2012 sha = arpptr;
2013 arpptr += dev->addr_len; /* sha */
2014 memcpy(&sip, arpptr, sizeof(sip));
2015 arpptr += sizeof(sip);
2016 arpptr += dev->addr_len; /* tha */
2017 memcpy(&tip, arpptr, sizeof(tip));
2018
2019 if (ipv4_is_loopback(tip) ||
2020 ipv4_is_multicast(tip))
2021 goto out;
2022
2023 n = neigh_lookup(&arp_tbl, &tip, dev);
2024
2025 if (n) {
e4f67add
DS		 2026 struct vxlan_fdb *f;
2027 struct sk_buff *reply;
2028
2029 if (!(n->nud_state & NUD_CONNECTED)) {
2030 neigh_release(n);
2031 goto out;
2032 }
2033
3ad7a4b1 2034 f = vxlan_find_mac(vxlan, n->ha, vni);
e4c7ed41 2035 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
e4f67add
DS		 2036 /* bridge-local neighbor */
2037 neigh_release(n);
2038 goto out;
2039 }
2040
2041 reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
2042 n->ha, sha);
2043
2044 neigh_release(n);
2045
7346135d
DS		 2046 if (reply == NULL)
2047 goto out;
2048
e4f67add
DS		 2049 skb_reset_mac_header(reply);
2050 __skb_pull(reply, skb_network_offset(reply));
2051 reply->ip_summed = CHECKSUM_UNNECESSARY;
2052 reply->pkt_type = PACKET_HOST;
2053
2054 if (netif_rx_ni(reply) == NET_RX_DROP)
2055 dev->stats.rx_dropped++;
dc5321d7 2056 } else if (vxlan->cfg.flags & VXLAN_F_L3MISS) {
e4c7ed41
CW		 2057 union vxlan_addr ipa = {
2058 .sin.sin_addr.s_addr = tip,
a45e92a5 2059 .sin.sin_family = AF_INET,
e4c7ed41
CW		 2060 };
2061
2062 vxlan_ip_miss(dev, &ipa);
2063 }
e4f67add
DS		 2064out:
2065 consume_skb(skb);
2066 return NETDEV_TX_OK;
2067}
2068
f564f45c 2069#if IS_ENABLED(CONFIG_IPV6)
4b29dba9
DS		 2070static struct sk_buff *vxlan_na_create(struct sk_buff *request,
2071 struct neighbour *n, bool isrouter)
2072{
2073 struct net_device *dev = request->dev;
2074 struct sk_buff *reply;
2075 struct nd_msg *ns, *na;
2076 struct ipv6hdr *pip6;
2077 u8 *daddr;
2078 int na_olen = 8; /* opt hdr + ETH_ALEN for target */
2079 int ns_olen;
2080 int i, len;
2081
f1fb08f6 2082 if (dev == NULL || !pskb_may_pull(request, request->len))
4b29dba9
DS		 2083 return NULL;
2084
2085 len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) +
2086 sizeof(*na) + na_olen + dev->needed_tailroom;
2087 reply = alloc_skb(len, GFP_ATOMIC);
2088 if (reply == NULL)
2089 return NULL;
2090
2091 reply->protocol = htons(ETH_P_IPV6);
2092 reply->dev = dev;
2093 skb_reserve(reply, LL_RESERVED_SPACE(request->dev));
2094 skb_push(reply, sizeof(struct ethhdr));
6297b91c 2095 skb_reset_mac_header(reply);
4b29dba9 2096
f1fb08f6 2097 ns = (struct nd_msg *)(ipv6_hdr(request) + 1);
4b29dba9
DS		 2098
2099 daddr = eth_hdr(request)->h_source;
f1fb08f6
VB		 2100 ns_olen = request->len - skb_network_offset(request) -
2101 sizeof(struct ipv6hdr) - sizeof(*ns);
4b29dba9 2102 for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) {
8066e6b4
IS		 2103 if (!ns->opt[i + 1]) {
2104 kfree_skb(reply);
2105 return NULL;
2106 }
4b29dba9
DS		 2107 if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) {
2108 daddr = ns->opt + i + sizeof(struct nd_opt_hdr);
2109 break;
2110 }
2111 }
2112
2113 /* Ethernet header */
2114 ether_addr_copy(eth_hdr(reply)->h_dest, daddr);
2115 ether_addr_copy(eth_hdr(reply)->h_source, n->ha);
2116 eth_hdr(reply)->h_proto = htons(ETH_P_IPV6);
2117 reply->protocol = htons(ETH_P_IPV6);
2118
2119 skb_pull(reply, sizeof(struct ethhdr));
6297b91c 2120 skb_reset_network_header(reply);
4b29dba9
DS		 2121 skb_put(reply, sizeof(struct ipv6hdr));
2122
2123 /* IPv6 header */
2124
2125 pip6 = ipv6_hdr(reply);
2126 memset(pip6, 0, sizeof(struct ipv6hdr));
2127 pip6->version = 6;
2128 pip6->priority = ipv6_hdr(request)->priority;
2129 pip6->nexthdr = IPPROTO_ICMPV6;
2130 pip6->hop_limit = 255;
2131 pip6->daddr = ipv6_hdr(request)->saddr;
2132 pip6->saddr = *(struct in6_addr *)n->primary_key;
2133
2134 skb_pull(reply, sizeof(struct ipv6hdr));
6297b91c 2135 skb_reset_transport_header(reply);
4b29dba9 2136
4b29dba9 2137 /* Neighbor Advertisement */
b080db58 2138 na = skb_put_zero(reply, sizeof(*na) + na_olen);
4b29dba9
DS		 2139 na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT;
2140 na->icmph.icmp6_router = isrouter;
2141 na->icmph.icmp6_override = 1;
2142 na->icmph.icmp6_solicited = 1;
2143 na->target = ns->target;
2144 ether_addr_copy(&na->opt[2], n->ha);
2145 na->opt[0] = ND_OPT_TARGET_LL_ADDR;
2146 na->opt[1] = na_olen >> 3;
2147
2148 na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr,
2149 &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6,
2150 csum_partial(na, sizeof(*na)+na_olen, 0));
2151
2152 pip6->payload_len = htons(sizeof(*na)+na_olen);
2153
2154 skb_push(reply, sizeof(struct ipv6hdr));
2155
2156 reply->ip_summed = CHECKSUM_UNNECESSARY;
2157
2158 return reply;
2159}
2160
3ad7a4b1 2161static int neigh_reduce(struct net_device *dev, struct sk_buff *skb, __be32 vni)
f564f45c
CW		 2162{
2163 struct vxlan_dev *vxlan = netdev_priv(dev);
8dcd81a9 2164 const struct in6_addr *daddr;
8bff3685 2165 const struct ipv6hdr *iphdr;
4b29dba9 2166 struct inet6_dev *in6_dev;
8bff3685
XL		 2167 struct neighbour *n;
2168 struct nd_msg *msg;
f564f45c
CW		 2169
2170 in6_dev = __in6_dev_get(dev);
2171 if (!in6_dev)
2172 goto out;
2173
f564f45c 2174 iphdr = ipv6_hdr(skb);
f564f45c 2175 daddr = &iphdr->daddr;
f1fb08f6 2176 msg = (struct nd_msg *)(iphdr + 1);
f564f45c 2177
4b29dba9
DS		 2178 if (ipv6_addr_loopback(daddr) ||
2179 ipv6_addr_is_multicast(&msg->target))
2180 goto out;
2181
2182 n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev);
f564f45c
CW		 2183
2184 if (n) {
2185 struct vxlan_fdb *f;
4b29dba9 2186 struct sk_buff *reply;
f564f45c
CW		 2187
2188 if (!(n->nud_state & NUD_CONNECTED)) {
2189 neigh_release(n);
2190 goto out;
2191 }
2192
3ad7a4b1 2193 f = vxlan_find_mac(vxlan, n->ha, vni);
f564f45c
CW		 2194 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
2195 /* bridge-local neighbor */
2196 neigh_release(n);
2197 goto out;
2198 }
2199
4b29dba9
DS		 2200 reply = vxlan_na_create(skb, n,
2201 !!(f ? f->flags & NTF_ROUTER : 0));
2202
f564f45c 2203 neigh_release(n);
4b29dba9
DS		 2204
2205 if (reply == NULL)
2206 goto out;
2207
2208 if (netif_rx_ni(reply) == NET_RX_DROP)
2209 dev->stats.rx_dropped++;
2210
dc5321d7 2211 } else if (vxlan->cfg.flags & VXLAN_F_L3MISS) {
4b29dba9
DS		 2212 union vxlan_addr ipa = {
2213 .sin6.sin6_addr = msg->target,
a45e92a5 2214 .sin6.sin6_family = AF_INET6,
4b29dba9
DS		 2215 };
2216
f564f45c
CW		 2217 vxlan_ip_miss(dev, &ipa);
2218 }
2219
2220out:
2221 consume_skb(skb);
2222 return NETDEV_TX_OK;
2223}
2224#endif
2225
e4f67add
DS		 2226static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb)
2227{
2228 struct vxlan_dev *vxlan = netdev_priv(dev);
2229 struct neighbour *n;
e4f67add
DS		 2230
2231 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
2232 return false;
2233
2234 n = NULL;
2235 switch (ntohs(eth_hdr(skb)->h_proto)) {
2236 case ETH_P_IP:
e15a00aa
CW		 2237 {
2238 struct iphdr *pip;
2239
e4f67add
DS		 2240 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
2241 return false;
2242 pip = ip_hdr(skb);
2243 n = neigh_lookup(&arp_tbl, &pip->daddr, dev);
dc5321d7 2244 if (!n && (vxlan->cfg.flags & VXLAN_F_L3MISS)) {
e4c7ed41
CW		 2245 union vxlan_addr ipa = {
2246 .sin.sin_addr.s_addr = pip->daddr,
a45e92a5 2247 .sin.sin_family = AF_INET,
e4c7ed41
CW		 2248 };
2249
2250 vxlan_ip_miss(dev, &ipa);
2251 return false;
2252 }
2253
e4f67add 2254 break;
e15a00aa
CW		 2255 }
2256#if IS_ENABLED(CONFIG_IPV6)
2257 case ETH_P_IPV6:
2258 {
2259 struct ipv6hdr *pip6;
2260
2261 if (!pskb_may_pull(skb, sizeof(struct ipv6hdr)))
2262 return false;
2263 pip6 = ipv6_hdr(skb);
2264 n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev);
dc5321d7 2265 if (!n && (vxlan->cfg.flags & VXLAN_F_L3MISS)) {
e15a00aa
CW		 2266 union vxlan_addr ipa = {
2267 .sin6.sin6_addr = pip6->daddr,
a45e92a5 2268 .sin6.sin6_family = AF_INET6,
e15a00aa
CW		 2269 };
2270
2271 vxlan_ip_miss(dev, &ipa);
2272 return false;
2273 }
2274
2275 break;
2276 }
2277#endif
e4f67add
DS		 2278 default:
2279 return false;
2280 }
2281
2282 if (n) {
2283 bool diff;
2284
7367d0b5 2285 diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha);
e4f67add
DS		 2286 if (diff) {
2287 memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest,
2288 dev->addr_len);
2289 memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len);
2290 }
2291 neigh_release(n);
2292 return diff;
e4c7ed41
CW		 2293 }
2294
e4f67add
DS		 2295 return false;
2296}
2297
af33c1ad 2298static void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, u32 vxflags,
3511494c
TG		 2299 struct vxlan_metadata *md)
2300{
2301 struct vxlanhdr_gbp *gbp;
2302
db79a621
TG		 2303 if (!md->gbp)
2304 return;
2305
3511494c 2306 gbp = (struct vxlanhdr_gbp *)vxh;
54bfd872 2307 vxh->vx_flags |= VXLAN_HF_GBP;
3511494c
TG		 2308
2309 if (md->gbp & VXLAN_GBP_DONT_LEARN)
2310 gbp->dont_learn = 1;
2311
2312 if (md->gbp & VXLAN_GBP_POLICY_APPLIED)
2313 gbp->policy_applied = 1;
2314
2315 gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK);
2316}
2317
e1e5314d
JB		 2318static int vxlan_build_gpe_hdr(struct vxlanhdr *vxh, u32 vxflags,
2319 __be16 protocol)
2320{
2321 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)vxh;
2322
2323 gpe->np_applied = 1;
fa20e0e3
JB		 2324 gpe->next_protocol = tun_p_from_eth_p(protocol);
2325 if (!gpe->next_protocol)
2326 return -EPFNOSUPPORT;
2327 return 0;
e1e5314d
JB		 2328}
2329
f491e56d
JB		 2330static int vxlan_build_skb(struct sk_buff *skb, struct dst_entry *dst,
2331 int iphdr_len, __be32 vni,
2332 struct vxlan_metadata *md, u32 vxflags,
b4ed5cad 2333 bool udp_sum)
e4c7ed41 2334{
e4c7ed41 2335 struct vxlanhdr *vxh;
e4c7ed41
CW		 2336 int min_headroom;
2337 int err;
dfd8645e 2338 int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL;
e1e5314d 2339 __be16 inner_protocol = htons(ETH_P_TEB);
dfd8645e 2340
af33c1ad 2341 if ((vxflags & VXLAN_F_REMCSUM_TX) &&
dfd8645e
TH		 2342 skb->ip_summed == CHECKSUM_PARTIAL) {
2343 int csum_start = skb_checksum_start_offset(skb);
2344
2345 if (csum_start <= VXLAN_MAX_REMCSUM_START &&
2346 !(csum_start & VXLAN_RCO_SHIFT_MASK) &&
2347 (skb->csum_offset == offsetof(struct udphdr, check) ||
b5708501 2348 skb->csum_offset == offsetof(struct tcphdr, check)))
dfd8645e 2349 type |= SKB_GSO_TUNNEL_REMCSUM;
dfd8645e 2350 }
e4c7ed41 2351
e4c7ed41 2352 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
4a4f86cc 2353 + VXLAN_HLEN + iphdr_len;
649c5b8b
PS		 2354
2355 /* Need space for new headers (invalidates iph ptr) */
2356 err = skb_cow_head(skb, min_headroom);
e1e5314d 2357 if (unlikely(err))
c46b7897 2358 return err;
649c5b8b 2359
aed069df
AD		 2360 err = iptunnel_handle_offloads(skb, type);
2361 if (err)
c46b7897 2362 return err;
b736a623 2363
d58ff351 2364 vxh = __skb_push(skb, sizeof(*vxh));
54bfd872
JB		 2365 vxh->vx_flags = VXLAN_HF_VNI;
2366 vxh->vx_vni = vxlan_vni_field(vni);
49560532 2367
dfd8645e 2368 if (type & SKB_GSO_TUNNEL_REMCSUM) {
54bfd872 2369 unsigned int start;
dfd8645e 2370
54bfd872
JB		 2371 start = skb_checksum_start_offset(skb) - sizeof(struct vxlanhdr);
2372 vxh->vx_vni |= vxlan_compute_rco(start, skb->csum_offset);
2373 vxh->vx_flags |= VXLAN_HF_RCO;
dfd8645e
TH		 2374
2375 if (!skb_is_gso(skb)) {
2376 skb->ip_summed = CHECKSUM_NONE;
2377 skb->encapsulation = 0;
2378 }
2379 }
2380
af33c1ad
TH		 2381 if (vxflags & VXLAN_F_GBP)
2382 vxlan_build_gbp_hdr(vxh, vxflags, md);
e1e5314d
JB		 2383 if (vxflags & VXLAN_F_GPE) {
2384 err = vxlan_build_gpe_hdr(vxh, vxflags, skb->protocol);
2385 if (err < 0)
c46b7897 2386 return err;
e1e5314d
JB		 2387 inner_protocol = skb->protocol;
2388 }
3511494c 2389
e1e5314d 2390 skb_set_inner_protocol(skb, inner_protocol);
039f5062 2391 return 0;
49560532 2392}
49560532 2393
655c3de1 2394static struct rtable *vxlan_get_route(struct vxlan_dev *vxlan, struct net_device *dev,
2395 struct vxlan_sock *sock4,
1a8496ba 2396 struct sk_buff *skb, int oif, u8 tos,
4ecb1d83 2397 __be32 daddr, __be32 *saddr, __be16 dport, __be16 sport,
0c1d70af 2398 struct dst_cache *dst_cache,
db3c6139 2399 const struct ip_tunnel_info *info)
1a8496ba 2400{
db3c6139 2401 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
1a8496ba
JB		 2402 struct rtable *rt = NULL;
2403 struct flowi4 fl4;
2404
655c3de1 2405 if (!sock4)
2406 return ERR_PTR(-EIO);
2407
db3c6139
DB		 2408 if (tos && !info)
2409 use_cache = false;
2410 if (use_cache) {
0c1d70af
PA		 2411 rt = dst_cache_get_ip4(dst_cache, saddr);
2412 if (rt)
2413 return rt;
2414 }
2415
1a8496ba
JB		 2416 memset(&fl4, 0, sizeof(fl4));
2417 fl4.flowi4_oif = oif;
2418 fl4.flowi4_tos = RT_TOS(tos);
2419 fl4.flowi4_mark = skb->mark;
2420 fl4.flowi4_proto = IPPROTO_UDP;
2421 fl4.daddr = daddr;
272d96a5 2422 fl4.saddr = *saddr;
4ecb1d83
MP		 2423 fl4.fl4_dport = dport;
2424 fl4.fl4_sport = sport;
1a8496ba
JB		 2425
2426 rt = ip_route_output_key(vxlan->net, &fl4);
478db1f1 2427 if (!IS_ERR(rt)) {
655c3de1 2428 if (rt->dst.dev == dev) {
2429 netdev_dbg(dev, "circular route to %pI4\n", &daddr);
2430 ip_rt_put(rt);
2431 return ERR_PTR(-ELOOP);
2432 }
2433
1a8496ba 2434 *saddr = fl4.saddr;
0c1d70af
PA		 2435 if (use_cache)
2436 dst_cache_set_ip4(dst_cache, &rt->dst, fl4.saddr);
655c3de1 2437 } else {
2438 netdev_dbg(dev, "no route to %pI4\n", &daddr);
2439 return ERR_PTR(-ENETUNREACH);
0c1d70af 2440 }
1a8496ba
JB		 2441 return rt;
2442}
2443
e5d4b29f
JB		 2444#if IS_ENABLED(CONFIG_IPV6)
2445static struct dst_entry *vxlan6_get_route(struct vxlan_dev *vxlan,
655c3de1 2446 struct net_device *dev,
03dc52a8 2447 struct vxlan_sock *sock6,
1400615d 2448 struct sk_buff *skb, int oif, u8 tos,
e7f70af1 2449 __be32 label,
e5d4b29f 2450 const struct in6_addr *daddr,
0c1d70af 2451 struct in6_addr *saddr,
4ecb1d83 2452 __be16 dport, __be16 sport,
db3c6139
DB		 2453 struct dst_cache *dst_cache,
2454 const struct ip_tunnel_info *info)
e5d4b29f 2455{
db3c6139 2456 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
e5d4b29f
JB		 2457 struct dst_entry *ndst;
2458 struct flowi6 fl6;
e5d4b29f 2459
c6fcc4fc 2460 if (!sock6)
2461 return ERR_PTR(-EIO);
2462
1400615d
DB		 2463 if (tos && !info)
2464 use_cache = false;
db3c6139 2465 if (use_cache) {
0c1d70af
PA		 2466 ndst = dst_cache_get_ip6(dst_cache, saddr);
2467 if (ndst)
2468 return ndst;
2469 }
2470
e5d4b29f
JB		 2471 memset(&fl6, 0, sizeof(fl6));
2472 fl6.flowi6_oif = oif;
2473 fl6.daddr = *daddr;
272d96a5 2474 fl6.saddr = *saddr;
eaa93bf4 2475 fl6.flowlabel = ip6_make_flowinfo(RT_TOS(tos), label);
e5d4b29f
JB		 2476 fl6.flowi6_mark = skb->mark;
2477 fl6.flowi6_proto = IPPROTO_UDP;
4ecb1d83
MP		 2478 fl6.fl6_dport = dport;
2479 fl6.fl6_sport = sport;
e5d4b29f 2480
6c8991f4
SD		 2481 ndst = ipv6_stub->ipv6_dst_lookup_flow(vxlan->net, sock6->sock->sk,
2482 &fl6, NULL);
2483 if (unlikely(IS_ERR(ndst))) {
655c3de1 2484 netdev_dbg(dev, "no route to %pI6\n", daddr);
2485 return ERR_PTR(-ENETUNREACH);
2486 }
2487
2488 if (unlikely(ndst->dev == dev)) {
2489 netdev_dbg(dev, "circular route to %pI6\n", daddr);
2490 dst_release(ndst);
2491 return ERR_PTR(-ELOOP);
2492 }
e5d4b29f
JB		 2493
2494 *saddr = fl6.saddr;
db3c6139 2495 if (use_cache)
0c1d70af 2496 dst_cache_set_ip6(dst_cache, ndst, saddr);
e5d4b29f
JB		 2497 return ndst;
2498}
2499#endif
2500
9dcc71e1
SS		 2501/* Bypass encapsulation if the destination is local */
2502static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan,
fc68c995
SB		 2503 struct vxlan_dev *dst_vxlan, __be32 vni,
2504 bool snoop)
9dcc71e1 2505{
8f84985f 2506 struct pcpu_sw_netstats *tx_stats, *rx_stats;
e4c7ed41
CW		 2507 union vxlan_addr loopback;
2508 union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip;
4179cb5a 2509 struct net_device *dev;
ce6502a8 2510 int len = skb->len;
9dcc71e1 2511
8f84985f
LR		 2512 tx_stats = this_cpu_ptr(src_vxlan->dev->tstats);
2513 rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats);
9dcc71e1
SS		 2514 skb->pkt_type = PACKET_HOST;
2515 skb->encapsulation = 0;
2516 skb->dev = dst_vxlan->dev;
2517 __skb_pull(skb, skb_network_offset(skb));
2518
e4c7ed41
CW		 2519 if (remote_ip->sa.sa_family == AF_INET) {
2520 loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2521 loopback.sa.sa_family = AF_INET;
2522#if IS_ENABLED(CONFIG_IPV6)
2523 } else {
2524 loopback.sin6.sin6_addr = in6addr_loopback;
2525 loopback.sa.sa_family = AF_INET6;
2526#endif
2527 }
2528
4179cb5a
ED		 2529 rcu_read_lock();
2530 dev = skb->dev;
2531 if (unlikely(!(dev->flags & IFF_UP))) {
2532 kfree_skb(skb);
2533 goto drop;
2534 }
2535
fc68c995 2536 if ((dst_vxlan->cfg.flags & VXLAN_F_LEARN) && snoop)
4179cb5a 2537 vxlan_snoop(dev, &loopback, eth_hdr(skb)->h_source, 0, vni);
9dcc71e1
SS		 2538
2539 u64_stats_update_begin(&tx_stats->syncp);
2540 tx_stats->tx_packets++;
ce6502a8 2541 tx_stats->tx_bytes += len;
9dcc71e1
SS		 2542 u64_stats_update_end(&tx_stats->syncp);
2543
2544 if (netif_rx(skb) == NET_RX_SUCCESS) {
2545 u64_stats_update_begin(&rx_stats->syncp);
2546 rx_stats->rx_packets++;
ce6502a8 2547 rx_stats->rx_bytes += len;
9dcc71e1
SS		 2548 u64_stats_update_end(&rx_stats->syncp);
2549 } else {
4179cb5a 2550drop:
ce6502a8 2551 dev->stats.rx_dropped++;
9dcc71e1 2552 }
4179cb5a 2553 rcu_read_unlock();
9dcc71e1
SS		 2554}
2555
fee1fad7 2556static int encap_bypass_if_local(struct sk_buff *skb, struct net_device *dev,
49f810f0
MS		 2557 struct vxlan_dev *vxlan,
2558 union vxlan_addr *daddr,
2559 __be16 dst_port, int dst_ifindex, __be32 vni,
2560 struct dst_entry *dst,
fee1fad7 2561 u32 rt_flags)
2562{
2563#if IS_ENABLED(CONFIG_IPV6)
2564 /* IPv6 rt-flags are checked against RTF_LOCAL, but the value of
2565 * RTF_LOCAL is equal to RTCF_LOCAL. So to keep code simple
2566 * we can use RTCF_LOCAL which works for ipv4 and ipv6 route entry.
2567 */
2568 BUILD_BUG_ON(RTCF_LOCAL != RTF_LOCAL);
2569#endif
2570 /* Bypass encapsulation if the destination is local */
2571 if (rt_flags & RTCF_LOCAL &&
2572 !(rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
2573 struct vxlan_dev *dst_vxlan;
2574
2575 dst_release(dst);
49f810f0 2576 dst_vxlan = vxlan_find_vni(vxlan->net, dst_ifindex, vni,
fee1fad7 2577 daddr->sa.sa_family, dst_port,
dc5321d7 2578 vxlan->cfg.flags);
fee1fad7 2579 if (!dst_vxlan) {
2580 dev->stats.tx_errors++;
2581 kfree_skb(skb);
2582
2583 return -ENOENT;
2584 }
fc68c995 2585 vxlan_encap_bypass(skb, vxlan, dst_vxlan, vni, true);
fee1fad7 2586 return 1;
2587 }
2588
2589 return 0;
2590}
2591
4ad16930 2592static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev,
3ad7a4b1
RP		 2593 __be32 default_vni, struct vxlan_rdst *rdst,
2594 bool did_rsc)
d342894c 2595{
d71785ff 2596 struct dst_cache *dst_cache;
3093fbe7 2597 struct ip_tunnel_info *info;
d342894c 2598 struct vxlan_dev *vxlan = netdev_priv(dev);
0770b53b 2599 const struct iphdr *old_iph = ip_hdr(skb);
e4c7ed41 2600 union vxlan_addr *dst;
272d96a5 2601 union vxlan_addr remote_ip, local_ip;
ee122c79
TG		 2602 struct vxlan_metadata _md;
2603 struct vxlan_metadata *md = &_md;
e4c7ed41 2604 __be16 src_port = 0, dst_port;
655c3de1 2605 struct dst_entry *ndst = NULL;
e7f70af1 2606 __be32 vni, label;
d342894c 2607 __u8 tos, ttl;
49f810f0 2608 int ifindex;
0e6fbc5b 2609 int err;
dc5321d7 2610 u32 flags = vxlan->cfg.flags;
b4ed5cad 2611 bool udp_sum = false;
f491e56d 2612 bool xnet = !net_eq(vxlan->net, dev_net(vxlan->dev));
d342894c 2613
61adedf3 2614 info = skb_tunnel_info(skb);
3093fbe7 2615
ee122c79 2616 if (rdst) {
0770b53b 2617 dst = &rdst->remote_ip;
2618 if (vxlan_addr_any(dst)) {
2619 if (did_rsc) {
2620 /* short-circuited back to local bridge */
fc68c995
SB		 2621 vxlan_encap_bypass(skb, vxlan, vxlan,
2622 default_vni, true);
0770b53b 2623 return;
2624 }
2625 goto drop;
2626 }
2627
0dfbdf41 2628 dst_port = rdst->remote_port ? rdst->remote_port : vxlan->cfg.dst_port;
3ad7a4b1 2629 vni = (rdst->remote_vni) ? : default_vni;
49f810f0 2630 ifindex = rdst->remote_ifindex;
1158632b 2631 local_ip = vxlan->cfg.saddr;
d71785ff 2632 dst_cache = &rdst->dst_cache;
0770b53b 2633 md->gbp = skb->mark;
72f6d71e
HL		 2634 if (flags & VXLAN_F_TTL_INHERIT) {
2635 ttl = ip_tunnel_get_ttl(old_iph, skb);
2636 } else {
2637 ttl = vxlan->cfg.ttl;
2638 if (!ttl && vxlan_addr_multicast(dst))
2639 ttl = 1;
2640 }
0770b53b 2641
2642 tos = vxlan->cfg.tos;
2643 if (tos == 1)
2644 tos = ip_tunnel_get_dsfield(old_iph, skb);
2645
2646 if (dst->sa.sa_family == AF_INET)
2647 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM_TX);
2648 else
2649 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM6_TX);
2650 label = vxlan->cfg.label;
ee122c79
TG		 2651 } else {
2652 if (!info) {
2653 WARN_ONCE(1, "%s: Missing encapsulation instructions\n",
2654 dev->name);
2655 goto drop;
2656 }
b1be00a6 2657 remote_ip.sa.sa_family = ip_tunnel_info_af(info);
272d96a5 2658 if (remote_ip.sa.sa_family == AF_INET) {
a725e514 2659 remote_ip.sin.sin_addr.s_addr = info->key.u.ipv4.dst;
272d96a5 2660 local_ip.sin.sin_addr.s_addr = info->key.u.ipv4.src;
2661 } else {
a725e514 2662 remote_ip.sin6.sin6_addr = info->key.u.ipv6.dst;
272d96a5 2663 local_ip.sin6.sin6_addr = info->key.u.ipv6.src;
2664 }
ee122c79 2665 dst = &remote_ip;
0770b53b 2666 dst_port = info->key.tp_dst ? : vxlan->cfg.dst_port;
2667 vni = tunnel_id_to_key32(info->key.tun_id);
49f810f0 2668 ifindex = 0;
d71785ff 2669 dst_cache = &info->dst_cache;
eadf52cf
XL		 2670 if (info->key.tun_flags & TUNNEL_VXLAN_OPT) {
2671 if (info->options_len < sizeof(*md))
2672 goto drop;
0770b53b 2673 md = ip_tunnel_info_opts(info);
eadf52cf 2674 }
a725e514
JB		 2675 ttl = info->key.ttl;
2676 tos = info->key.tos;
e7f70af1 2677 label = info->key.label;
b4ed5cad 2678 udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
a725e514 2679 }
0770b53b 2680 src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
2681 vxlan->cfg.port_max, true);
a725e514 2682
56de859e 2683 rcu_read_lock();
e4c7ed41 2684 if (dst->sa.sa_family == AF_INET) {
c6fcc4fc 2685 struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock);
c46b7897 2686 struct rtable *rt;
0770b53b 2687 __be16 df = 0;
c6fcc4fc 2688
aab8cc36
AB		 2689 if (!ifindex)
2690 ifindex = sock4->sock->sk->sk_bound_dev_if;
2691
49f810f0 2692 rt = vxlan_get_route(vxlan, dev, sock4, skb, ifindex, tos,
272d96a5 2693 dst->sin.sin_addr.s_addr,
1158632b 2694 &local_ip.sin.sin_addr.s_addr,
4ecb1d83 2695 dst_port, src_port,
d71785ff 2696 dst_cache, info);
8ebd115b
DM		 2697 if (IS_ERR(rt)) {
2698 err = PTR_ERR(rt);
c46b7897 2699 goto tx_error;
8ebd115b 2700 }
e4c7ed41 2701
fee1fad7 2702 if (!info) {
b4d30697 2703 /* Bypass encapsulation if the destination is local */
fee1fad7 2704 err = encap_bypass_if_local(skb, dev, vxlan, dst,
49f810f0
MS		 2705 dst_port, ifindex, vni,
2706 &rt->dst, rt->rt_flags);
fee1fad7 2707 if (err)
56de859e 2708 goto out_unlock;
b4d30697
SB		 2709
2710 if (vxlan->cfg.df == VXLAN_DF_SET) {
2711 df = htons(IP_DF);
2712 } else if (vxlan->cfg.df == VXLAN_DF_INHERIT) {
2713 struct ethhdr *eth = eth_hdr(skb);
2714
2715 if (ntohs(eth->h_proto) == ETH_P_IPV6 ||
2716 (ntohs(eth->h_proto) == ETH_P_IP &&
2717 old_iph->frag_off & htons(IP_DF)))
2718 df = htons(IP_DF);
2719 }
fee1fad7 2720 } else if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT) {
6ceb31ca 2721 df = htons(IP_DF);
fee1fad7 2722 }
6ceb31ca 2723
c46b7897 2724 ndst = &rt->dst;
fc68c995
SB		 2725 err = skb_tunnel_check_pmtu(skb, ndst, VXLAN_HEADROOM,
2726 netif_is_any_bridge_port(dev));
2727 if (err < 0) {
2728 goto tx_error;
2729 } else if (err) {
2730 if (info) {
2731 struct in_addr src, dst;
2732
2733 src = remote_ip.sin.sin_addr;
2734 dst = local_ip.sin.sin_addr;
2735 info->key.u.ipv4.src = src.s_addr;
2736 info->key.u.ipv4.dst = dst.s_addr;
2737 }
2738 vxlan_encap_bypass(skb, vxlan, vxlan, vni, false);
2739 dst_release(ndst);
2740 goto out_unlock;
2741 }
a93bf0ff 2742
a0dced17 2743 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
e4c7ed41 2744 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
c46b7897 2745 err = vxlan_build_skb(skb, ndst, sizeof(struct iphdr),
54bfd872 2746 vni, md, flags, udp_sum);
f491e56d 2747 if (err < 0)
c46b7897 2748 goto tx_error;
f491e56d 2749
1158632b 2750 udp_tunnel_xmit_skb(rt, sock4->sock->sk, skb, local_ip.sin.sin_addr.s_addr,
f491e56d
JB		 2751 dst->sin.sin_addr.s_addr, tos, ttl, df,
2752 src_port, dst_port, xnet, !udp_sum);
e4c7ed41
CW		 2753#if IS_ENABLED(CONFIG_IPV6)
2754 } else {
c6fcc4fc 2755 struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock);
e4c7ed41 2756
aab8cc36
AB		 2757 if (!ifindex)
2758 ifindex = sock6->sock->sk->sk_bound_dev_if;
2759
49f810f0 2760 ndst = vxlan6_get_route(vxlan, dev, sock6, skb, ifindex, tos,
272d96a5 2761 label, &dst->sin6.sin6_addr,
1158632b 2762 &local_ip.sin6.sin6_addr,
4ecb1d83 2763 dst_port, src_port,
db3c6139 2764 dst_cache, info);
e5d4b29f 2765 if (IS_ERR(ndst)) {
8ebd115b 2766 err = PTR_ERR(ndst);
c46b7897 2767 ndst = NULL;
9dcc71e1 2768 goto tx_error;
e4c7ed41 2769 }
655c3de1 2770
fee1fad7 2771 if (!info) {
2772 u32 rt6i_flags = ((struct rt6_info *)ndst)->rt6i_flags;
49560532 2773
fee1fad7 2774 err = encap_bypass_if_local(skb, dev, vxlan, dst,
49f810f0
MS		 2775 dst_port, ifindex, vni,
2776 ndst, rt6i_flags);
fee1fad7 2777 if (err)
56de859e 2778 goto out_unlock;
fee1fad7 2779 }
35e2d115 2780
fc68c995
SB		 2781 err = skb_tunnel_check_pmtu(skb, ndst, VXLAN6_HEADROOM,
2782 netif_is_any_bridge_port(dev));
2783 if (err < 0) {
2784 goto tx_error;
2785 } else if (err) {
2786 if (info) {
2787 struct in6_addr src, dst;
2788
2789 src = remote_ip.sin6.sin6_addr;
2790 dst = local_ip.sin6.sin6_addr;
2791 info->key.u.ipv6.src = src;
2792 info->key.u.ipv6.dst = dst;
2793 }
2794
2795 vxlan_encap_bypass(skb, vxlan, vxlan, vni, false);
2796 dst_release(ndst);
2797 goto out_unlock;
2798 }
a93bf0ff 2799
a0dced17 2800 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
e4c7ed41 2801 ttl = ttl ? : ip6_dst_hoplimit(ndst);
f491e56d
JB		 2802 skb_scrub_packet(skb, xnet);
2803 err = vxlan_build_skb(skb, ndst, sizeof(struct ipv6hdr),
54bfd872 2804 vni, md, flags, udp_sum);
c46b7897 2805 if (err < 0)
2806 goto tx_error;
2807
0770b53b 2808 udp_tunnel6_xmit_skb(ndst, sock6->sock->sk, skb, dev,
1158632b 2809 &local_ip.sin6.sin6_addr,
272d96a5 2810 &dst->sin6.sin6_addr, tos, ttl,
e7f70af1 2811 label, src_port, dst_port, !udp_sum);
e4c7ed41
CW		 2812#endif
2813 }
56de859e
JK		 2814out_unlock:
2815 rcu_read_unlock();
4ad16930 2816 return;
d342894c 2817
2818drop:
2819 dev->stats.tx_dropped++;
c46b7897 2820 dev_kfree_skb(skb);
2821 return;
d342894c 2822
2823tx_error:
56de859e 2824 rcu_read_unlock();
655c3de1 2825 if (err == -ELOOP)
2826 dev->stats.collisions++;
2827 else if (err == -ENETUNREACH)
2828 dev->stats.tx_carrier_errors++;
c46b7897 2829 dst_release(ndst);
d342894c 2830 dev->stats.tx_errors++;
c46b7897 2831 kfree_skb(skb);
d342894c 2832}
2833
1274e1cc
RP		 2834static void vxlan_xmit_nh(struct sk_buff *skb, struct net_device *dev,
2835 struct vxlan_fdb *f, __be32 vni, bool did_rsc)
2836{
2837 struct vxlan_rdst nh_rdst;
2838 struct nexthop *nh;
2839 bool do_xmit;
2840 u32 hash;
2841
2842 memset(&nh_rdst, 0, sizeof(struct vxlan_rdst));
2843 hash = skb_get_hash(skb);
2844
2845 rcu_read_lock();
2846 nh = rcu_dereference(f->nh);
2847 if (!nh) {
2848 rcu_read_unlock();
2849 goto drop;
2850 }
2851 do_xmit = vxlan_fdb_nh_path_select(nh, hash, &nh_rdst);
2852 rcu_read_unlock();
2853
2854 if (likely(do_xmit))
2855 vxlan_xmit_one(skb, dev, vni, &nh_rdst, did_rsc);
2856 else
2857 goto drop;
2858
2859 return;
2860
2861drop:
2862 dev->stats.tx_dropped++;
2863 dev_kfree_skb(skb);
2864}
2865
6681712d
DS		 2866/* Transmit local packets over Vxlan
2867 *
2868 * Outer IP header inherits ECN and DF from inner header.
2869 * Outer UDP destination is the VXLAN assigned port.
2870 * source port is based on hash of flow
2871 */
2872static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev)
2873{
2874 struct vxlan_dev *vxlan = netdev_priv(dev);
8bff3685 2875 struct vxlan_rdst *rdst, *fdst = NULL;
3093fbe7 2876 const struct ip_tunnel_info *info;
6681712d 2877 bool did_rsc = false;
6681712d 2878 struct vxlan_fdb *f;
8bff3685 2879 struct ethhdr *eth;
3ad7a4b1 2880 __be32 vni = 0;
6681712d 2881
61adedf3 2882 info = skb_tunnel_info(skb);
3093fbe7 2883
6681712d 2884 skb_reset_mac_header(skb);
6681712d 2885
dc5321d7 2886 if (vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA) {
3ad7a4b1
RP		 2887 if (info && info->mode & IP_TUNNEL_INFO_BRIDGE &&
2888 info->mode & IP_TUNNEL_INFO_TX) {
2889 vni = tunnel_id_to_key32(info->key.tun_id);
2890 } else {
2891 if (info && info->mode & IP_TUNNEL_INFO_TX)
2892 vxlan_xmit_one(skb, dev, vni, NULL, false);
2893 else
2894 kfree_skb(skb);
2895 return NETDEV_TX_OK;
2896 }
47e5d1b0
JB		 2897 }
2898
dc5321d7 2899 if (vxlan->cfg.flags & VXLAN_F_PROXY) {
47e5d1b0 2900 eth = eth_hdr(skb);
f564f45c 2901 if (ntohs(eth->h_proto) == ETH_P_ARP)
3ad7a4b1 2902 return arp_reduce(dev, skb, vni);
f564f45c 2903#if IS_ENABLED(CONFIG_IPV6)
8bff3685
XL		 2904 else if (ntohs(eth->h_proto) == ETH_P_IPV6 &&
2905 pskb_may_pull(skb, sizeof(struct ipv6hdr) +
2906 sizeof(struct nd_msg)) &&
2907 ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) {
2908 struct nd_msg *m = (struct nd_msg *)(ipv6_hdr(skb) + 1);
2909
2910 if (m->icmph.icmp6_code == 0 &&
2911 m->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION)
f1fb08f6 2912 return neigh_reduce(dev, skb, vni);
f564f45c
CW		 2913 }
2914#endif
2915 }
6681712d 2916
47e5d1b0 2917 eth = eth_hdr(skb);
3ad7a4b1 2918 f = vxlan_find_mac(vxlan, eth->h_dest, vni);
ae884082
DS		 2919 did_rsc = false;
2920
dc5321d7 2921 if (f && (f->flags & NTF_ROUTER) && (vxlan->cfg.flags & VXLAN_F_RSC) &&
e15a00aa
CW		 2922 (ntohs(eth->h_proto) == ETH_P_IP ||
2923 ntohs(eth->h_proto) == ETH_P_IPV6)) {
ae884082
DS		 2924 did_rsc = route_shortcircuit(dev, skb);
2925 if (did_rsc)
3ad7a4b1 2926 f = vxlan_find_mac(vxlan, eth->h_dest, vni);
ae884082
DS		 2927 }
2928
6681712d 2929 if (f == NULL) {
3ad7a4b1 2930 f = vxlan_find_mac(vxlan, all_zeros_mac, vni);
afbd8bae 2931 if (f == NULL) {
dc5321d7 2932 if ((vxlan->cfg.flags & VXLAN_F_L2MISS) &&
afbd8bae
MR		 2933 !is_multicast_ether_addr(eth->h_dest))
2934 vxlan_fdb_miss(vxlan, eth->h_dest);
2935
2936 dev->stats.tx_dropped++;
8f646c92 2937 kfree_skb(skb);
afbd8bae
MR		 2938 return NETDEV_TX_OK;
2939 }
2940 }
6681712d 2941
1274e1cc
RP		 2942 if (rcu_access_pointer(f->nh)) {
2943 vxlan_xmit_nh(skb, dev, f,
2944 (vni ? : vxlan->default_dst.remote_vni), did_rsc);
2945 } else {
2946 list_for_each_entry_rcu(rdst, &f->remotes, list) {
2947 struct sk_buff *skb1;
6681712d 2948
1274e1cc
RP		 2949 if (!fdst) {
2950 fdst = rdst;
2951 continue;
2952 }
2953 skb1 = skb_clone(skb, GFP_ATOMIC);
2954 if (skb1)
2955 vxlan_xmit_one(skb1, dev, vni, rdst, did_rsc);
8f646c92 2956 }
1274e1cc
RP		 2957 if (fdst)
2958 vxlan_xmit_one(skb, dev, vni, fdst, did_rsc);
2959 else
2960 kfree_skb(skb);
6681712d
DS		 2961 }
2962
4ad16930 2963 return NETDEV_TX_OK;
6681712d
DS		 2964}
2965
d342894c 2966/* Walk the forwarding table and purge stale entries */
df7e828c 2967static void vxlan_cleanup(struct timer_list *t)
d342894c 2968{
df7e828c 2969 struct vxlan_dev *vxlan = from_timer(vxlan, t, age_timer);
d342894c 2970 unsigned long next_timer = jiffies + FDB_AGE_INTERVAL;
2971 unsigned int h;
2972
2973 if (!netif_running(vxlan->dev))
2974 return;
2975
d342894c 2976 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2977 struct hlist_node *p, *n;
14e1d0fa 2978
fe1e0713 2979 spin_lock(&vxlan->hash_lock[h]);
d342894c 2980 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2981 struct vxlan_fdb *f
2982 = container_of(p, struct vxlan_fdb, hlist);
2983 unsigned long timeout;
2984
efb5f68f 2985 if (f->state & (NUD_PERMANENT | NUD_NOARP))
d342894c 2986 continue;
2987
def499c9
RP		 2988 if (f->flags & NTF_EXT_LEARNED)
2989 continue;
2990
0dfbdf41 2991 timeout = f->used + vxlan->cfg.age_interval * HZ;
d342894c 2992 if (time_before_eq(timeout, jiffies)) {
2993 netdev_dbg(vxlan->dev,
2994 "garbage collect %pM\n",
2995 f->eth_addr);
2996 f->state = NUD_STALE;
0e6160f3 2997 vxlan_fdb_destroy(vxlan, f, true, true);
d342894c 2998 } else if (time_before(timeout, next_timer))
2999 next_timer = timeout;
3000 }
fe1e0713 3001 spin_unlock(&vxlan->hash_lock[h]);
d342894c 3002 }
d342894c 3003
3004 mod_timer(&vxlan->age_timer, next_timer);
3005}
3006
a53cb29b
MB		 3007static void vxlan_vs_del_dev(struct vxlan_dev *vxlan)
3008{
3009 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
3010
3011 spin_lock(&vn->sock_lock);
69e76661
JB		 3012 hlist_del_init_rcu(&vxlan->hlist4.hlist);
3013#if IS_ENABLED(CONFIG_IPV6)
3014 hlist_del_init_rcu(&vxlan->hlist6.hlist);
3015#endif
a53cb29b
MB		 3016 spin_unlock(&vn->sock_lock);
3017}
3018
69e76661
JB		 3019static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan,
3020 struct vxlan_dev_node *node)
9c2e24e1 3021{
56ef9c90 3022 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
54bfd872 3023 __be32 vni = vxlan->default_dst.remote_vni;
9c2e24e1 3024
69e76661 3025 node->vxlan = vxlan;
56ef9c90 3026 spin_lock(&vn->sock_lock);
69e76661 3027 hlist_add_head_rcu(&node->hlist, vni_head(vs, vni));
56ef9c90 3028 spin_unlock(&vn->sock_lock);
9c2e24e1
PS		 3029}
3030
d342894c 3031/* Setup stats when device is created */
3032static int vxlan_init(struct net_device *dev)
3033{
384d91c2
TY		 3034 struct vxlan_dev *vxlan = netdev_priv(dev);
3035 int err;
3036
1c213bd2 3037 dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
e8171045 3038 if (!dev->tstats)
d342894c 3039 return -ENOMEM;
3040
384d91c2
TY		 3041 err = gro_cells_init(&vxlan->gro_cells, dev);
3042 if (err) {
3043 free_percpu(dev->tstats);
3044 return err;
3045 }
3046
d342894c 3047 return 0;
3048}
3049
3ad7a4b1 3050static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan, __be32 vni)
afbd8bae
MR		 3051{
3052 struct vxlan_fdb *f;
fe1e0713 3053 u32 hash_index = fdb_head_index(vxlan, all_zeros_mac, vni);
afbd8bae 3054
fe1e0713 3055 spin_lock_bh(&vxlan->hash_lock[hash_index]);
3ad7a4b1 3056 f = __vxlan_find_mac(vxlan, all_zeros_mac, vni);
afbd8bae 3057 if (f)
0e6160f3 3058 vxlan_fdb_destroy(vxlan, f, true, true);
fe1e0713 3059 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
afbd8bae
MR		 3060}
3061
ebf4063e
SH		 3062static void vxlan_uninit(struct net_device *dev)
3063{
3064 struct vxlan_dev *vxlan = netdev_priv(dev);
ebf4063e 3065
ad6c9986
SB		 3066 gro_cells_destroy(&vxlan->gro_cells);
3067
3ad7a4b1 3068 vxlan_fdb_delete_default(vxlan, vxlan->cfg.vni);
afbd8bae 3069
ebf4063e
SH		 3070 free_percpu(dev->tstats);
3071}
3072
d342894c 3073/* Start ageing timer and join group when device is brought up */
3074static int vxlan_open(struct net_device *dev)
3075{
3076 struct vxlan_dev *vxlan = netdev_priv(dev);
205f356d 3077 int ret;
56ef9c90 3078
205f356d
JB		 3079 ret = vxlan_sock_add(vxlan);
3080 if (ret < 0)
3081 return ret;
d342894c 3082
79d4a94f 3083 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) {
56ef9c90 3084 ret = vxlan_igmp_join(vxlan);
bef0057b
MRL		 3085 if (ret == -EADDRINUSE)
3086 ret = 0;
56ef9c90 3087 if (ret) {
205f356d 3088 vxlan_sock_release(vxlan);
56ef9c90
MRL		 3089 return ret;
3090 }
d342894c 3091 }
3092
0dfbdf41 3093 if (vxlan->cfg.age_interval)
d342894c 3094 mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL);
3095
56ef9c90 3096 return ret;
d342894c 3097}
3098
3099/* Purge the forwarding table */
8b3f9337 3100static void vxlan_flush(struct vxlan_dev *vxlan, bool do_all)
d342894c 3101{
31fec5aa 3102 unsigned int h;
d342894c 3103
d342894c 3104 for (h = 0; h < FDB_HASH_SIZE; ++h) {
3105 struct hlist_node *p, *n;
fe1e0713
L		 3106
3107 spin_lock_bh(&vxlan->hash_lock[h]);
d342894c 3108 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
3109 struct vxlan_fdb *f
3110 = container_of(p, struct vxlan_fdb, hlist);
8b3f9337
RP		 3111 if (!do_all && (f->state & (NUD_PERMANENT | NUD_NOARP)))
3112 continue;
afbd8bae 3113 /* the all_zeros_mac entry is deleted at vxlan_uninit */
fda2ec62
TY		 3114 if (is_zero_ether_addr(f->eth_addr) &&
3115 f->vni == vxlan->cfg.vni)
3116 continue;
3117 vxlan_fdb_destroy(vxlan, f, true, true);
d342894c 3118 }
fe1e0713 3119 spin_unlock_bh(&vxlan->hash_lock[h]);
d342894c 3120 }
d342894c 3121}
3122
3123/* Cleanup timer and forwarding table on shutdown */
3124static int vxlan_stop(struct net_device *dev)
3125{
3126 struct vxlan_dev *vxlan = netdev_priv(dev);
f01ec1c0 3127 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
56ef9c90 3128 int ret = 0;
d342894c 3129
24c0e683 3130 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip) &&
f13b1689 3131 !vxlan_group_used(vn, vxlan))
56ef9c90 3132 ret = vxlan_igmp_leave(vxlan);
d342894c 3133
3134 del_timer_sync(&vxlan->age_timer);
3135
8b3f9337 3136 vxlan_flush(vxlan, false);
205f356d 3137 vxlan_sock_release(vxlan);
d342894c 3138
56ef9c90 3139 return ret;
d342894c 3140}
3141
d342894c 3142/* Stub, nothing needs to be done. */
3143static void vxlan_set_multicast_list(struct net_device *dev)
3144{
3145}
3146
91572088 3147static int vxlan_change_mtu(struct net_device *dev, int new_mtu)
345010b5 3148{
91572088
JW		 3149 struct vxlan_dev *vxlan = netdev_priv(dev);
3150 struct vxlan_rdst *dst = &vxlan->default_dst;
3151 struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
3152 dst->remote_ifindex);
ce44a4ae 3153 bool use_ipv6 = !!(vxlan->cfg.flags & VXLAN_F_IPV6);
345010b5 3154
91572088
JW		 3155 /* This check is different than dev->max_mtu, because it looks at
3156 * the lowerdev->mtu, rather than the static dev->max_mtu
3157 */
3158 if (lowerdev) {
3159 int max_mtu = lowerdev->mtu -
3160 (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
3161 if (new_mtu > max_mtu)
72564b59 3162 return -EINVAL;
72564b59
DW		 3163 }
3164
345010b5
DB		 3165 dev->mtu = new_mtu;
3166 return 0;
3167}
3168
fc4099f1
PS		 3169static int vxlan_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
3170{
3171 struct vxlan_dev *vxlan = netdev_priv(dev);
3172 struct ip_tunnel_info *info = skb_tunnel_info(skb);
3173 __be16 sport, dport;
3174
3175 sport = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
3176 vxlan->cfg.port_max, true);
3177 dport = info->key.tp_dst ? : vxlan->cfg.dst_port;
3178
239e944f 3179 if (ip_tunnel_info_af(info) == AF_INET) {
c6fcc4fc 3180 struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock);
1a8496ba
JB		 3181 struct rtable *rt;
3182
655c3de1 3183 rt = vxlan_get_route(vxlan, dev, sock4, skb, 0, info->key.tos,
1a8496ba 3184 info->key.u.ipv4.dst,
22f0708a
PA		 3185 &info->key.u.ipv4.src, dport, sport,
3186 &info->dst_cache, info);
1a8496ba
JB		 3187 if (IS_ERR(rt))
3188 return PTR_ERR(rt);
3189 ip_rt_put(rt);
239e944f
JB		 3190 } else {
3191#if IS_ENABLED(CONFIG_IPV6)
03dc52a8 3192 struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock);
239e944f
JB		 3193 struct dst_entry *ndst;
3194
655c3de1 3195 ndst = vxlan6_get_route(vxlan, dev, sock6, skb, 0, info->key.tos,
e7f70af1 3196 info->key.label, &info->key.u.ipv6.dst,
22f0708a
PA		 3197 &info->key.u.ipv6.src, dport, sport,
3198 &info->dst_cache, info);
239e944f
JB		 3199 if (IS_ERR(ndst))
3200 return PTR_ERR(ndst);
3201 dst_release(ndst);
239e944f
JB		 3202#else /* !CONFIG_IPV6 */
3203 return -EPFNOSUPPORT;
3204#endif
3205 }
1a8496ba
JB		 3206 info->key.tp_src = sport;
3207 info->key.tp_dst = dport;
239e944f 3208 return 0;
fc4099f1
PS		 3209}
3210
0c867c9b 3211static const struct net_device_ops vxlan_netdev_ether_ops = {
d342894c 3212 .ndo_init = vxlan_init,
ebf4063e 3213 .ndo_uninit = vxlan_uninit,
d342894c 3214 .ndo_open = vxlan_open,
3215 .ndo_stop = vxlan_stop,
3216 .ndo_start_xmit = vxlan_xmit,
e8171045 3217 .ndo_get_stats64 = ip_tunnel_get_stats64,
d342894c 3218 .ndo_set_rx_mode = vxlan_set_multicast_list,
345010b5 3219 .ndo_change_mtu = vxlan_change_mtu,
d342894c 3220 .ndo_validate_addr = eth_validate_addr,
3221 .ndo_set_mac_address = eth_mac_addr,
3222 .ndo_fdb_add = vxlan_fdb_add,
3223 .ndo_fdb_del = vxlan_fdb_delete,
3224 .ndo_fdb_dump = vxlan_fdb_dump,
474c3c89 3225 .ndo_fdb_get = vxlan_fdb_get,
fc4099f1 3226 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
8f1af75d 3227 .ndo_change_proto_down = dev_change_proto_down_generic,
d342894c 3228};
3229
e1e5314d
JB		 3230static const struct net_device_ops vxlan_netdev_raw_ops = {
3231 .ndo_init = vxlan_init,
3232 .ndo_uninit = vxlan_uninit,
3233 .ndo_open = vxlan_open,
3234 .ndo_stop = vxlan_stop,
3235 .ndo_start_xmit = vxlan_xmit,
3236 .ndo_get_stats64 = ip_tunnel_get_stats64,
3237 .ndo_change_mtu = vxlan_change_mtu,
3238 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
3239};
3240
d342894c 3241/* Info for udev, that this is a virtual tunnel endpoint */
3242static struct device_type vxlan_type = {
3243 .name = "vxlan",
3244};
3245
e5de25dc 3246/* Calls the ndo_udp_tunnel_add of the caller in order to
35e42379 3247 * supply the listening VXLAN udp ports. Callers are expected
e5de25dc 3248 * to implement the ndo_udp_tunnel_add.
53cf5275 3249 */
2d2b13fc 3250static void vxlan_offload_rx_ports(struct net_device *dev, bool push)
53cf5275
JG		 3251{
3252 struct vxlan_sock *vs;
3253 struct net *net = dev_net(dev);
3254 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
35e42379 3255 unsigned int i;
53cf5275
JG		 3256
3257 spin_lock(&vn->sock_lock);
3258 for (i = 0; i < PORT_HASH_SIZE; ++i) {
2d2b13fc
SD		 3259 hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) {
3260 unsigned short type;
3261
3262 if (vs->flags & VXLAN_F_GPE)
3263 type = UDP_TUNNEL_TYPE_VXLAN_GPE;
3264 else
3265 type = UDP_TUNNEL_TYPE_VXLAN;
3266
3267 if (push)
3268 udp_tunnel_push_rx_port(dev, vs->sock, type);
3269 else
3270 udp_tunnel_drop_rx_port(dev, vs->sock, type);
3271 }
53cf5275
JG		 3272 }
3273 spin_unlock(&vn->sock_lock);
3274}
53cf5275 3275
d342894c 3276/* Initialize the device structure. */
3277static void vxlan_setup(struct net_device *dev)
3278{
3279 struct vxlan_dev *vxlan = netdev_priv(dev);
31fec5aa 3280 unsigned int h;
d342894c 3281
65226ef8
JB		 3282 eth_hw_addr_random(dev);
3283 ether_setup(dev);
3284
cf124db5 3285 dev->needs_free_netdev = true;
d342894c 3286 SET_NETDEV_DEVTYPE(dev, &vxlan_type);
3287
d342894c 3288 dev->features |= NETIF_F_LLTX;
d6727fe3 3289 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
0afb1666 3290 dev->features |= NETIF_F_RXCSUM;
05c0db08 3291 dev->features |= NETIF_F_GSO_SOFTWARE;
0afb1666 3292
1eaa8178 3293 dev->vlan_features = dev->features;
0afb1666 3294 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM;
05c0db08 3295 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
02875878 3296 netif_keep_dst(dev);
0c867c9b 3297 dev->priv_flags |= IFF_NO_QUEUE;
d342894c 3298
a985343b
MS		 3299 /* MTU range: 68 - 65535 */
3300 dev->min_mtu = ETH_MIN_MTU;
3301 dev->max_mtu = ETH_MAX_MTU;
3302
553675fb 3303 INIT_LIST_HEAD(&vxlan->next);
d342894c 3304
df7e828c 3305 timer_setup(&vxlan->age_timer, vxlan_cleanup, TIMER_DEFERRABLE);
d342894c 3306
3307 vxlan->dev = dev;
3308
fe1e0713
L		 3309 for (h = 0; h < FDB_HASH_SIZE; ++h) {
3310 spin_lock_init(&vxlan->hash_lock[h]);
d342894c 3311 INIT_HLIST_HEAD(&vxlan->fdb_head[h]);
fe1e0713 3312 }
d342894c 3313}
3314
0c867c9b
JB		 3315static void vxlan_ether_setup(struct net_device *dev)
3316{
0c867c9b
JB		 3317 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
3318 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
3319 dev->netdev_ops = &vxlan_netdev_ether_ops;
3320}
3321
e1e5314d
JB		 3322static void vxlan_raw_setup(struct net_device *dev)
3323{
65226ef8 3324 dev->header_ops = NULL;
e1e5314d
JB		 3325 dev->type = ARPHRD_NONE;
3326 dev->hard_header_len = 0;
3327 dev->addr_len = 0;
e1e5314d
JB		 3328 dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
3329 dev->netdev_ops = &vxlan_netdev_raw_ops;
3330}
3331
d342894c 3332static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = {
3333 [IFLA_VXLAN_ID] = { .type = NLA_U32 },
c593642c 3334 [IFLA_VXLAN_GROUP] = { .len = sizeof_field(struct iphdr, daddr) },
e4c7ed41 3335 [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) },
d342894c 3336 [IFLA_VXLAN_LINK] = { .type = NLA_U32 },
c593642c 3337 [IFLA_VXLAN_LOCAL] = { .len = sizeof_field(struct iphdr, saddr) },
e4c7ed41 3338 [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) },
d342894c 3339 [IFLA_VXLAN_TOS] = { .type = NLA_U8 },
3340 [IFLA_VXLAN_TTL] = { .type = NLA_U8 },
e7f70af1 3341 [IFLA_VXLAN_LABEL] = { .type = NLA_U32 },
d342894c 3342 [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 },
3343 [IFLA_VXLAN_AGEING] = { .type = NLA_U32 },
3344 [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 },
05f47d69 3345 [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) },
e4f67add
DS		 3346 [IFLA_VXLAN_PROXY] = { .type = NLA_U8 },
3347 [IFLA_VXLAN_RSC] = { .type = NLA_U8 },
3348 [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 },
3349 [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 },
f8a9b1bc 3350 [IFLA_VXLAN_COLLECT_METADATA] = { .type = NLA_U8 },
823aa873 3351 [IFLA_VXLAN_PORT] = { .type = NLA_U16 },
5c91ae08
TH		 3352 [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 },
3353 [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 },
3354 [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 },
dfd8645e
TH		 3355 [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 },
3356 [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 },
3511494c 3357 [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, },
e1e5314d 3358 [IFLA_VXLAN_GPE] = { .type = NLA_FLAG, },
0ace2ca8 3359 [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG },
72f6d71e 3360 [IFLA_VXLAN_TTL_INHERIT] = { .type = NLA_FLAG },
b4d30697 3361 [IFLA_VXLAN_DF] = { .type = NLA_U8 },
d342894c 3362};
3363
a8b8a889
MS		 3364static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[],
3365 struct netlink_ext_ack *extack)
d342894c 3366{
3367 if (tb[IFLA_ADDRESS]) {
3368 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) {
653ef6a3
GM		 3369 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_ADDRESS],
3370 "Provided link layer address is not Ethernet");
d342894c 3371 return -EINVAL;
3372 }
3373
3374 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) {
653ef6a3
GM		 3375 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_ADDRESS],
3376 "Provided Ethernet address is not unicast");
d342894c 3377 return -EADDRNOTAVAIL;
3378 }
3379 }
3380
a985343b 3381 if (tb[IFLA_MTU]) {
019b13ae 3382 u32 mtu = nla_get_u32(tb[IFLA_MTU]);
a985343b 3383
653ef6a3
GM		 3384 if (mtu < ETH_MIN_MTU || mtu > ETH_MAX_MTU) {
3385 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_MTU],
3386 "MTU must be between 68 and 65535");
a985343b 3387 return -EINVAL;
653ef6a3 3388 }
a985343b
MS		 3389 }
3390
653ef6a3
GM		 3391 if (!data) {
3392 NL_SET_ERR_MSG(extack,
3393 "Required attributes not provided to perform the operation");
d342894c 3394 return -EINVAL;
653ef6a3 3395 }
d342894c 3396
3397 if (data[IFLA_VXLAN_ID]) {
a985343b
MS		 3398 u32 id = nla_get_u32(data[IFLA_VXLAN_ID]);
3399
653ef6a3 3400 if (id >= VXLAN_N_VID) {
cc8e7c69 3401 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_ID],
653ef6a3 3402 "VXLAN ID must be lower than 16777216");
d342894c 3403 return -ERANGE;
653ef6a3 3404 }
d342894c 3405 }
3406
05f47d69 3407 if (data[IFLA_VXLAN_PORT_RANGE]) {
3408 const struct ifla_vxlan_port_range *p
3409 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
3410
3411 if (ntohs(p->high) < ntohs(p->low)) {
cc8e7c69 3412 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_PORT_RANGE],
653ef6a3 3413 "Invalid source port range");
05f47d69 3414 return -EINVAL;
3415 }
3416 }
3417
b4d30697
SB		 3418 if (data[IFLA_VXLAN_DF]) {
3419 enum ifla_vxlan_df df = nla_get_u8(data[IFLA_VXLAN_DF]);
3420
3421 if (df < 0 || df > VXLAN_DF_MAX) {
cc8e7c69 3422 NL_SET_ERR_MSG_ATTR(extack, data[IFLA_VXLAN_DF],
b4d30697
SB		 3423 "Invalid DF attribute");
3424 return -EINVAL;
3425 }
3426 }
3427
d342894c 3428 return 0;
3429}
3430
1b13c97f
YB		 3431static void vxlan_get_drvinfo(struct net_device *netdev,
3432 struct ethtool_drvinfo *drvinfo)
3433{
3434 strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version));
3435 strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver));
3436}
3437
36fe3a61
MS		 3438static int vxlan_get_link_ksettings(struct net_device *dev,
3439 struct ethtool_link_ksettings *cmd)
3440{
3441 struct vxlan_dev *vxlan = netdev_priv(dev);
3442 struct vxlan_rdst *dst = &vxlan->default_dst;
3443 struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
3444 dst->remote_ifindex);
3445
3446 if (!lowerdev) {
3447 cmd->base.duplex = DUPLEX_UNKNOWN;
3448 cmd->base.port = PORT_OTHER;
3449 cmd->base.speed = SPEED_UNKNOWN;
3450
3451 return 0;
3452 }
3453
3454 return __ethtool_get_link_ksettings(lowerdev, cmd);
3455}
3456
1b13c97f 3457static const struct ethtool_ops vxlan_ethtool_ops = {
36fe3a61
MS		 3458 .get_drvinfo = vxlan_get_drvinfo,
3459 .get_link = ethtool_op_get_link,
3460 .get_link_ksettings = vxlan_get_link_ksettings,
1b13c97f
YB		 3461};
3462
3ee64f39 3463static struct socket *vxlan_create_sock(struct net *net, bool ipv6,
aab8cc36 3464 __be16 port, u32 flags, int ifindex)
553675fb 3465{
e4c7ed41 3466 struct socket *sock;
3ee64f39
TH		 3467 struct udp_port_cfg udp_conf;
3468 int err;
e4c7ed41 3469
3ee64f39 3470 memset(&udp_conf, 0, sizeof(udp_conf));
e4c7ed41 3471
3ee64f39
TH		 3472 if (ipv6) {
3473 udp_conf.family = AF_INET6;
3ee64f39 3474 udp_conf.use_udp6_rx_checksums =
3dc2b6a8 3475 !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX);
a43a9ef6 3476 udp_conf.ipv6_v6only = 1;
3ee64f39
TH		 3477 } else {
3478 udp_conf.family = AF_INET;
e4c7ed41 3479 }
e4c7ed41 3480
3ee64f39 3481 udp_conf.local_udp_port = port;
aab8cc36 3482 udp_conf.bind_ifindex = ifindex;
553675fb 3483
3ee64f39
TH		 3484 /* Open UDP socket */
3485 err = udp_sock_create(net, &udp_conf, &sock);
3486 if (err < 0)
3487 return ERR_PTR(err);
e4c7ed41 3488
39deb2c7 3489 return sock;
e4c7ed41
CW		 3490}
3491
3492/* Create new listen socket if needed */
b1be00a6 3493static struct vxlan_sock *vxlan_socket_create(struct net *net, bool ipv6,
aab8cc36
AB		 3494 __be16 port, u32 flags,
3495 int ifindex)
e4c7ed41
CW		 3496{
3497 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
3498 struct vxlan_sock *vs;
3499 struct socket *sock;
e4c7ed41 3500 unsigned int h;
acbf74a7 3501 struct udp_tunnel_sock_cfg tunnel_cfg;
e4c7ed41 3502
dc01e7d3 3503 vs = kzalloc(sizeof(*vs), GFP_KERNEL);
e4c7ed41
CW		 3504 if (!vs)
3505 return ERR_PTR(-ENOMEM);
3506
3507 for (h = 0; h < VNI_HASH_SIZE; ++h)
3508 INIT_HLIST_HEAD(&vs->vni_list[h]);
3509
aab8cc36 3510 sock = vxlan_create_sock(net, ipv6, port, flags, ifindex);
39deb2c7 3511 if (IS_ERR(sock)) {
553675fb 3512 kfree(vs);
e50fddc8 3513 return ERR_CAST(sock);
553675fb 3514 }
e4c7ed41
CW		 3515
3516 vs->sock = sock;
66af846f 3517 refcount_set(&vs->refcnt, 1);
af33c1ad 3518 vs->flags = (flags & VXLAN_F_RCV_FLAGS);
553675fb 3519
9c2e24e1
PS		 3520 spin_lock(&vn->sock_lock);
3521 hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
e7b3db5e 3522 udp_tunnel_notify_add_rx_port(sock,
b9adcd69
AD		 3523 (vs->flags & VXLAN_F_GPE) ?
3524 UDP_TUNNEL_TYPE_VXLAN_GPE :
e7b3db5e 3525 UDP_TUNNEL_TYPE_VXLAN);
9c2e24e1 3526 spin_unlock(&vn->sock_lock);
553675fb 3527
3528 /* Mark socket as an encapsulation socket. */
5602c48c 3529 memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
acbf74a7
AZ		 3530 tunnel_cfg.sk_user_data = vs;
3531 tunnel_cfg.encap_type = 1;
f2d1968e 3532 tunnel_cfg.encap_rcv = vxlan_rcv;
c3a43b9f 3533 tunnel_cfg.encap_err_lookup = vxlan_err_lookup;
acbf74a7 3534 tunnel_cfg.encap_destroy = NULL;
5602c48c
TH		 3535 tunnel_cfg.gro_receive = vxlan_gro_receive;
3536 tunnel_cfg.gro_complete = vxlan_gro_complete;
acbf74a7
AZ		 3537
3538 setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
e4c7ed41 3539
9c2e24e1
PS		 3540 return vs;
3541}
3542
b1be00a6 3543static int __vxlan_sock_add(struct vxlan_dev *vxlan, bool ipv6)
9c2e24e1 3544{
205f356d
JB		 3545 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
3546 struct vxlan_sock *vs = NULL;
69e76661 3547 struct vxlan_dev_node *node;
aab8cc36
AB		 3548 int l3mdev_index = 0;
3549
3550 if (vxlan->cfg.remote_ifindex)
3551 l3mdev_index = l3mdev_master_upper_ifindex_by_index(
3552 vxlan->net, vxlan->cfg.remote_ifindex);
9c2e24e1 3553
205f356d 3554 if (!vxlan->cfg.no_share) {
56ef9c90 3555 spin_lock(&vn->sock_lock);
205f356d 3556 vs = vxlan_find_sock(vxlan->net, ipv6 ? AF_INET6 : AF_INET,
aab8cc36
AB		 3557 vxlan->cfg.dst_port, vxlan->cfg.flags,
3558 l3mdev_index);
66af846f 3559 if (vs && !refcount_inc_not_zero(&vs->refcnt)) {
56ef9c90 3560 spin_unlock(&vn->sock_lock);
205f356d 3561 return -EBUSY;
56ef9c90
MRL		 3562 }
3563 spin_unlock(&vn->sock_lock);
3564 }
205f356d 3565 if (!vs)
b1be00a6 3566 vs = vxlan_socket_create(vxlan->net, ipv6,
aab8cc36
AB		 3567 vxlan->cfg.dst_port, vxlan->cfg.flags,
3568 l3mdev_index);
205f356d
JB		 3569 if (IS_ERR(vs))
3570 return PTR_ERR(vs);
b1be00a6 3571#if IS_ENABLED(CONFIG_IPV6)
69e76661 3572 if (ipv6) {
c6fcc4fc 3573 rcu_assign_pointer(vxlan->vn6_sock, vs);
69e76661
JB		 3574 node = &vxlan->hlist6;
3575 } else
b1be00a6 3576#endif
69e76661 3577 {
c6fcc4fc 3578 rcu_assign_pointer(vxlan->vn4_sock, vs);
69e76661
JB		 3579 node = &vxlan->hlist4;
3580 }
3581 vxlan_vs_add_dev(vs, vxlan, node);
205f356d 3582 return 0;
553675fb 3583}
3584
b1be00a6
JB		 3585static int vxlan_sock_add(struct vxlan_dev *vxlan)
3586{
dc5321d7
MS		 3587 bool metadata = vxlan->cfg.flags & VXLAN_F_COLLECT_METADATA;
3588 bool ipv6 = vxlan->cfg.flags & VXLAN_F_IPV6 || metadata;
d074bf96 3589 bool ipv4 = !ipv6 || metadata;
b1be00a6
JB		 3590 int ret = 0;
3591
c6fcc4fc 3592 RCU_INIT_POINTER(vxlan->vn4_sock, NULL);
b1be00a6 3593#if IS_ENABLED(CONFIG_IPV6)
c6fcc4fc 3594 RCU_INIT_POINTER(vxlan->vn6_sock, NULL);
d074bf96 3595 if (ipv6) {
b1be00a6 3596 ret = __vxlan_sock_add(vxlan, true);
d074bf96
JB		 3597 if (ret < 0 && ret != -EAFNOSUPPORT)
3598 ipv4 = false;
3599 }
b1be00a6 3600#endif
d074bf96 3601 if (ipv4)
b1be00a6
JB		 3602 ret = __vxlan_sock_add(vxlan, false);
3603 if (ret < 0)
3604 vxlan_sock_release(vxlan);
3605 return ret;
3606}
3607
a985343b
MS		 3608static int vxlan_config_validate(struct net *src_net, struct vxlan_config *conf,
3609 struct net_device **lower,
653ef6a3
GM		 3610 struct vxlan_dev *old,
3611 struct netlink_ext_ack *extack)
d342894c 3612{
33564bbb 3613 struct vxlan_net *vn = net_generic(src_net, vxlan_net_id);
a985343b 3614 struct vxlan_dev *tmp;
e4c7ed41 3615 bool use_ipv6 = false;
d342894c 3616
a985343b
MS		 3617 if (conf->flags & VXLAN_F_GPE) {
3618 /* For now, allow GPE only together with
3619 * COLLECT_METADATA. This can be relaxed later; in such
3620 * case, the other side of the PtP link will have to be
3621 * provided.
3622 */
3623 if ((conf->flags & ~VXLAN_F_ALLOWED_GPE) ||
3624 !(conf->flags & VXLAN_F_COLLECT_METADATA)) {
653ef6a3
GM		 3625 NL_SET_ERR_MSG(extack,
3626 "VXLAN GPE does not support this combination of attributes");
a985343b 3627 return -EINVAL;
3555621d 3628 }
e1e5314d 3629 }
0c867c9b 3630
ce44a4ae
MS		 3631 if (!conf->remote_ip.sa.sa_family && !conf->saddr.sa.sa_family) {
3632 /* Unless IPv6 is explicitly requested, assume IPv4 */
a985343b 3633 conf->remote_ip.sa.sa_family = AF_INET;
ce44a4ae
MS		 3634 conf->saddr.sa.sa_family = AF_INET;
3635 } else if (!conf->remote_ip.sa.sa_family) {
3636 conf->remote_ip.sa.sa_family = conf->saddr.sa.sa_family;
3637 } else if (!conf->saddr.sa.sa_family) {
3638 conf->saddr.sa.sa_family = conf->remote_ip.sa.sa_family;
3639 }
3640
653ef6a3
GM		 3641 if (conf->saddr.sa.sa_family != conf->remote_ip.sa.sa_family) {
3642 NL_SET_ERR_MSG(extack,
3643 "Local and remote address must be from the same family");
ce44a4ae 3644 return -EINVAL;
653ef6a3 3645 }
e4c7ed41 3646
653ef6a3
GM		 3647 if (vxlan_addr_multicast(&conf->saddr)) {
3648 NL_SET_ERR_MSG(extack, "Local address cannot be multicast");
0f22a3c6 3649 return -EINVAL;
653ef6a3 3650 }
0f22a3c6 3651
ce44a4ae 3652 if (conf->saddr.sa.sa_family == AF_INET6) {
653ef6a3
GM		 3653 if (!IS_ENABLED(CONFIG_IPV6)) {
3654 NL_SET_ERR_MSG(extack,
3655 "IPv6 support not enabled in the kernel");
057ba29b 3656 return -EPFNOSUPPORT;
653ef6a3 3657 }
e4c7ed41 3658 use_ipv6 = true;
a985343b 3659 conf->flags |= VXLAN_F_IPV6;
0f22a3c6
MS		 3660
3661 if (!(conf->flags & VXLAN_F_COLLECT_METADATA)) {
3662 int local_type =
3663 ipv6_addr_type(&conf->saddr.sin6.sin6_addr);
3664 int remote_type =
3665 ipv6_addr_type(&conf->remote_ip.sin6.sin6_addr);
3666
3667 if (local_type & IPV6_ADDR_LINKLOCAL) {
3668 if (!(remote_type & IPV6_ADDR_LINKLOCAL) &&
653ef6a3
GM		 3669 (remote_type != IPV6_ADDR_ANY)) {
3670 NL_SET_ERR_MSG(extack,
3671 "Invalid combination of local and remote address scopes");
0f22a3c6 3672 return -EINVAL;
653ef6a3 3673 }
0f22a3c6
MS		 3674
3675 conf->flags |= VXLAN_F_IPV6_LINKLOCAL;
3676 } else {
3677 if (remote_type ==
653ef6a3
GM		 3678 (IPV6_ADDR_UNICAST | IPV6_ADDR_LINKLOCAL)) {
3679 NL_SET_ERR_MSG(extack,
3680 "Invalid combination of local and remote address scopes");
0f22a3c6 3681 return -EINVAL;
653ef6a3 3682 }
0f22a3c6
MS		 3683
3684 conf->flags &= ~VXLAN_F_IPV6_LINKLOCAL;
3685 }
3686 }
057ba29b 3687 }
d342894c 3688
653ef6a3
GM		 3689 if (conf->label && !use_ipv6) {
3690 NL_SET_ERR_MSG(extack,
3691 "Label attribute only applies to IPv6 VXLAN devices");
e7f70af1 3692 return -EINVAL;
653ef6a3 3693 }
e7f70af1 3694
a985343b
MS		 3695 if (conf->remote_ifindex) {
3696 struct net_device *lowerdev;
34e02aa1 3697
a985343b 3698 lowerdev = __dev_get_by_index(src_net, conf->remote_ifindex);
653ef6a3
GM		 3699 if (!lowerdev) {
3700 NL_SET_ERR_MSG(extack,
3701 "Invalid local interface, device not found");
34e02aa1 3702 return -ENODEV;
653ef6a3 3703 }
d342894c 3704
e4c7ed41
CW		 3705#if IS_ENABLED(CONFIG_IPV6)
3706 if (use_ipv6) {
3707 struct inet6_dev *idev = __in6_dev_get(lowerdev);
653ef6a3
GM		 3708 if (idev && idev->cnf.disable_ipv6) {
3709 NL_SET_ERR_MSG(extack,
3710 "IPv6 support disabled by administrator");
e4c7ed41 3711 return -EPERM;
653ef6a3 3712 }
e4c7ed41
CW		 3713 }
3714#endif
3715
a985343b
MS		 3716 *lower = lowerdev;
3717 } else {
653ef6a3
GM		 3718 if (vxlan_addr_multicast(&conf->remote_ip)) {
3719 NL_SET_ERR_MSG(extack,
3720 "Local interface required for multicast remote destination");
3721
a985343b 3722 return -EINVAL;
653ef6a3 3723 }
1ba56fb4 3724
0f22a3c6 3725#if IS_ENABLED(CONFIG_IPV6)
653ef6a3
GM		 3726 if (conf->flags & VXLAN_F_IPV6_LINKLOCAL) {
3727 NL_SET_ERR_MSG(extack,
3728 "Local interface required for link-local local/remote addresses");
0f22a3c6 3729 return -EINVAL;
653ef6a3 3730 }
0f22a3c6
MS		 3731#endif
3732
a985343b 3733 *lower = NULL;
9dc2ad10 3734 }
d342894c 3735
a985343b
MS		 3736 if (!conf->dst_port) {
3737 if (conf->flags & VXLAN_F_GPE)
3738 conf->dst_port = htons(4790); /* IANA VXLAN-GPE port */
3739 else
3740 conf->dst_port = htons(vxlan_port);
d6acfeb1
FM		 3741 }
3742
a985343b
MS		 3743 if (!conf->age_interval)
3744 conf->age_interval = FDB_AGE_DEFAULT;
91572088 3745
a985343b
MS		 3746 list_for_each_entry(tmp, &vn->vxlan_list, next) {
3747 if (tmp == old)
3748 continue;
91572088 3749
49f810f0
MS		 3750 if (tmp->cfg.vni != conf->vni)
3751 continue;
3752 if (tmp->cfg.dst_port != conf->dst_port)
3753 continue;
3754 if ((tmp->cfg.flags & (VXLAN_F_RCV_FLAGS | VXLAN_F_IPV6)) !=
ce44a4ae 3755 (conf->flags & (VXLAN_F_RCV_FLAGS | VXLAN_F_IPV6)))
49f810f0
MS		 3756 continue;
3757
3758 if ((conf->flags & VXLAN_F_IPV6_LINKLOCAL) &&
3759 tmp->cfg.remote_ifindex != conf->remote_ifindex)
3760 continue;
3761
653ef6a3
GM		 3762 NL_SET_ERR_MSG(extack,
3763 "A VXLAN device with the specified VNI already exists");
49f810f0 3764 return -EEXIST;
a985343b 3765 }
91572088 3766
a985343b
MS		 3767 return 0;
3768}
3769
3770static void vxlan_config_apply(struct net_device *dev,
3771 struct vxlan_config *conf,
889ce937
SD		 3772 struct net_device *lowerdev,
3773 struct net *src_net,
3774 bool changelink)
a985343b
MS		 3775{
3776 struct vxlan_dev *vxlan = netdev_priv(dev);
3777 struct vxlan_rdst *dst = &vxlan->default_dst;
3778 unsigned short needed_headroom = ETH_HLEN;
3779 bool use_ipv6 = !!(conf->flags & VXLAN_F_IPV6);
3780 int max_mtu = ETH_MAX_MTU;
3781
3782 if (!changelink) {
3783 if (conf->flags & VXLAN_F_GPE)
3784 vxlan_raw_setup(dev);
3785 else
3786 vxlan_ether_setup(dev);
3787
3788 if (conf->mtu)
3789 dev->mtu = conf->mtu;
889ce937
SD		 3790
3791 vxlan->net = src_net;
a985343b
MS		 3792 }
3793
3794 dst->remote_vni = conf->vni;
3795
3796 memcpy(&dst->remote_ip, &conf->remote_ip, sizeof(conf->remote_ip));
3797
3798 if (lowerdev) {
3799 dst->remote_ifindex = conf->remote_ifindex;
3800
3801 dev->gso_max_size = lowerdev->gso_max_size;
3802 dev->gso_max_segs = lowerdev->gso_max_segs;
91572088 3803
a985343b 3804 needed_headroom = lowerdev->hard_header_len;
91572088 3805
a985343b
MS		 3806 max_mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM :
3807 VXLAN_HEADROOM);
f870c1ff
AK		 3808 if (max_mtu < ETH_MIN_MTU)
3809 max_mtu = ETH_MIN_MTU;
3810
3811 if (!changelink && !conf->mtu)
3812 dev->mtu = max_mtu;
7e059158
DW		 3813 }
3814
a985343b
MS		 3815 if (dev->mtu > max_mtu)
3816 dev->mtu = max_mtu;
3817
b1be00a6
JB		 3818 if (use_ipv6 || conf->flags & VXLAN_F_COLLECT_METADATA)
3819 needed_headroom += VXLAN6_HEADROOM;
3820 else
3821 needed_headroom += VXLAN_HEADROOM;
3822 dev->needed_headroom = needed_headroom;
3823
0dfbdf41 3824 memcpy(&vxlan->cfg, conf, sizeof(*conf));
a985343b 3825}
0dfbdf41 3826
a985343b 3827static int vxlan_dev_configure(struct net *src_net, struct net_device *dev,
653ef6a3
GM		 3828 struct vxlan_config *conf, bool changelink,
3829 struct netlink_ext_ack *extack)
a985343b
MS		 3830{
3831 struct vxlan_dev *vxlan = netdev_priv(dev);
3832 struct net_device *lowerdev;
3833 int ret;
0dfbdf41 3834
653ef6a3 3835 ret = vxlan_config_validate(src_net, conf, &lowerdev, vxlan, extack);
a985343b
MS		 3836 if (ret)
3837 return ret;
8bcdc4f3 3838
889ce937 3839 vxlan_config_apply(dev, conf, lowerdev, src_net, changelink);
0dfbdf41 3840
0dfbdf41
TG		 3841 return 0;
3842}
3843
c80498e3 3844static int __vxlan_dev_create(struct net *net, struct net_device *dev,
653ef6a3
GM		 3845 struct vxlan_config *conf,
3846 struct netlink_ext_ack *extack)
c80498e3
ND		 3847{
3848 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
3849 struct vxlan_dev *vxlan = netdev_priv(dev);
0ce1822c 3850 struct net_device *remote_dev = NULL;
0241b836 3851 struct vxlan_fdb *f = NULL;
6db92468 3852 bool unregister = false;
0ce1822c 3853 struct vxlan_rdst *dst;
c80498e3
ND		 3854 int err;
3855
0ce1822c 3856 dst = &vxlan->default_dst;
653ef6a3 3857 err = vxlan_dev_configure(net, dev, conf, false, extack);
c80498e3
ND		 3858 if (err)
3859 return err;
3860
3861 dev->ethtool_ops = &vxlan_ethtool_ops;
3862
3863 /* create an fdb entry for a valid default destination */
0ce1822c 3864 if (!vxlan_addr_any(&dst->remote_ip)) {
0241b836 3865 err = vxlan_fdb_create(vxlan, all_zeros_mac,
0ce1822c 3866 &dst->remote_ip,
c80498e3 3867 NUD_REACHABLE | NUD_PERMANENT,
c80498e3 3868 vxlan->cfg.dst_port,
0ce1822c
TY		 3869 dst->remote_vni,
3870 dst->remote_vni,
3871 dst->remote_ifindex,
1274e1cc 3872 NTF_SELF, 0, &f, extack);
c80498e3
ND		 3873 if (err)
3874 return err;
3875 }
3876
3877 err = register_netdevice(dev);
0241b836
RP		 3878 if (err)
3879 goto errout;
6db92468 3880 unregister = true;
0241b836 3881
0ce1822c
TY		 3882 if (dst->remote_ifindex) {
3883 remote_dev = __dev_get_by_index(net, dst->remote_ifindex);
3884 if (!remote_dev)
3885 goto errout;
3886
3887 err = netdev_upper_dev_link(remote_dev, dev, extack);
3888 if (err)
3889 goto errout;
3890 }
3891
0241b836 3892 err = rtnl_configure_link(dev, NULL);
6db92468 3893 if (err)
0ce1822c 3894 goto unlink;
c80498e3 3895
61f46fe8 3896 if (f) {
0ce1822c 3897 vxlan_fdb_insert(vxlan, all_zeros_mac, dst->remote_vni, f);
7c31e54a
TY		 3898
3899 /* notify default fdb entry */
61f46fe8 3900 err = vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f),
4c59b7d1 3901 RTM_NEWNEIGH, true, extack);
7c31e54a
TY		 3902 if (err) {
3903 vxlan_fdb_destroy(vxlan, f, false, false);
0ce1822c
TY		 3904 if (remote_dev)
3905 netdev_upper_dev_unlink(remote_dev, dev);
7c31e54a
TY		 3906 goto unregister;
3907 }
61f46fe8 3908 }
0241b836 3909
c80498e3 3910 list_add(&vxlan->next, &vn->vxlan_list);
0ce1822c
TY		 3911 if (remote_dev)
3912 dst->remote_dev = remote_dev;
c80498e3 3913 return 0;
0ce1822c
TY		 3914unlink:
3915 if (remote_dev)
3916 netdev_upper_dev_unlink(remote_dev, dev);
0241b836 3917errout:
6db92468
PM		 3918 /* unregister_netdevice() destroys the default FDB entry with deletion
3919 * notification. But the addition notification was not sent yet, so
3920 * destroy the entry by hand here.
3921 */
0241b836 3922 if (f)
7c31e54a
TY		 3923 __vxlan_fdb_free(f);
3924unregister:
6db92468
PM		 3925 if (unregister)
3926 unregister_netdevice(dev);
0241b836 3927 return err;
c80498e3
ND		 3928}
3929
70fb0828
RP		 3930/* Set/clear flags based on attribute */
3931static int vxlan_nl2flag(struct vxlan_config *conf, struct nlattr *tb[],
3932 int attrtype, unsigned long mask, bool changelink,
3933 bool changelink_supported,
3934 struct netlink_ext_ack *extack)
3935{
3936 unsigned long flags;
3937
3938 if (!tb[attrtype])
3939 return 0;
3940
3941 if (changelink && !changelink_supported) {
3942 vxlan_flag_attr_error(attrtype, extack);
3943 return -EOPNOTSUPP;
3944 }
3945
3946 if (vxlan_policy[attrtype].type == NLA_FLAG)
3947 flags = conf->flags | mask;
3948 else if (nla_get_u8(tb[attrtype]))
3949 flags = conf->flags | mask;
3950 else
3951 flags = conf->flags & ~mask;
3952
3953 conf->flags = flags;
3954
3955 return 0;
3956}
3957
8bcdc4f3
RP		 3958static int vxlan_nl2conf(struct nlattr *tb[], struct nlattr *data[],
3959 struct net_device *dev, struct vxlan_config *conf,
70fb0828 3960 bool changelink, struct netlink_ext_ack *extack)
0dfbdf41 3961{
8bcdc4f3 3962 struct vxlan_dev *vxlan = netdev_priv(dev);
70fb0828 3963 int err = 0;
0dfbdf41 3964
8bcdc4f3 3965 memset(conf, 0, sizeof(*conf));
e277de5f 3966
8bcdc4f3
RP		 3967 /* if changelink operation, start with old existing cfg */
3968 if (changelink)
3969 memcpy(conf, &vxlan->cfg, sizeof(*conf));
3970
3971 if (data[IFLA_VXLAN_ID]) {
3972 __be32 vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID]));
3973
70fb0828
RP		 3974 if (changelink && (vni != conf->vni)) {
3975 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_ID], "Cannot change VNI");
8bcdc4f3 3976 return -EOPNOTSUPP;
70fb0828 3977 }
8bcdc4f3
RP		 3978 conf->vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID]));
3979 }
0dfbdf41
TG		 3980
3981 if (data[IFLA_VXLAN_GROUP]) {
70fb0828
RP		 3982 if (changelink && (conf->remote_ip.sa.sa_family != AF_INET)) {
3983 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_GROUP], "New group address family does not match old group");
ce44a4ae 3984 return -EOPNOTSUPP;
70fb0828 3985 }
ce44a4ae 3986
8bcdc4f3 3987 conf->remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]);
ce44a4ae 3988 conf->remote_ip.sa.sa_family = AF_INET;
0dfbdf41 3989 } else if (data[IFLA_VXLAN_GROUP6]) {
70fb0828
RP		 3990 if (!IS_ENABLED(CONFIG_IPV6)) {
3991 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_GROUP6], "IPv6 support not enabled in the kernel");
0dfbdf41 3992 return -EPFNOSUPPORT;
70fb0828 3993 }
0dfbdf41 3994
70fb0828
RP		 3995 if (changelink && (conf->remote_ip.sa.sa_family != AF_INET6)) {
3996 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_GROUP6], "New group address family does not match old group");
ce44a4ae 3997 return -EOPNOTSUPP;
70fb0828 3998 }
ce44a4ae 3999
8bcdc4f3
RP		 4000 conf->remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]);
4001 conf->remote_ip.sa.sa_family = AF_INET6;
0dfbdf41
TG		 4002 }
4003
4004 if (data[IFLA_VXLAN_LOCAL]) {
70fb0828
RP		 4005 if (changelink && (conf->saddr.sa.sa_family != AF_INET)) {
4006 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LOCAL], "New local address family does not match old");
ce44a4ae 4007 return -EOPNOTSUPP;
70fb0828 4008 }
ce44a4ae 4009
8bcdc4f3
RP		 4010 conf->saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]);
4011 conf->saddr.sa.sa_family = AF_INET;
0dfbdf41 4012 } else if (data[IFLA_VXLAN_LOCAL6]) {
70fb0828
RP		 4013 if (!IS_ENABLED(CONFIG_IPV6)) {
4014 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LOCAL6], "IPv6 support not enabled in the kernel");
0dfbdf41 4015 return -EPFNOSUPPORT;
70fb0828 4016 }
0dfbdf41 4017
70fb0828
RP		 4018 if (changelink && (conf->saddr.sa.sa_family != AF_INET6)) {
4019 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LOCAL6], "New local address family does not match old");
ce44a4ae 4020 return -EOPNOTSUPP;
70fb0828 4021 }
ce44a4ae 4022
0dfbdf41 4023 /* TODO: respect scope id */
8bcdc4f3
RP		 4024 conf->saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]);
4025 conf->saddr.sa.sa_family = AF_INET6;
0dfbdf41
TG		 4026 }
4027
4028 if (data[IFLA_VXLAN_LINK])
8bcdc4f3 4029 conf->remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]);
0dfbdf41 4030
d342894c 4031 if (data[IFLA_VXLAN_TOS])
8bcdc4f3 4032 conf->tos = nla_get_u8(data[IFLA_VXLAN_TOS]);
d342894c 4033
afb97186 4034 if (data[IFLA_VXLAN_TTL])
8bcdc4f3 4035 conf->ttl = nla_get_u8(data[IFLA_VXLAN_TTL]);
afb97186 4036
72f6d71e 4037 if (data[IFLA_VXLAN_TTL_INHERIT]) {
70fb0828
RP		 4038 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_TTL_INHERIT,
4039 VXLAN_F_TTL_INHERIT, changelink, false,
4040 extack);
4041 if (err)
4042 return err;
4043
72f6d71e
HL		 4044 }
4045
e7f70af1 4046 if (data[IFLA_VXLAN_LABEL])
8bcdc4f3 4047 conf->label = nla_get_be32(data[IFLA_VXLAN_LABEL]) &
e7f70af1
DB		 4048 IPV6_FLOWLABEL_MASK;
4049
8bcdc4f3 4050 if (data[IFLA_VXLAN_LEARNING]) {
70fb0828
RP		 4051 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_LEARNING,
4052 VXLAN_F_LEARN, changelink, true,
4053 extack);
4054 if (err)
4055 return err;
8bcdc4f3
RP		 4056 } else if (!changelink) {
4057 /* default to learn on a new device */
4058 conf->flags |= VXLAN_F_LEARN;
4059 }
d342894c 4060
40051c4d 4061 if (data[IFLA_VXLAN_AGEING])
8bcdc4f3 4062 conf->age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]);
d342894c 4063
8bcdc4f3 4064 if (data[IFLA_VXLAN_PROXY]) {
70fb0828
RP		 4065 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_PROXY,
4066 VXLAN_F_PROXY, changelink, false,
4067 extack);
4068 if (err)
4069 return err;
8bcdc4f3 4070 }
e4f67add 4071
8bcdc4f3 4072 if (data[IFLA_VXLAN_RSC]) {
70fb0828
RP		 4073 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_RSC,
4074 VXLAN_F_RSC, changelink, false,
4075 extack);
4076 if (err)
4077 return err;
8bcdc4f3 4078 }
e4f67add 4079
8bcdc4f3 4080 if (data[IFLA_VXLAN_L2MISS]) {
70fb0828
RP		 4081 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_L2MISS,
4082 VXLAN_F_L2MISS, changelink, false,
4083 extack);
4084 if (err)
4085 return err;
8bcdc4f3 4086 }
e4f67add 4087
8bcdc4f3 4088 if (data[IFLA_VXLAN_L3MISS]) {
70fb0828
RP		 4089 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_L3MISS,
4090 VXLAN_F_L3MISS, changelink, false,
4091 extack);
4092 if (err)
4093 return err;
8bcdc4f3 4094 }
e4f67add 4095
8bcdc4f3 4096 if (data[IFLA_VXLAN_LIMIT]) {
70fb0828
RP		 4097 if (changelink) {
4098 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_LIMIT],
4099 "Cannot change limit");
8bcdc4f3 4100 return -EOPNOTSUPP;
70fb0828 4101 }
8bcdc4f3
RP		 4102 conf->addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]);
4103 }
d342894c 4104
8bcdc4f3 4105 if (data[IFLA_VXLAN_COLLECT_METADATA]) {
70fb0828
RP		 4106 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_COLLECT_METADATA,
4107 VXLAN_F_COLLECT_METADATA, changelink, false,
4108 extack);
4109 if (err)
4110 return err;
8bcdc4f3 4111 }
f8a9b1bc 4112
05f47d69 4113 if (data[IFLA_VXLAN_PORT_RANGE]) {
8bcdc4f3
RP		 4114 if (!changelink) {
4115 const struct ifla_vxlan_port_range *p
4116 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
4117 conf->port_min = ntohs(p->low);
4118 conf->port_max = ntohs(p->high);
4119 } else {
70fb0828
RP		 4120 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_PORT_RANGE],
4121 "Cannot change port range");
8bcdc4f3
RP		 4122 return -EOPNOTSUPP;
4123 }
05f47d69 4124 }
4125
8bcdc4f3 4126 if (data[IFLA_VXLAN_PORT]) {
70fb0828
RP		 4127 if (changelink) {
4128 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_PORT],
4129 "Cannot change port");
8bcdc4f3 4130 return -EOPNOTSUPP;
70fb0828 4131 }
8bcdc4f3
RP		 4132 conf->dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]);
4133 }
823aa873 4134
8bcdc4f3 4135 if (data[IFLA_VXLAN_UDP_CSUM]) {
70fb0828
RP		 4136 if (changelink) {
4137 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_VXLAN_UDP_CSUM],
4138 "Cannot change UDP_CSUM flag");
8bcdc4f3 4139 return -EOPNOTSUPP;
70fb0828 4140 }
8bcdc4f3
RP		 4141 if (!nla_get_u8(data[IFLA_VXLAN_UDP_CSUM]))
4142 conf->flags |= VXLAN_F_UDP_ZERO_CSUM_TX;
4143 }
359a0ea9 4144
8bcdc4f3 4145 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX]) {
70fb0828
RP		 4146 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_UDP_ZERO_CSUM6_TX,
4147 VXLAN_F_UDP_ZERO_CSUM6_TX, changelink,
4148 false, extack);
4149 if (err)
4150 return err;
8bcdc4f3 4151 }
359a0ea9 4152
8bcdc4f3 4153 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX]) {
70fb0828
RP		 4154 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_UDP_ZERO_CSUM6_RX,
4155 VXLAN_F_UDP_ZERO_CSUM6_RX, changelink,
4156 false, extack);
4157 if (err)
4158 return err;
8bcdc4f3 4159 }
359a0ea9 4160
8bcdc4f3 4161 if (data[IFLA_VXLAN_REMCSUM_TX]) {
70fb0828
RP		 4162 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_REMCSUM_TX,
4163 VXLAN_F_REMCSUM_TX, changelink, false,
4164 extack);
4165 if (err)
4166 return err;
8bcdc4f3 4167 }
dfd8645e 4168
8bcdc4f3 4169 if (data[IFLA_VXLAN_REMCSUM_RX]) {
70fb0828
RP		 4170 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_REMCSUM_RX,
4171 VXLAN_F_REMCSUM_RX, changelink, false,
4172 extack);
4173 if (err)
4174 return err;
8bcdc4f3
RP		 4175 }
4176
4177 if (data[IFLA_VXLAN_GBP]) {
70fb0828
RP		 4178 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_GBP,
4179 VXLAN_F_GBP, changelink, false, extack);
4180 if (err)
4181 return err;
8bcdc4f3
RP		 4182 }
4183
4184 if (data[IFLA_VXLAN_GPE]) {
70fb0828
RP		 4185 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_GPE,
4186 VXLAN_F_GPE, changelink, false,
4187 extack);
4188 if (err)
4189 return err;
8bcdc4f3
RP		 4190 }
4191
4192 if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL]) {
70fb0828
RP		 4193 err = vxlan_nl2flag(conf, data, IFLA_VXLAN_REMCSUM_NOPARTIAL,
4194 VXLAN_F_REMCSUM_NOPARTIAL, changelink,
4195 false, extack);
4196 if (err)
4197 return err;
8bcdc4f3
RP		 4198 }
4199
4200 if (tb[IFLA_MTU]) {
70fb0828
RP		 4201 if (changelink) {
4202 NL_SET_ERR_MSG_ATTR(extack, tb[IFLA_MTU],
4203 "Cannot change mtu");
8bcdc4f3 4204 return -EOPNOTSUPP;
70fb0828 4205 }
8bcdc4f3
RP		 4206 conf->mtu = nla_get_u32(tb[IFLA_MTU]);
4207 }
4208
b4d30697
SB		 4209 if (data[IFLA_VXLAN_DF])
4210 conf->df = nla_get_u8(data[IFLA_VXLAN_DF]);
4211
8bcdc4f3
RP		 4212 return 0;
4213}
4214
4215static int vxlan_newlink(struct net *src_net, struct net_device *dev,
7a3f4a18
MS		 4216 struct nlattr *tb[], struct nlattr *data[],
4217 struct netlink_ext_ack *extack)
8bcdc4f3 4218{
8bcdc4f3
RP		 4219 struct vxlan_config conf;
4220 int err;
4221
70fb0828 4222 err = vxlan_nl2conf(tb, data, dev, &conf, false, extack);
8bcdc4f3
RP		 4223 if (err)
4224 return err;
4225
653ef6a3 4226 return __vxlan_dev_create(src_net, dev, &conf, extack);
8bcdc4f3 4227}
dfd8645e 4228
8bcdc4f3 4229static int vxlan_changelink(struct net_device *dev, struct nlattr *tb[],
ad744b22
MS		 4230 struct nlattr *data[],
4231 struct netlink_ext_ack *extack)
8bcdc4f3
RP		 4232{
4233 struct vxlan_dev *vxlan = netdev_priv(dev);
8db9427d 4234 struct net_device *lowerdev;
8bcdc4f3 4235 struct vxlan_config conf;
0ce1822c 4236 struct vxlan_rdst *dst;
8bcdc4f3
RP		 4237 int err;
4238
0ce1822c 4239 dst = &vxlan->default_dst;
70fb0828 4240 err = vxlan_nl2conf(tb, data, dev, &conf, true, extack);
8bcdc4f3
RP		 4241 if (err)
4242 return err;
3511494c 4243
8db9427d
PM		 4244 err = vxlan_config_validate(vxlan->net, &conf, &lowerdev,
4245 vxlan, extack);
8bcdc4f3
RP		 4246 if (err)
4247 return err;
0ace2ca8 4248
c6761cf5
TY		 4249 if (dst->remote_dev == lowerdev)
4250 lowerdev = NULL;
4251
0ce1822c
TY		 4252 err = netdev_adjacent_change_prepare(dst->remote_dev, lowerdev, dev,
4253 extack);
4254 if (err)
4255 return err;
4256
8bcdc4f3 4257 /* handle default dst entry */
038a5a99 4258 if (!vxlan_addr_equal(&conf.remote_ip, &dst->remote_ip)) {
fe1e0713
L		 4259 u32 hash_index = fdb_head_index(vxlan, all_zeros_mac, conf.vni);
4260
4261 spin_lock_bh(&vxlan->hash_lock[hash_index]);
038a5a99 4262 if (!vxlan_addr_any(&conf.remote_ip)) {
ce5e098f 4263 err = vxlan_fdb_update(vxlan, all_zeros_mac,
038a5a99 4264 &conf.remote_ip,
8bcdc4f3 4265 NUD_REACHABLE | NUD_PERMANENT,
ce5e098f 4266 NLM_F_APPEND | NLM_F_CREATE,
8bcdc4f3 4267 vxlan->cfg.dst_port,
038a5a99
PM		 4268 conf.vni, conf.vni,
4269 conf.remote_ifindex,
1274e1cc 4270 NTF_SELF, 0, true, extack);
8bcdc4f3 4271 if (err) {
fe1e0713 4272 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
0ce1822c
TY		 4273 netdev_adjacent_change_abort(dst->remote_dev,
4274 lowerdev, dev);
8bcdc4f3
RP		 4275 return err;
4276 }
4277 }
1cdc98c2
PM		 4278 if (!vxlan_addr_any(&dst->remote_ip))
4279 __vxlan_fdb_delete(vxlan, all_zeros_mac,
4280 dst->remote_ip,
4281 vxlan->cfg.dst_port,
4282 dst->remote_vni,
4283 dst->remote_vni,
4284 dst->remote_ifindex,
4285 true);
fe1e0713 4286 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
8bcdc4f3 4287 }
ce577668 4288
038a5a99
PM		 4289 if (conf.age_interval != vxlan->cfg.age_interval)
4290 mod_timer(&vxlan->age_timer, jiffies);
4291
0ce1822c 4292 netdev_adjacent_change_commit(dst->remote_dev, lowerdev, dev);
845e0ebb 4293 if (lowerdev && lowerdev != dst->remote_dev)
0ce1822c 4294 dst->remote_dev = lowerdev;
038a5a99 4295 vxlan_config_apply(dev, &conf, lowerdev, vxlan->net, true);
8bcdc4f3 4296 return 0;
d342894c 4297}
4298
4299static void vxlan_dellink(struct net_device *dev, struct list_head *head)
4300{
4301 struct vxlan_dev *vxlan = netdev_priv(dev);
4302
8b3f9337
RP		 4303 vxlan_flush(vxlan, true);
4304
553675fb 4305 list_del(&vxlan->next);
d342894c 4306 unregister_netdevice_queue(dev, head);
0ce1822c
TY		 4307 if (vxlan->default_dst.remote_dev)
4308 netdev_upper_dev_unlink(vxlan->default_dst.remote_dev, dev);
d342894c 4309}
4310
4311static size_t vxlan_get_size(const struct net_device *dev)
4312{
4313
4314 return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */
e4c7ed41 4315 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */
d342894c 4316 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */
e4c7ed41 4317 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */
d342894c 4318 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */
8fd78069 4319 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL_INHERIT */
d342894c 4320 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */
b4d30697 4321 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_DF */
e7f70af1 4322 nla_total_size(sizeof(__be32)) + /* IFLA_VXLAN_LABEL */
d342894c 4323 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */
e4f67add
DS		 4324 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */
4325 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */
4326 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */
4327 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */
da8b43c0 4328 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_COLLECT_METADATA */
d342894c 4329 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */
4330 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */
05f47d69 4331 nla_total_size(sizeof(struct ifla_vxlan_port_range)) +
359a0ea9
TH		 4332 nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */
4333 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */
4334 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */
4335 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */
dfd8645e
TH		 4336 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */
4337 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */
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,
98c81476 4414 !!(vxlan->cfg.flags & VXLAN_F_REMCSUM_RX)))
d342894c 4415 goto nla_put_failure;
4416
05f47d69 4417 if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports))
4418 goto nla_put_failure;
4419
dc5321d7 4420 if (vxlan->cfg.flags & VXLAN_F_GBP &&
3511494c
TG		 4421 nla_put_flag(skb, IFLA_VXLAN_GBP))
4422 goto nla_put_failure;
4423
dc5321d7 4424 if (vxlan->cfg.flags & VXLAN_F_GPE &&
e1e5314d
JB		 4425 nla_put_flag(skb, IFLA_VXLAN_GPE))
4426 goto nla_put_failure;
4427
dc5321d7 4428 if (vxlan->cfg.flags & VXLAN_F_REMCSUM_NOPARTIAL &&
0ace2ca8
TH		 4429 nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL))
4430 goto nla_put_failure;
4431
d342894c 4432 return 0;
4433
4434nla_put_failure:
4435 return -EMSGSIZE;
4436}
4437
1728d4fa
ND		 4438static struct net *vxlan_get_link_net(const struct net_device *dev)
4439{
4440 struct vxlan_dev *vxlan = netdev_priv(dev);
4441
4442 return vxlan->net;
4443}
4444
d342894c 4445static struct rtnl_link_ops vxlan_link_ops __read_mostly = {
4446 .kind = "vxlan",
4447 .maxtype = IFLA_VXLAN_MAX,
4448 .policy = vxlan_policy,
4449 .priv_size = sizeof(struct vxlan_dev),
4450 .setup = vxlan_setup,
4451 .validate = vxlan_validate,
4452 .newlink = vxlan_newlink,
8bcdc4f3 4453 .changelink = vxlan_changelink,
d342894c 4454 .dellink = vxlan_dellink,
4455 .get_size = vxlan_get_size,
4456 .fill_info = vxlan_fill_info,
1728d4fa 4457 .get_link_net = vxlan_get_link_net,
d342894c 4458};
4459
cf5da330
ND		 4460struct net_device *vxlan_dev_create(struct net *net, const char *name,
4461 u8 name_assign_type,
4462 struct vxlan_config *conf)
4463{
4464 struct nlattr *tb[IFLA_MAX + 1];
4465 struct net_device *dev;
4466 int err;
4467
4468 memset(&tb, 0, sizeof(tb));
4469
4470 dev = rtnl_create_link(net, name, name_assign_type,
d0522f1c 4471 &vxlan_link_ops, tb, NULL);
cf5da330
ND		 4472 if (IS_ERR(dev))
4473 return dev;
4474
653ef6a3 4475 err = __vxlan_dev_create(net, dev, conf, NULL);
cf5da330
ND		 4476 if (err < 0) {
4477 free_netdev(dev);
4478 return ERR_PTR(err);
4479 }
4480
4481 err = rtnl_configure_link(dev, NULL);
4482 if (err < 0) {
4483 LIST_HEAD(list_kill);
4484
4485 vxlan_dellink(dev, &list_kill);
4486 unregister_netdevice_many(&list_kill);
4487 return ERR_PTR(err);
4488 }
4489
4490 return dev;
4491}
4492EXPORT_SYMBOL_GPL(vxlan_dev_create);
4493
acaf4e70
DB		 4494static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn,
4495 struct net_device *dev)
4496{
4497 struct vxlan_dev *vxlan, *next;
4498 LIST_HEAD(list_kill);
4499
4500 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
4501 struct vxlan_rdst *dst = &vxlan->default_dst;
4502
4503 /* In case we created vxlan device with carrier
4504 * and we loose the carrier due to module unload
4505 * we also need to remove vxlan device. In other
4506 * cases, it's not necessary and remote_ifindex
4507 * is 0 here, so no matches.
4508 */
4509 if (dst->remote_ifindex == dev->ifindex)
4510 vxlan_dellink(vxlan->dev, &list_kill);
4511 }
4512
4513 unregister_netdevice_many(&list_kill);
4514}
4515
b7aade15
HFS		 4516static int vxlan_netdevice_event(struct notifier_block *unused,
4517 unsigned long event, void *ptr)
acaf4e70
DB		 4518{
4519 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
783c1463 4520 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
acaf4e70 4521
04584957 4522 if (event == NETDEV_UNREGISTER) {
cc4e3835
JK		 4523 if (!dev->udp_tunnel_nic_info)
4524 vxlan_offload_rx_ports(dev, false);
acaf4e70 4525 vxlan_handle_lowerdev_unregister(vn, dev);
04584957 4526 } else if (event == NETDEV_REGISTER) {
cc4e3835
JK		 4527 if (!dev->udp_tunnel_nic_info)
4528 vxlan_offload_rx_ports(dev, true);
04584957
SD		 4529 } else if (event == NETDEV_UDP_TUNNEL_PUSH_INFO ||
4530 event == NETDEV_UDP_TUNNEL_DROP_INFO) {
2d2b13fc 4531 vxlan_offload_rx_ports(dev, event == NETDEV_UDP_TUNNEL_PUSH_INFO);
04584957 4532 }
acaf4e70
DB		 4533
4534 return NOTIFY_DONE;
4535}
4536
4537static struct notifier_block vxlan_notifier_block __read_mostly = {
b7aade15 4538 .notifier_call = vxlan_netdevice_event,
acaf4e70
DB		 4539};
4540
0efe1173
PM		 4541static void
4542vxlan_fdb_offloaded_set(struct net_device *dev,
4543 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
4544{
4545 struct vxlan_dev *vxlan = netdev_priv(dev);
4546 struct vxlan_rdst *rdst;
4547 struct vxlan_fdb *f;
fe1e0713
L		 4548 u32 hash_index;
4549
4550 hash_index = fdb_head_index(vxlan, fdb_info->eth_addr, fdb_info->vni);
0efe1173 4551
fe1e0713 4552 spin_lock_bh(&vxlan->hash_lock[hash_index]);
0efe1173
PM		 4553
4554 f = vxlan_find_mac(vxlan, fdb_info->eth_addr, fdb_info->vni);
4555 if (!f)
4556 goto out;
4557
4558 rdst = vxlan_fdb_find_rdst(f, &fdb_info->remote_ip,
4559 fdb_info->remote_port,
4560 fdb_info->remote_vni,
4561 fdb_info->remote_ifindex);
4562 if (!rdst)
4563 goto out;
4564
4565 rdst->offloaded = fdb_info->offloaded;
4566
4567out:
fe1e0713 4568 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
0efe1173
PM		 4569}
4570
5728ae0d
PM		 4571static int
4572vxlan_fdb_external_learn_add(struct net_device *dev,
4573 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
4574{
4575 struct vxlan_dev *vxlan = netdev_priv(dev);
4c59b7d1 4576 struct netlink_ext_ack *extack;
fe1e0713 4577 u32 hash_index;
5728ae0d
PM		 4578 int err;
4579
fe1e0713 4580 hash_index = fdb_head_index(vxlan, fdb_info->eth_addr, fdb_info->vni);
4c59b7d1
PM		 4581 extack = switchdev_notifier_info_to_extack(&fdb_info->info);
4582
fe1e0713 4583 spin_lock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4584 err = vxlan_fdb_update(vxlan, fdb_info->eth_addr, &fdb_info->remote_ip,
4585 NUD_REACHABLE,
4586 NLM_F_CREATE | NLM_F_REPLACE,
4587 fdb_info->remote_port,
4588 fdb_info->vni,
4589 fdb_info->remote_vni,
4590 fdb_info->remote_ifindex,
4591 NTF_USE | NTF_SELF | NTF_EXT_LEARNED,
1274e1cc 4592 0, false, extack);
fe1e0713 4593 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4594
4595 return err;
4596}
4597
4598static int
4599vxlan_fdb_external_learn_del(struct net_device *dev,
4600 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
4601{
4602 struct vxlan_dev *vxlan = netdev_priv(dev);
4603 struct vxlan_fdb *f;
fe1e0713 4604 u32 hash_index;
5728ae0d
PM		 4605 int err = 0;
4606
fe1e0713
L		 4607 hash_index = fdb_head_index(vxlan, fdb_info->eth_addr, fdb_info->vni);
4608 spin_lock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4609
4610 f = vxlan_find_mac(vxlan, fdb_info->eth_addr, fdb_info->vni);
4611 if (!f)
4612 err = -ENOENT;
4613 else if (f->flags & NTF_EXT_LEARNED)
4614 err = __vxlan_fdb_delete(vxlan, fdb_info->eth_addr,
4615 fdb_info->remote_ip,
4616 fdb_info->remote_port,
4617 fdb_info->vni,
4618 fdb_info->remote_vni,
4619 fdb_info->remote_ifindex,
4620 false);
4621
fe1e0713 4622 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
5728ae0d
PM		 4623
4624 return err;
4625}
4626
0efe1173
PM		 4627static int vxlan_switchdev_event(struct notifier_block *unused,
4628 unsigned long event, void *ptr)
4629{
4630 struct net_device *dev = switchdev_notifier_info_to_dev(ptr);
5728ae0d
PM		 4631 struct switchdev_notifier_vxlan_fdb_info *fdb_info;
4632 int err = 0;
0efe1173
PM		 4633
4634 switch (event) {
4635 case SWITCHDEV_VXLAN_FDB_OFFLOADED:
4636 vxlan_fdb_offloaded_set(dev, ptr);
4637 break;
5728ae0d
PM		 4638 case SWITCHDEV_VXLAN_FDB_ADD_TO_BRIDGE:
4639 fdb_info = ptr;
4640 err = vxlan_fdb_external_learn_add(dev, fdb_info);
4641 if (err) {
4642 err = notifier_from_errno(err);
4643 break;
4644 }
4645 fdb_info->offloaded = true;
4646 vxlan_fdb_offloaded_set(dev, fdb_info);
4647 break;
4648 case SWITCHDEV_VXLAN_FDB_DEL_TO_BRIDGE:
4649 fdb_info = ptr;
4650 err = vxlan_fdb_external_learn_del(dev, fdb_info);
4651 if (err) {
4652 err = notifier_from_errno(err);
4653 break;
4654 }
4655 fdb_info->offloaded = false;
4656 vxlan_fdb_offloaded_set(dev, fdb_info);
4657 break;
0efe1173
PM		 4658 }
4659
5728ae0d 4660 return err;
0efe1173
PM		 4661}
4662
4663static struct notifier_block vxlan_switchdev_notifier_block __read_mostly = {
4664 .notifier_call = vxlan_switchdev_event,
4665};
4666
79472fe8
RP		 4667static void vxlan_fdb_nh_flush(struct nexthop *nh)
4668{
4669 struct vxlan_fdb *fdb;
4670 struct vxlan_dev *vxlan;
4671 u32 hash_index;
4672
4673 rcu_read_lock();
4674 list_for_each_entry_rcu(fdb, &nh->fdb_list, nh_list) {
4675 vxlan = rcu_dereference(fdb->vdev);
4676 WARN_ON(!vxlan);
4677 hash_index = fdb_head_index(vxlan, fdb->eth_addr,
4678 vxlan->default_dst.remote_vni);
4679 spin_lock_bh(&vxlan->hash_lock[hash_index]);
4680 if (!hlist_unhashed(&fdb->hlist))
4681 vxlan_fdb_destroy(vxlan, fdb, false, false);
4682 spin_unlock_bh(&vxlan->hash_lock[hash_index]);
4683 }
4684 rcu_read_unlock();
4685}
4686
c7cdbe2e
RP		 4687static int vxlan_nexthop_event(struct notifier_block *nb,
4688 unsigned long event, void *ptr)
4689{
4690 struct nexthop *nh = ptr;
c7cdbe2e
RP		 4691
4692 if (!nh || event != NEXTHOP_EVENT_DEL)
4693 return NOTIFY_DONE;
4694
79472fe8 4695 vxlan_fdb_nh_flush(nh);
c7cdbe2e
RP		 4696
4697 return NOTIFY_DONE;
4698}
4699
4700static struct notifier_block vxlan_nexthop_notifier_block __read_mostly = {
4701 .notifier_call = vxlan_nexthop_event,
4702};
4703
d342894c 4704static __net_init int vxlan_init_net(struct net *net)
4705{
4706 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
31fec5aa 4707 unsigned int h;
d342894c 4708
553675fb 4709 INIT_LIST_HEAD(&vn->vxlan_list);
1c51a915 4710 spin_lock_init(&vn->sock_lock);
d342894c 4711
553675fb 4712 for (h = 0; h < PORT_HASH_SIZE; ++h)
4713 INIT_HLIST_HEAD(&vn->sock_list[h]);
d342894c 4714
c7cdbe2e 4715 return register_nexthop_notifier(net, &vxlan_nexthop_notifier_block);
d342894c 4716}
4717
57b61127 4718static void vxlan_destroy_tunnels(struct net *net, struct list_head *head)
f01ec1c0
ND		 4719{
4720 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
4721 struct vxlan_dev *vxlan, *next;
4722 struct net_device *dev, *aux;
0e4ec5ac 4723 unsigned int h;
f01ec1c0 4724
f01ec1c0
ND		 4725 for_each_netdev_safe(net, dev, aux)
4726 if (dev->rtnl_link_ops == &vxlan_link_ops)
57b61127 4727 unregister_netdevice_queue(dev, head);
f01ec1c0
ND		 4728
4729 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
4730 /* If vxlan->dev is in the same netns, it has already been added
4731 * to the list by the previous loop.
4732 */
cc4807bb 4733 if (!net_eq(dev_net(vxlan->dev), net))
57b61127 4734 unregister_netdevice_queue(vxlan->dev, head);
f01ec1c0
ND		 4735 }
4736
0e4ec5ac
VA		 4737 for (h = 0; h < PORT_HASH_SIZE; ++h)
4738 WARN_ON_ONCE(!hlist_empty(&vn->sock_list[h]));
f01ec1c0
ND		 4739}
4740
57b61127
HY		 4741static void __net_exit vxlan_exit_batch_net(struct list_head *net_list)
4742{
4743 struct net *net;
4744 LIST_HEAD(list);
4745
4746 rtnl_lock();
c7cdbe2e
RP		 4747 list_for_each_entry(net, net_list, exit_list)
4748 unregister_nexthop_notifier(net, &vxlan_nexthop_notifier_block);
57b61127
HY		 4749 list_for_each_entry(net, net_list, exit_list)
4750 vxlan_destroy_tunnels(net, &list);
4751
4752 unregister_netdevice_many(&list);
4753 rtnl_unlock();
4754}
4755
d342894c 4756static struct pernet_operations vxlan_net_ops = {
4757 .init = vxlan_init_net,
57b61127 4758 .exit_batch = vxlan_exit_batch_net,
d342894c 4759 .id = &vxlan_net_id,
4760 .size = sizeof(struct vxlan_net),
4761};
4762
4763static int __init vxlan_init_module(void)
4764{
4765 int rc;
4766
4767 get_random_bytes(&vxlan_salt, sizeof(vxlan_salt));
4768
783c1463 4769 rc = register_pernet_subsys(&vxlan_net_ops);
d342894c 4770 if (rc)
4771 goto out1;
4772
acaf4e70 4773 rc = register_netdevice_notifier(&vxlan_notifier_block);
d342894c 4774 if (rc)
4775 goto out2;
4776
0efe1173 4777 rc = register_switchdev_notifier(&vxlan_switchdev_notifier_block);
acaf4e70
DB		 4778 if (rc)
4779 goto out3;
d342894c 4780
0efe1173
PM		 4781 rc = rtnl_link_register(&vxlan_link_ops);
4782 if (rc)
4783 goto out4;
4784
acaf4e70 4785 return 0;
0efe1173
PM		 4786out4:
4787 unregister_switchdev_notifier(&vxlan_switchdev_notifier_block);
acaf4e70
DB		 4788out3:
4789 unregister_netdevice_notifier(&vxlan_notifier_block);
d342894c 4790out2:
783c1463 4791 unregister_pernet_subsys(&vxlan_net_ops);
d342894c 4792out1:
4793 return rc;
4794}
7332a13b 4795late_initcall(vxlan_init_module);
d342894c 4796
4797static void __exit vxlan_cleanup_module(void)
4798{
b7153984 4799 rtnl_link_unregister(&vxlan_link_ops);
0efe1173 4800 unregister_switchdev_notifier(&vxlan_switchdev_notifier_block);
acaf4e70 4801 unregister_netdevice_notifier(&vxlan_notifier_block);
783c1463
DB		 4802 unregister_pernet_subsys(&vxlan_net_ops);
4803 /* rcu_barrier() is called by netns */
d342894c 4804}
4805module_exit(vxlan_cleanup_module);
4806
4807MODULE_LICENSE("GPL");
4808MODULE_VERSION(VXLAN_VERSION);
3b8df3c6 4809MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>");
ead5139a 4810MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic");
d342894c 4811MODULE_ALIAS_RTNL_LINK("vxlan");
Ubuntu Jammy Linux kernel mirror