blob: da4e3d6632f647f95b6fdbc0d22abf098afdbe01 [file] [log] [blame]
stephen hemmingerd3428942012-10-01 12:32:35 +00001/*
Rami Roseneb5ce432012-11-13 13:29:15 +00002 * VXLAN: Virtual eXtensible Local Area Network
stephen hemmingerd3428942012-10-01 12:32:35 +00003 *
stephen hemminger3b8df3c2013-04-27 11:31:52 +00004 * Copyright (c) 2012-2013 Vyatta Inc.
stephen hemmingerd3428942012-10-01 12:32:35 +00005 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
stephen hemmingerd3428942012-10-01 12:32:35 +00009 */
10
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/kernel.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000014#include <linux/module.h>
15#include <linux/errno.h>
16#include <linux/slab.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000017#include <linux/udp.h>
18#include <linux/igmp.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000019#include <linux/if_ether.h>
Yan Burman1b13c972013-01-29 23:43:07 +000020#include <linux/ethtool.h>
David Stevense4f67ad2012-11-20 02:50:14 +000021#include <net/arp.h>
22#include <net/ndisc.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000023#include <net/ip.h>
24#include <net/icmp.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000025#include <net/rtnetlink.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000026#include <net/inet_ecn.h>
27#include <net/net_namespace.h>
28#include <net/netns/generic.h>
Pravin B Shelar012a5722013-08-19 11:23:07 -070029#include <net/vxlan.h>
Or Gerlitzdc01e7d2014-01-20 13:59:21 +020030#include <net/protocol.h>
stephen hemminger40d29af2016-02-09 22:07:29 -080031
Cong Wange4c7ed42013-08-31 13:44:33 +080032#if IS_ENABLED(CONFIG_IPV6)
Cong Wange4c7ed42013-08-31 13:44:33 +080033#include <net/ip6_tunnel.h>
Cong Wang660d98c2013-09-02 10:06:52 +080034#include <net/ip6_checksum.h>
Cong Wange4c7ed42013-08-31 13:44:33 +080035#endif
stephen hemmingerd3428942012-10-01 12:32:35 +000036
37#define VXLAN_VERSION "0.1"
38
stephen hemminger553675f2013-05-16 11:35:20 +000039#define PORT_HASH_BITS 8
40#define PORT_HASH_SIZE (1<<PORT_HASH_BITS)
stephen hemmingerd3428942012-10-01 12:32:35 +000041#define FDB_AGE_DEFAULT 300 /* 5 min */
42#define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */
43
stephen hemminger23c578b2013-04-27 11:31:53 +000044/* UDP port for VXLAN traffic.
45 * The IANA assigned port is 4789, but the Linux default is 8472
Stephen Hemminger234f5b72013-06-17 14:16:41 -070046 * for compatibility with early adopters.
stephen hemminger23c578b2013-04-27 11:31:53 +000047 */
Stephen Hemminger9daaa392013-06-17 14:16:12 -070048static unsigned short vxlan_port __read_mostly = 8472;
49module_param_named(udp_port, vxlan_port, ushort, 0444);
stephen hemmingerd3428942012-10-01 12:32:35 +000050MODULE_PARM_DESC(udp_port, "Destination UDP port");
51
52static bool log_ecn_error = true;
53module_param(log_ecn_error, bool, 0644);
54MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
55
Pravin B Shelar60d9d4c2013-06-20 00:26:31 -070056static int vxlan_net_id;
Thomas Graf0dfbdf42015-07-21 10:44:02 +020057static struct rtnl_link_ops vxlan_link_ops;
stephen hemminger553675f2013-05-16 11:35:20 +000058
Li RongQing7256eac2016-01-29 09:43:47 +080059static const u8 all_zeros_mac[ETH_ALEN + 2];
Mike Rapoportafbd8ba2013-06-25 16:01:51 +030060
Jiri Benc205f3562015-09-24 13:50:01 +020061static int vxlan_sock_add(struct vxlan_dev *vxlan);
Thomas Graf614732e2015-07-21 10:44:06 +020062
stephen hemminger553675f2013-05-16 11:35:20 +000063/* per-network namespace private data for this module */
64struct vxlan_net {
65 struct list_head vxlan_list;
66 struct hlist_head sock_list[PORT_HASH_SIZE];
Stephen Hemminger1c51a912013-06-17 14:16:11 -070067 spinlock_t sock_lock;
stephen hemminger553675f2013-05-16 11:35:20 +000068};
69
stephen hemmingerd3428942012-10-01 12:32:35 +000070/* Forwarding table entry */
71struct vxlan_fdb {
72 struct hlist_node hlist; /* linked list of entries */
73 struct rcu_head rcu;
74 unsigned long updated; /* jiffies */
75 unsigned long used;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -070076 struct list_head remotes;
Sowmini Varadhan7177a3b2015-07-20 09:54:50 +020077 u8 eth_addr[ETH_ALEN];
stephen hemmingerd3428942012-10-01 12:32:35 +000078 u16 state; /* see ndm_state */
David Stevensae884082013-04-19 00:36:26 +000079 u8 flags; /* see ndm_flags */
stephen hemmingerd3428942012-10-01 12:32:35 +000080};
81
stephen hemmingerd3428942012-10-01 12:32:35 +000082/* salt for hash table */
83static u32 vxlan_salt __read_mostly;
84
Thomas Grafee122c72015-07-21 10:43:58 +020085static inline bool vxlan_collect_metadata(struct vxlan_sock *vs)
86{
Thomas Grafe7030872015-07-21 10:44:01 +020087 return vs->flags & VXLAN_F_COLLECT_METADATA ||
88 ip_tunnel_collect_metadata();
Thomas Grafee122c72015-07-21 10:43:58 +020089}
90
Cong Wange4c7ed42013-08-31 13:44:33 +080091#if IS_ENABLED(CONFIG_IPV6)
92static inline
93bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
94{
Jiri Bencf0ef3122015-03-29 16:17:37 +020095 if (a->sa.sa_family != b->sa.sa_family)
96 return false;
97 if (a->sa.sa_family == AF_INET6)
98 return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr);
99 else
100 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
Cong Wange4c7ed42013-08-31 13:44:33 +0800101}
102
103static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
104{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200105 if (ipa->sa.sa_family == AF_INET6)
106 return ipv6_addr_any(&ipa->sin6.sin6_addr);
107 else
108 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
Cong Wange4c7ed42013-08-31 13:44:33 +0800109}
110
111static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
112{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200113 if (ipa->sa.sa_family == AF_INET6)
114 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
115 else
116 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
Cong Wange4c7ed42013-08-31 13:44:33 +0800117}
118
119static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
120{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200121 if (nla_len(nla) >= sizeof(struct in6_addr)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200122 ip->sin6.sin6_addr = nla_get_in6_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200123 ip->sa.sa_family = AF_INET6;
124 return 0;
125 } else if (nla_len(nla) >= sizeof(__be32)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200126 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200127 ip->sa.sa_family = AF_INET;
128 return 0;
129 } else {
130 return -EAFNOSUPPORT;
131 }
Cong Wange4c7ed42013-08-31 13:44:33 +0800132}
133
134static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
Jiri Bencf0ef3122015-03-29 16:17:37 +0200135 const union vxlan_addr *ip)
Cong Wange4c7ed42013-08-31 13:44:33 +0800136{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200137 if (ip->sa.sa_family == AF_INET6)
Jiri Benc930345e2015-03-29 16:59:25 +0200138 return nla_put_in6_addr(skb, attr, &ip->sin6.sin6_addr);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200139 else
Jiri Benc930345e2015-03-29 16:59:25 +0200140 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +0800141}
142
143#else /* !CONFIG_IPV6 */
144
145static inline
146bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
147{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200148 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
Cong Wange4c7ed42013-08-31 13:44:33 +0800149}
150
151static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
152{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200153 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
Cong Wange4c7ed42013-08-31 13:44:33 +0800154}
155
156static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
157{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200158 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
Cong Wange4c7ed42013-08-31 13:44:33 +0800159}
160
161static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
162{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200163 if (nla_len(nla) >= sizeof(struct in6_addr)) {
164 return -EAFNOSUPPORT;
165 } else if (nla_len(nla) >= sizeof(__be32)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200166 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200167 ip->sa.sa_family = AF_INET;
168 return 0;
169 } else {
170 return -EAFNOSUPPORT;
171 }
Cong Wange4c7ed42013-08-31 13:44:33 +0800172}
173
174static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
Jiri Bencf0ef3122015-03-29 16:17:37 +0200175 const union vxlan_addr *ip)
Cong Wange4c7ed42013-08-31 13:44:33 +0800176{
Jiri Benc930345e2015-03-29 16:59:25 +0200177 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +0800178}
179#endif
180
stephen hemminger553675f2013-05-16 11:35:20 +0000181/* Virtual Network hash table head */
Jiri Benc54bfd872016-02-16 21:58:58 +0100182static inline struct hlist_head *vni_head(struct vxlan_sock *vs, __be32 vni)
stephen hemminger553675f2013-05-16 11:35:20 +0000183{
Jiri Benc54bfd872016-02-16 21:58:58 +0100184 return &vs->vni_list[hash_32((__force u32)vni, VNI_HASH_BITS)];
stephen hemminger553675f2013-05-16 11:35:20 +0000185}
186
187/* Socket hash table head */
188static inline struct hlist_head *vs_head(struct net *net, __be16 port)
stephen hemmingerd3428942012-10-01 12:32:35 +0000189{
190 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
191
stephen hemminger553675f2013-05-16 11:35:20 +0000192 return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)];
193}
194
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700195/* First remote destination for a forwarding entry.
196 * Guaranteed to be non-NULL because remotes are never deleted.
197 */
stephen hemminger5ca54612013-08-04 17:17:39 -0700198static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700199{
stephen hemminger5ca54612013-08-04 17:17:39 -0700200 return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list);
201}
202
203static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb)
204{
205 return list_first_entry(&fdb->remotes, struct vxlan_rdst, list);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700206}
207
Thomas Grafac5132d2015-01-15 03:53:56 +0100208/* Find VXLAN socket based on network namespace, address family and UDP port
209 * and enabled unshareable flags.
210 */
211static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family,
212 __be16 port, u32 flags)
stephen hemminger553675f2013-05-16 11:35:20 +0000213{
214 struct vxlan_sock *vs;
Tom Herbertaf33c1a2015-01-20 11:23:05 -0800215
216 flags &= VXLAN_F_RCV_FLAGS;
stephen hemminger553675f2013-05-16 11:35:20 +0000217
218 hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) {
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -0200219 if (inet_sk(vs->sock->sk)->inet_sport == port &&
Jiri Benc705cc622015-08-20 13:56:28 +0200220 vxlan_get_sk_family(vs) == family &&
Tom Herbertaf33c1a2015-01-20 11:23:05 -0800221 vs->flags == flags)
stephen hemminger553675f2013-05-16 11:35:20 +0000222 return vs;
223 }
224 return NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000225}
226
Jiri Benc54bfd872016-02-16 21:58:58 +0100227static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, __be32 vni)
stephen hemmingerd3428942012-10-01 12:32:35 +0000228{
229 struct vxlan_dev *vxlan;
stephen hemmingerd3428942012-10-01 12:32:35 +0000230
Jiri Bencb9167b22016-02-16 21:59:03 +0100231 /* For flow based devices, map all packets to VNI 0 */
232 if (vs->flags & VXLAN_F_COLLECT_METADATA)
233 vni = 0;
234
Jiri Benc54bfd872016-02-16 21:58:58 +0100235 hlist_for_each_entry_rcu(vxlan, vni_head(vs, vni), hlist) {
236 if (vxlan->default_dst.remote_vni == vni)
stephen hemmingerd3428942012-10-01 12:32:35 +0000237 return vxlan;
238 }
239
240 return NULL;
241}
242
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700243/* Look up VNI in a per net namespace table */
Jiri Benc54bfd872016-02-16 21:58:58 +0100244static struct vxlan_dev *vxlan_find_vni(struct net *net, __be32 vni,
Thomas Grafac5132d2015-01-15 03:53:56 +0100245 sa_family_t family, __be16 port,
246 u32 flags)
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700247{
248 struct vxlan_sock *vs;
249
Thomas Grafac5132d2015-01-15 03:53:56 +0100250 vs = vxlan_find_sock(net, family, port, flags);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700251 if (!vs)
252 return NULL;
253
Jiri Benc54bfd872016-02-16 21:58:58 +0100254 return vxlan_vs_find_vni(vs, vni);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700255}
256
stephen hemmingerd3428942012-10-01 12:32:35 +0000257/* Fill in neighbour message in skbuff. */
258static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan,
Stephen Hemminger234f5b72013-06-17 14:16:41 -0700259 const struct vxlan_fdb *fdb,
260 u32 portid, u32 seq, int type, unsigned int flags,
261 const struct vxlan_rdst *rdst)
stephen hemmingerd3428942012-10-01 12:32:35 +0000262{
263 unsigned long now = jiffies;
264 struct nda_cacheinfo ci;
265 struct nlmsghdr *nlh;
266 struct ndmsg *ndm;
David Stevense4f67ad2012-11-20 02:50:14 +0000267 bool send_ip, send_eth;
stephen hemmingerd3428942012-10-01 12:32:35 +0000268
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));
David Stevense4f67ad2012-11-20 02:50:14 +0000275
276 send_eth = send_ip = true;
277
278 if (type == RTM_GETNEIGH) {
279 ndm->ndm_family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +0800280 send_ip = !vxlan_addr_any(&rdst->remote_ip);
David Stevense4f67ad2012-11-20 02:50:14 +0000281 send_eth = !is_zero_ether_addr(fdb->eth_addr);
282 } else
283 ndm->ndm_family = AF_BRIDGE;
stephen hemmingerd3428942012-10-01 12:32:35 +0000284 ndm->ndm_state = fdb->state;
285 ndm->ndm_ifindex = vxlan->dev->ifindex;
David Stevensae884082013-04-19 00:36:26 +0000286 ndm->ndm_flags = fdb->flags;
Jun Zhao545469f2014-07-26 00:38:59 +0800287 ndm->ndm_type = RTN_UNICAST;
stephen hemmingerd3428942012-10-01 12:32:35 +0000288
Nicolas Dichtel193523b2015-01-20 15:15:47 +0100289 if (!net_eq(dev_net(vxlan->dev), vxlan->net) &&
Nicolas Dichtel49670822015-01-26 14:10:53 +0100290 nla_put_s32(skb, NDA_LINK_NETNSID,
Nicolas Dichtel7a0877d2015-05-07 11:02:49 +0200291 peernet2id_alloc(dev_net(vxlan->dev), vxlan->net)))
Nicolas Dichtel193523b2015-01-20 15:15:47 +0100292 goto nla_put_failure;
293
David Stevense4f67ad2012-11-20 02:50:14 +0000294 if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000295 goto nla_put_failure;
296
Cong Wange4c7ed42013-08-31 13:44:33 +0800297 if (send_ip && vxlan_nla_put_addr(skb, NDA_DST, &rdst->remote_ip))
David Stevens66817122013-03-15 04:35:51 +0000298 goto nla_put_failure;
299
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200300 if (rdst->remote_port && rdst->remote_port != vxlan->cfg.dst_port &&
David Stevens66817122013-03-15 04:35:51 +0000301 nla_put_be16(skb, NDA_PORT, rdst->remote_port))
302 goto nla_put_failure;
Atzm Watanabec7995c42013-04-16 02:50:52 +0000303 if (rdst->remote_vni != vxlan->default_dst.remote_vni &&
Jiri Benc54bfd872016-02-16 21:58:58 +0100304 nla_put_u32(skb, NDA_VNI, be32_to_cpu(rdst->remote_vni)))
David Stevens66817122013-03-15 04:35:51 +0000305 goto nla_put_failure;
306 if (rdst->remote_ifindex &&
307 nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +0000308 goto nla_put_failure;
309
310 ci.ndm_used = jiffies_to_clock_t(now - fdb->used);
311 ci.ndm_confirmed = 0;
312 ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated);
313 ci.ndm_refcnt = 0;
314
315 if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci))
316 goto nla_put_failure;
317
Johannes Berg053c0952015-01-16 22:09:00 +0100318 nlmsg_end(skb, nlh);
319 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +0000320
321nla_put_failure:
322 nlmsg_cancel(skb, nlh);
323 return -EMSGSIZE;
324}
325
326static inline size_t vxlan_nlmsg_size(void)
327{
328 return NLMSG_ALIGN(sizeof(struct ndmsg))
329 + nla_total_size(ETH_ALEN) /* NDA_LLADDR */
Cong Wange4c7ed42013-08-31 13:44:33 +0800330 + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */
stephen hemminger73cf3312013-04-27 11:31:54 +0000331 + nla_total_size(sizeof(__be16)) /* NDA_PORT */
David Stevens66817122013-03-15 04:35:51 +0000332 + nla_total_size(sizeof(__be32)) /* NDA_VNI */
333 + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */
Nicolas Dichtel49670822015-01-26 14:10:53 +0100334 + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */
stephen hemmingerd3428942012-10-01 12:32:35 +0000335 + nla_total_size(sizeof(struct nda_cacheinfo));
336}
337
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200338static void vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
339 struct vxlan_rdst *rd, int type)
stephen hemmingerd3428942012-10-01 12:32:35 +0000340{
341 struct net *net = dev_net(vxlan->dev);
342 struct sk_buff *skb;
343 int err = -ENOBUFS;
344
345 skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC);
346 if (skb == NULL)
347 goto errout;
348
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200349 err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd);
stephen hemmingerd3428942012-10-01 12:32:35 +0000350 if (err < 0) {
351 /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */
352 WARN_ON(err == -EMSGSIZE);
353 kfree_skb(skb);
354 goto errout;
355 }
356
357 rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC);
358 return;
359errout:
360 if (err < 0)
361 rtnl_set_sk_err(net, RTNLGRP_NEIGH, err);
362}
363
Cong Wange4c7ed42013-08-31 13:44:33 +0800364static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa)
David Stevense4f67ad2012-11-20 02:50:14 +0000365{
366 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700367 struct vxlan_fdb f = {
368 .state = NUD_STALE,
369 };
370 struct vxlan_rdst remote = {
Cong Wange4c7ed42013-08-31 13:44:33 +0800371 .remote_ip = *ipa, /* goes to NDA_DST */
Jiri Benc54bfd872016-02-16 21:58:58 +0100372 .remote_vni = cpu_to_be32(VXLAN_N_VID),
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700373 };
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700374
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200375 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH);
David Stevense4f67ad2012-11-20 02:50:14 +0000376}
377
378static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN])
379{
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700380 struct vxlan_fdb f = {
381 .state = NUD_STALE,
382 };
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200383 struct vxlan_rdst remote = { };
David Stevense4f67ad2012-11-20 02:50:14 +0000384
David Stevense4f67ad2012-11-20 02:50:14 +0000385 memcpy(f.eth_addr, eth_addr, ETH_ALEN);
386
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200387 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH);
David Stevense4f67ad2012-11-20 02:50:14 +0000388}
389
stephen hemmingerd3428942012-10-01 12:32:35 +0000390/* Hash Ethernet address */
391static u32 eth_hash(const unsigned char *addr)
392{
393 u64 value = get_unaligned((u64 *)addr);
394
395 /* only want 6 bytes */
396#ifdef __BIG_ENDIAN
stephen hemmingerd3428942012-10-01 12:32:35 +0000397 value >>= 16;
stephen hemminger321fb992012-10-09 20:35:47 +0000398#else
399 value <<= 16;
stephen hemmingerd3428942012-10-01 12:32:35 +0000400#endif
401 return hash_64(value, FDB_HASH_BITS);
402}
403
404/* Hash chain to use given mac address */
405static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan,
406 const u8 *mac)
407{
408 return &vxlan->fdb_head[eth_hash(mac)];
409}
410
411/* Look up Ethernet address in forwarding table */
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000412static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan,
stephen hemmingerd3428942012-10-01 12:32:35 +0000413 const u8 *mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000414{
415 struct hlist_head *head = vxlan_fdb_head(vxlan, mac);
416 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000417
Sasha Levinb67bfe02013-02-27 17:06:00 -0800418 hlist_for_each_entry_rcu(f, head, hlist) {
Joe Perches7367d0b2013-09-01 11:51:23 -0700419 if (ether_addr_equal(mac, f->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000420 return f;
421 }
422
423 return NULL;
424}
425
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000426static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan,
427 const u8 *mac)
428{
429 struct vxlan_fdb *f;
430
431 f = __vxlan_find_mac(vxlan, mac);
432 if (f)
433 f->used = jiffies;
434
435 return f;
436}
437
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300438/* caller should hold vxlan->hash_lock */
439static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800440 union vxlan_addr *ip, __be16 port,
Jiri Benc54bfd872016-02-16 21:58:58 +0100441 __be32 vni, __u32 ifindex)
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300442{
443 struct vxlan_rdst *rd;
444
445 list_for_each_entry(rd, &f->remotes, list) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800446 if (vxlan_addr_equal(&rd->remote_ip, ip) &&
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300447 rd->remote_port == port &&
448 rd->remote_vni == vni &&
449 rd->remote_ifindex == ifindex)
450 return rd;
451 }
452
453 return NULL;
454}
455
Thomas Richter906dc182013-07-19 17:20:07 +0200456/* Replace destination of unicast mac */
457static int vxlan_fdb_replace(struct vxlan_fdb *f,
Jiri Benc54bfd872016-02-16 21:58:58 +0100458 union vxlan_addr *ip, __be16 port, __be32 vni,
459 __u32 ifindex)
Thomas Richter906dc182013-07-19 17:20:07 +0200460{
461 struct vxlan_rdst *rd;
462
463 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
464 if (rd)
465 return 0;
466
467 rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list);
468 if (!rd)
469 return 0;
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100470
471 dst_cache_reset(&rd->dst_cache);
Cong Wange4c7ed42013-08-31 13:44:33 +0800472 rd->remote_ip = *ip;
Thomas Richter906dc182013-07-19 17:20:07 +0200473 rd->remote_port = port;
474 rd->remote_vni = vni;
475 rd->remote_ifindex = ifindex;
476 return 1;
477}
478
David Stevens66817122013-03-15 04:35:51 +0000479/* Add/update destinations for multicast */
480static int vxlan_fdb_append(struct vxlan_fdb *f,
Jiri Benc54bfd872016-02-16 21:58:58 +0100481 union vxlan_addr *ip, __be16 port, __be32 vni,
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200482 __u32 ifindex, struct vxlan_rdst **rdp)
David Stevens66817122013-03-15 04:35:51 +0000483{
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700484 struct vxlan_rdst *rd;
David Stevens66817122013-03-15 04:35:51 +0000485
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300486 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
487 if (rd)
488 return 0;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700489
David Stevens66817122013-03-15 04:35:51 +0000490 rd = kmalloc(sizeof(*rd), GFP_ATOMIC);
491 if (rd == NULL)
492 return -ENOBUFS;
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100493
494 if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) {
495 kfree(rd);
496 return -ENOBUFS;
497 }
498
Cong Wange4c7ed42013-08-31 13:44:33 +0800499 rd->remote_ip = *ip;
David Stevens66817122013-03-15 04:35:51 +0000500 rd->remote_port = port;
501 rd->remote_vni = vni;
502 rd->remote_ifindex = ifindex;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700503
504 list_add_tail_rcu(&rd->list, &f->remotes);
505
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200506 *rdp = rd;
David Stevens66817122013-03-15 04:35:51 +0000507 return 1;
508}
509
Tom Herbertdfd86452015-01-12 17:00:38 -0800510static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb,
511 unsigned int off,
512 struct vxlanhdr *vh, size_t hdrlen,
Jiri Benc54bfd872016-02-16 21:58:58 +0100513 __be32 vni_field,
514 struct gro_remcsum *grc,
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800515 bool nopartial)
Tom Herbertdfd86452015-01-12 17:00:38 -0800516{
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700517 size_t start, offset;
Tom Herbertdfd86452015-01-12 17:00:38 -0800518
519 if (skb->remcsum_offload)
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700520 return vh;
Tom Herbertdfd86452015-01-12 17:00:38 -0800521
522 if (!NAPI_GRO_CB(skb)->csum_valid)
523 return NULL;
524
Jiri Benc54bfd872016-02-16 21:58:58 +0100525 start = vxlan_rco_start(vni_field);
526 offset = start + vxlan_rco_offset(vni_field);
Tom Herbertdfd86452015-01-12 17:00:38 -0800527
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700528 vh = skb_gro_remcsum_process(skb, (void *)vh, off, hdrlen,
529 start, offset, grc, nopartial);
Tom Herbertdfd86452015-01-12 17:00:38 -0800530
531 skb->remcsum_offload = 1;
532
533 return vh;
534}
535
Tom Herbert5602c482016-04-05 08:22:53 -0700536static struct sk_buff **vxlan_gro_receive(struct sock *sk,
537 struct sk_buff **head,
538 struct sk_buff *skb)
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200539{
540 struct sk_buff *p, **pp = NULL;
541 struct vxlanhdr *vh, *vh2;
Jesse Gross9b174d82014-12-30 19:10:15 -0800542 unsigned int hlen, off_vx;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200543 int flush = 1;
Tom Herbert5602c482016-04-05 08:22:53 -0700544 struct vxlan_sock *vs = rcu_dereference_sk_user_data(sk);
Jiri Benc54bfd872016-02-16 21:58:58 +0100545 __be32 flags;
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800546 struct gro_remcsum grc;
547
548 skb_gro_remcsum_init(&grc);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200549
550 off_vx = skb_gro_offset(skb);
551 hlen = off_vx + sizeof(*vh);
552 vh = skb_gro_header_fast(skb, off_vx);
553 if (skb_gro_header_hard(skb, hlen)) {
554 vh = skb_gro_header_slow(skb, hlen, off_vx);
555 if (unlikely(!vh))
556 goto out;
557 }
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200558
Tom Herbertdfd86452015-01-12 17:00:38 -0800559 skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr));
560
Jiri Benc54bfd872016-02-16 21:58:58 +0100561 flags = vh->vx_flags;
Tom Herbertdfd86452015-01-12 17:00:38 -0800562
563 if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) {
564 vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr),
Jiri Benc54bfd872016-02-16 21:58:58 +0100565 vh->vx_vni, &grc,
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800566 !!(vs->flags &
567 VXLAN_F_REMCSUM_NOPARTIAL));
Tom Herbertdfd86452015-01-12 17:00:38 -0800568
569 if (!vh)
570 goto out;
571 }
572
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700573 skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */
574
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200575 for (p = *head; p; p = p->next) {
576 if (!NAPI_GRO_CB(p)->same_flow)
577 continue;
578
579 vh2 = (struct vxlanhdr *)(p->data + off_vx);
Thomas Graf35114942015-01-15 03:53:55 +0100580 if (vh->vx_flags != vh2->vx_flags ||
581 vh->vx_vni != vh2->vx_vni) {
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200582 NAPI_GRO_CB(p)->same_flow = 0;
583 continue;
584 }
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200585 }
586
Jesse Gross9b174d82014-12-30 19:10:15 -0800587 pp = eth_gro_receive(head, skb);
Alexander Duyckc194cf92016-03-09 09:24:23 -0800588 flush = 0;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200589
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200590out:
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800591 skb_gro_remcsum_cleanup(skb, &grc);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200592 NAPI_GRO_CB(skb)->flush |= flush;
593
594 return pp;
595}
596
Tom Herbert5602c482016-04-05 08:22:53 -0700597static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200598{
Jarno Rajahalme229740c2016-05-03 16:10:21 -0700599 /* Sets 'skb->inner_mac_header' since we are always called with
600 * 'skb->encapsulation' set.
601 */
Jesse Gross9b174d82014-12-30 19:10:15 -0800602 return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr));
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200603}
604
stephen hemmingerd3428942012-10-01 12:32:35 +0000605/* Add new entry to forwarding table -- assumes lock held */
606static int vxlan_fdb_create(struct vxlan_dev *vxlan,
Cong Wange4c7ed42013-08-31 13:44:33 +0800607 const u8 *mac, union vxlan_addr *ip,
David Stevens66817122013-03-15 04:35:51 +0000608 __u16 state, __u16 flags,
Jiri Benc54bfd872016-02-16 21:58:58 +0100609 __be16 port, __be32 vni, __u32 ifindex,
David Stevensae884082013-04-19 00:36:26 +0000610 __u8 ndm_flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000611{
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200612 struct vxlan_rdst *rd = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000613 struct vxlan_fdb *f;
614 int notify = 0;
615
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000616 f = __vxlan_find_mac(vxlan, mac);
stephen hemmingerd3428942012-10-01 12:32:35 +0000617 if (f) {
618 if (flags & NLM_F_EXCL) {
619 netdev_dbg(vxlan->dev,
620 "lost race to create %pM\n", mac);
621 return -EEXIST;
622 }
623 if (f->state != state) {
624 f->state = state;
625 f->updated = jiffies;
626 notify = 1;
627 }
David Stevensae884082013-04-19 00:36:26 +0000628 if (f->flags != ndm_flags) {
629 f->flags = ndm_flags;
630 f->updated = jiffies;
631 notify = 1;
632 }
Thomas Richter906dc182013-07-19 17:20:07 +0200633 if ((flags & NLM_F_REPLACE)) {
634 /* Only change unicasts */
635 if (!(is_multicast_ether_addr(f->eth_addr) ||
636 is_zero_ether_addr(f->eth_addr))) {
Li RongQing60840422015-04-22 15:49:10 +0800637 notify |= vxlan_fdb_replace(f, ip, port, vni,
Thomas Richter906dc182013-07-19 17:20:07 +0200638 ifindex);
Thomas Richter906dc182013-07-19 17:20:07 +0200639 } else
640 return -EOPNOTSUPP;
641 }
David Stevens66817122013-03-15 04:35:51 +0000642 if ((flags & NLM_F_APPEND) &&
Mike Rapoport58e4c762013-06-25 16:01:56 +0300643 (is_multicast_ether_addr(f->eth_addr) ||
644 is_zero_ether_addr(f->eth_addr))) {
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200645 int rc = vxlan_fdb_append(f, ip, port, vni, ifindex,
646 &rd);
David Stevens66817122013-03-15 04:35:51 +0000647
648 if (rc < 0)
649 return rc;
650 notify |= rc;
651 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000652 } else {
653 if (!(flags & NLM_F_CREATE))
654 return -ENOENT;
655
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200656 if (vxlan->cfg.addrmax &&
657 vxlan->addrcnt >= vxlan->cfg.addrmax)
stephen hemmingerd3428942012-10-01 12:32:35 +0000658 return -ENOSPC;
659
Thomas Richter906dc182013-07-19 17:20:07 +0200660 /* Disallow replace to add a multicast entry */
661 if ((flags & NLM_F_REPLACE) &&
662 (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac)))
663 return -EOPNOTSUPP;
664
Cong Wange4c7ed42013-08-31 13:44:33 +0800665 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
stephen hemmingerd3428942012-10-01 12:32:35 +0000666 f = kmalloc(sizeof(*f), GFP_ATOMIC);
667 if (!f)
668 return -ENOMEM;
669
670 notify = 1;
stephen hemmingerd3428942012-10-01 12:32:35 +0000671 f->state = state;
David Stevensae884082013-04-19 00:36:26 +0000672 f->flags = ndm_flags;
stephen hemmingerd3428942012-10-01 12:32:35 +0000673 f->updated = f->used = jiffies;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700674 INIT_LIST_HEAD(&f->remotes);
stephen hemmingerd3428942012-10-01 12:32:35 +0000675 memcpy(f->eth_addr, mac, ETH_ALEN);
676
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200677 vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700678
stephen hemmingerd3428942012-10-01 12:32:35 +0000679 ++vxlan->addrcnt;
680 hlist_add_head_rcu(&f->hlist,
681 vxlan_fdb_head(vxlan, mac));
682 }
683
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200684 if (notify) {
685 if (rd == NULL)
686 rd = first_remote_rtnl(f);
687 vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH);
688 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000689
690 return 0;
691}
692
Wei Yongjun6706c822013-04-11 19:00:35 +0000693static void vxlan_fdb_free(struct rcu_head *head)
David Stevens66817122013-03-15 04:35:51 +0000694{
695 struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700696 struct vxlan_rdst *rd, *nd;
David Stevens66817122013-03-15 04:35:51 +0000697
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100698 list_for_each_entry_safe(rd, nd, &f->remotes, list) {
699 dst_cache_destroy(&rd->dst_cache);
David Stevens66817122013-03-15 04:35:51 +0000700 kfree(rd);
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100701 }
David Stevens66817122013-03-15 04:35:51 +0000702 kfree(f);
703}
704
stephen hemmingerd3428942012-10-01 12:32:35 +0000705static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f)
706{
707 netdev_dbg(vxlan->dev,
708 "delete %pM\n", f->eth_addr);
709
710 --vxlan->addrcnt;
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200711 vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_DELNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +0000712
713 hlist_del_rcu(&f->hlist);
David Stevens66817122013-03-15 04:35:51 +0000714 call_rcu(&f->rcu, vxlan_fdb_free);
stephen hemmingerd3428942012-10-01 12:32:35 +0000715}
716
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300717static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan,
Jiri Benc54bfd872016-02-16 21:58:58 +0100718 union vxlan_addr *ip, __be16 *port, __be32 *vni,
719 u32 *ifindex)
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300720{
721 struct net *net = dev_net(vxlan->dev);
Cong Wange4c7ed42013-08-31 13:44:33 +0800722 int err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300723
724 if (tb[NDA_DST]) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800725 err = vxlan_nla_get_addr(ip, tb[NDA_DST]);
726 if (err)
727 return err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300728 } else {
Cong Wange4c7ed42013-08-31 13:44:33 +0800729 union vxlan_addr *remote = &vxlan->default_dst.remote_ip;
730 if (remote->sa.sa_family == AF_INET) {
731 ip->sin.sin_addr.s_addr = htonl(INADDR_ANY);
732 ip->sa.sa_family = AF_INET;
733#if IS_ENABLED(CONFIG_IPV6)
734 } else {
735 ip->sin6.sin6_addr = in6addr_any;
736 ip->sa.sa_family = AF_INET6;
737#endif
738 }
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300739 }
740
741 if (tb[NDA_PORT]) {
742 if (nla_len(tb[NDA_PORT]) != sizeof(__be16))
743 return -EINVAL;
744 *port = nla_get_be16(tb[NDA_PORT]);
745 } else {
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200746 *port = vxlan->cfg.dst_port;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300747 }
748
749 if (tb[NDA_VNI]) {
750 if (nla_len(tb[NDA_VNI]) != sizeof(u32))
751 return -EINVAL;
Jiri Benc54bfd872016-02-16 21:58:58 +0100752 *vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI]));
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300753 } else {
754 *vni = vxlan->default_dst.remote_vni;
755 }
756
757 if (tb[NDA_IFINDEX]) {
758 struct net_device *tdev;
759
760 if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32))
761 return -EINVAL;
762 *ifindex = nla_get_u32(tb[NDA_IFINDEX]);
Ying Xue73763942014-01-15 10:23:41 +0800763 tdev = __dev_get_by_index(net, *ifindex);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300764 if (!tdev)
765 return -EADDRNOTAVAIL;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300766 } else {
767 *ifindex = 0;
768 }
769
770 return 0;
771}
772
stephen hemmingerd3428942012-10-01 12:32:35 +0000773/* Add static entry (via netlink) */
774static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[],
775 struct net_device *dev,
Jiri Pirkof6f64242014-11-28 14:34:15 +0100776 const unsigned char *addr, u16 vid, u16 flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000777{
778 struct vxlan_dev *vxlan = netdev_priv(dev);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300779 /* struct net *net = dev_net(vxlan->dev); */
Cong Wange4c7ed42013-08-31 13:44:33 +0800780 union vxlan_addr ip;
stephen hemminger73cf3312013-04-27 11:31:54 +0000781 __be16 port;
Jiri Benc54bfd872016-02-16 21:58:58 +0100782 __be32 vni;
783 u32 ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +0000784 int err;
785
786 if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) {
787 pr_info("RTM_NEWNEIGH with invalid state %#x\n",
788 ndm->ndm_state);
789 return -EINVAL;
790 }
791
792 if (tb[NDA_DST] == NULL)
793 return -EINVAL;
794
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300795 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
796 if (err)
797 return err;
David Stevens66817122013-03-15 04:35:51 +0000798
Mike Rapoport5933a7b2014-04-01 09:23:01 +0300799 if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family)
800 return -EAFNOSUPPORT;
801
stephen hemmingerd3428942012-10-01 12:32:35 +0000802 spin_lock_bh(&vxlan->hash_lock);
Cong Wange4c7ed42013-08-31 13:44:33 +0800803 err = vxlan_fdb_create(vxlan, addr, &ip, ndm->ndm_state, flags,
stephen hemminger73cf3312013-04-27 11:31:54 +0000804 port, vni, ifindex, ndm->ndm_flags);
stephen hemmingerd3428942012-10-01 12:32:35 +0000805 spin_unlock_bh(&vxlan->hash_lock);
806
807 return err;
808}
809
810/* Delete entry (via netlink) */
Vlad Yasevich1690be62013-02-13 12:00:18 +0000811static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[],
812 struct net_device *dev,
Jiri Pirkof6f64242014-11-28 14:34:15 +0100813 const unsigned char *addr, u16 vid)
stephen hemmingerd3428942012-10-01 12:32:35 +0000814{
815 struct vxlan_dev *vxlan = netdev_priv(dev);
816 struct vxlan_fdb *f;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300817 struct vxlan_rdst *rd = NULL;
Cong Wange4c7ed42013-08-31 13:44:33 +0800818 union vxlan_addr ip;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300819 __be16 port;
Jiri Benc54bfd872016-02-16 21:58:58 +0100820 __be32 vni;
821 u32 ifindex;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300822 int err;
823
824 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
825 if (err)
826 return err;
827
828 err = -ENOENT;
stephen hemmingerd3428942012-10-01 12:32:35 +0000829
830 spin_lock_bh(&vxlan->hash_lock);
831 f = vxlan_find_mac(vxlan, addr);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300832 if (!f)
833 goto out;
834
Cong Wange4c7ed42013-08-31 13:44:33 +0800835 if (!vxlan_addr_any(&ip)) {
836 rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300837 if (!rd)
838 goto out;
stephen hemmingerd3428942012-10-01 12:32:35 +0000839 }
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300840
841 err = 0;
842
843 /* remove a destination if it's not the only one on the list,
844 * otherwise destroy the fdb entry
845 */
846 if (rd && !list_is_singular(&f->remotes)) {
847 list_del_rcu(&rd->list);
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200848 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH);
Wei Yongjundcdc7a52013-08-17 07:32:09 +0800849 kfree_rcu(rd, rcu);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300850 goto out;
851 }
852
853 vxlan_fdb_destroy(vxlan, f);
854
855out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000856 spin_unlock_bh(&vxlan->hash_lock);
857
858 return err;
859}
860
861/* Dump forwarding table */
862static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb,
Jamal Hadi Salim5d5eacb2014-07-10 07:01:58 -0400863 struct net_device *dev,
864 struct net_device *filter_dev, int idx)
stephen hemmingerd3428942012-10-01 12:32:35 +0000865{
866 struct vxlan_dev *vxlan = netdev_priv(dev);
867 unsigned int h;
868
869 for (h = 0; h < FDB_HASH_SIZE; ++h) {
870 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000871 int err;
872
Sasha Levinb67bfe02013-02-27 17:06:00 -0800873 hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) {
David Stevens66817122013-03-15 04:35:51 +0000874 struct vxlan_rdst *rd;
stephen hemmingerd3428942012-10-01 12:32:35 +0000875
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700876 list_for_each_entry_rcu(rd, &f->remotes, list) {
Atzm Watanabe07a51cd2015-08-10 23:39:09 +0900877 if (idx < cb->args[0])
878 goto skip;
879
David Stevens66817122013-03-15 04:35:51 +0000880 err = vxlan_fdb_info(skb, vxlan, f,
881 NETLINK_CB(cb->skb).portid,
882 cb->nlh->nlmsg_seq,
883 RTM_NEWNEIGH,
884 NLM_F_MULTI, rd);
MINOURA Makoto / 箕浦 真472681d2016-02-25 14:20:48 +0900885 if (err < 0) {
886 cb->args[1] = err;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700887 goto out;
MINOURA Makoto / 箕浦 真472681d2016-02-25 14:20:48 +0900888 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700889skip:
Atzm Watanabe07a51cd2015-08-10 23:39:09 +0900890 ++idx;
891 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000892 }
893 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700894out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000895 return idx;
896}
897
898/* Watch incoming packets to learn mapping between Ethernet address
899 * and Tunnel endpoint.
Simon Hormanc4b49512015-04-02 11:17:58 +0900900 * Return true if packet is bogus and should be dropped.
stephen hemmingerd3428942012-10-01 12:32:35 +0000901 */
stephen hemminger26a41ae62013-06-17 12:09:58 -0700902static bool vxlan_snoop(struct net_device *dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800903 union vxlan_addr *src_ip, const u8 *src_mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000904{
905 struct vxlan_dev *vxlan = netdev_priv(dev);
906 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000907
908 f = vxlan_find_mac(vxlan, src_mac);
909 if (likely(f)) {
stephen hemminger5ca54612013-08-04 17:17:39 -0700910 struct vxlan_rdst *rdst = first_remote_rcu(f);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700911
Cong Wange4c7ed42013-08-31 13:44:33 +0800912 if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip)))
stephen hemminger26a41ae62013-06-17 12:09:58 -0700913 return false;
914
915 /* Don't migrate static entries, drop packets */
stephen hemmingereb064c32013-06-18 14:27:01 -0700916 if (f->state & NUD_NOARP)
stephen hemminger26a41ae62013-06-17 12:09:58 -0700917 return true;
stephen hemmingerd3428942012-10-01 12:32:35 +0000918
919 if (net_ratelimit())
920 netdev_info(dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800921 "%pM migrated from %pIS to %pIS\n",
Rasmus Villemoesa4870f72015-02-07 03:17:31 +0100922 src_mac, &rdst->remote_ip.sa, &src_ip->sa);
stephen hemmingerd3428942012-10-01 12:32:35 +0000923
Cong Wange4c7ed42013-08-31 13:44:33 +0800924 rdst->remote_ip = *src_ip;
stephen hemmingerd3428942012-10-01 12:32:35 +0000925 f->updated = jiffies;
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200926 vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +0000927 } else {
928 /* learned new entry */
929 spin_lock(&vxlan->hash_lock);
stephen hemminger3bf74b12013-06-17 12:09:57 -0700930
931 /* close off race between vxlan_flush and incoming packets */
932 if (netif_running(dev))
933 vxlan_fdb_create(vxlan, src_mac, src_ip,
934 NUD_REACHABLE,
935 NLM_F_EXCL|NLM_F_CREATE,
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200936 vxlan->cfg.dst_port,
stephen hemminger3bf74b12013-06-17 12:09:57 -0700937 vxlan->default_dst.remote_vni,
938 0, NTF_SELF);
stephen hemmingerd3428942012-10-01 12:32:35 +0000939 spin_unlock(&vxlan->hash_lock);
940 }
stephen hemminger26a41ae62013-06-17 12:09:58 -0700941
942 return false;
stephen hemmingerd3428942012-10-01 12:32:35 +0000943}
944
stephen hemmingerd3428942012-10-01 12:32:35 +0000945/* See if multicast group is already in use by other ID */
Gao feng95ab0992013-12-10 16:37:33 +0800946static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev)
stephen hemmingerd3428942012-10-01 12:32:35 +0000947{
stephen hemminger553675f2013-05-16 11:35:20 +0000948 struct vxlan_dev *vxlan;
Jiri Bencb1be00a2015-09-24 13:50:02 +0200949 unsigned short family = dev->default_dst.remote_ip.sa.sa_family;
stephen hemmingerd3428942012-10-01 12:32:35 +0000950
Gao feng95ab0992013-12-10 16:37:33 +0800951 /* The vxlan_sock is only used by dev, leaving group has
952 * no effect on other vxlan devices.
953 */
Jiri Bencb1be00a2015-09-24 13:50:02 +0200954 if (family == AF_INET && dev->vn4_sock &&
955 atomic_read(&dev->vn4_sock->refcnt) == 1)
Gao feng95ab0992013-12-10 16:37:33 +0800956 return false;
Jiri Bencb1be00a2015-09-24 13:50:02 +0200957#if IS_ENABLED(CONFIG_IPV6)
958 if (family == AF_INET6 && dev->vn6_sock &&
959 atomic_read(&dev->vn6_sock->refcnt) == 1)
960 return false;
961#endif
Gao feng95ab0992013-12-10 16:37:33 +0800962
stephen hemminger553675f2013-05-16 11:35:20 +0000963 list_for_each_entry(vxlan, &vn->vxlan_list, next) {
Gao feng95ab0992013-12-10 16:37:33 +0800964 if (!netif_running(vxlan->dev) || vxlan == dev)
stephen hemminger553675f2013-05-16 11:35:20 +0000965 continue;
stephen hemmingerd3428942012-10-01 12:32:35 +0000966
Jiri Bencb1be00a2015-09-24 13:50:02 +0200967 if (family == AF_INET && vxlan->vn4_sock != dev->vn4_sock)
Gao feng95ab0992013-12-10 16:37:33 +0800968 continue;
Jiri Bencb1be00a2015-09-24 13:50:02 +0200969#if IS_ENABLED(CONFIG_IPV6)
970 if (family == AF_INET6 && vxlan->vn6_sock != dev->vn6_sock)
971 continue;
972#endif
Gao feng95ab0992013-12-10 16:37:33 +0800973
974 if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip,
975 &dev->default_dst.remote_ip))
976 continue;
977
978 if (vxlan->default_dst.remote_ifindex !=
979 dev->default_dst.remote_ifindex)
980 continue;
981
982 return true;
stephen hemminger553675f2013-05-16 11:35:20 +0000983 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000984
985 return false;
986}
987
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +0200988static bool __vxlan_sock_release_prep(struct vxlan_sock *vs)
stephen hemmingerd3428942012-10-01 12:32:35 +0000989{
Jiri Bencb1be00a2015-09-24 13:50:02 +0200990 struct vxlan_net *vn;
Pravin B Shelar012a5722013-08-19 11:23:07 -0700991
Jiri Bencb1be00a2015-09-24 13:50:02 +0200992 if (!vs)
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +0200993 return false;
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700994 if (!atomic_dec_and_test(&vs->refcnt))
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +0200995 return false;
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700996
Jiri Bencb1be00a2015-09-24 13:50:02 +0200997 vn = net_generic(sock_net(vs->sock->sk), vxlan_net_id);
Stephen Hemminger1c51a912013-06-17 14:16:11 -0700998 spin_lock(&vn->sock_lock);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700999 hlist_del_rcu(&vs->hlist);
Alexander Duycke7b3db52016-06-16 12:20:52 -07001000 udp_tunnel_notify_del_rx_port(vs->sock,
Alexander Duyckb9adcd692016-06-16 12:23:19 -07001001 (vs->flags & VXLAN_F_GPE) ?
1002 UDP_TUNNEL_TYPE_VXLAN_GPE :
Alexander Duycke7b3db52016-06-16 12:20:52 -07001003 UDP_TUNNEL_TYPE_VXLAN);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001004 spin_unlock(&vn->sock_lock);
1005
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001006 return true;
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001007}
1008
Jiri Bencb1be00a2015-09-24 13:50:02 +02001009static void vxlan_sock_release(struct vxlan_dev *vxlan)
1010{
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001011 bool ipv4 = __vxlan_sock_release_prep(vxlan->vn4_sock);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001012#if IS_ENABLED(CONFIG_IPV6)
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001013 bool ipv6 = __vxlan_sock_release_prep(vxlan->vn6_sock);
1014#endif
1015
1016 synchronize_net();
1017
1018 if (ipv4) {
1019 udp_tunnel_sock_release(vxlan->vn4_sock->sock);
1020 kfree(vxlan->vn4_sock);
1021 }
1022
1023#if IS_ENABLED(CONFIG_IPV6)
1024 if (ipv6) {
1025 udp_tunnel_sock_release(vxlan->vn6_sock->sock);
1026 kfree(vxlan->vn6_sock);
1027 }
Jiri Bencb1be00a2015-09-24 13:50:02 +02001028#endif
1029}
1030
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001031/* Update multicast group membership when first VNI on
Simon Hormanc4b49512015-04-02 11:17:58 +09001032 * multicast address is brought up
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001033 */
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001034static int vxlan_igmp_join(struct vxlan_dev *vxlan)
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001035{
Jiri Bencb1be00a2015-09-24 13:50:02 +02001036 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001037 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1038 int ifindex = vxlan->default_dst.remote_ifindex;
Marcelo Ricardo Leitner149d7542015-03-20 10:26:21 -03001039 int ret = -EINVAL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001040
Cong Wange4c7ed42013-08-31 13:44:33 +08001041 if (ip->sa.sa_family == AF_INET) {
1042 struct ip_mreqn mreq = {
1043 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1044 .imr_ifindex = ifindex,
1045 };
1046
Jiri Bencb1be00a2015-09-24 13:50:02 +02001047 sk = vxlan->vn4_sock->sock->sk;
1048 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001049 ret = ip_mc_join_group(sk, &mreq);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001050 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001051#if IS_ENABLED(CONFIG_IPV6)
1052 } else {
Jiri Bencb1be00a2015-09-24 13:50:02 +02001053 sk = vxlan->vn6_sock->sock->sk;
1054 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001055 ret = ipv6_stub->ipv6_sock_mc_join(sk, ifindex,
1056 &ip->sin6.sin6_addr);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001057 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001058#endif
1059 }
stephen hemminger3fc2de22013-07-18 08:40:15 -07001060
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001061 return ret;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001062}
1063
1064/* Inverse of vxlan_igmp_join when last VNI is brought down */
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001065static int vxlan_igmp_leave(struct vxlan_dev *vxlan)
stephen hemminger3fc2de22013-07-18 08:40:15 -07001066{
Jiri Bencb1be00a2015-09-24 13:50:02 +02001067 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001068 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1069 int ifindex = vxlan->default_dst.remote_ifindex;
Marcelo Ricardo Leitner149d7542015-03-20 10:26:21 -03001070 int ret = -EINVAL;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001071
Cong Wange4c7ed42013-08-31 13:44:33 +08001072 if (ip->sa.sa_family == AF_INET) {
1073 struct ip_mreqn mreq = {
1074 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1075 .imr_ifindex = ifindex,
1076 };
1077
Jiri Bencb1be00a2015-09-24 13:50:02 +02001078 sk = vxlan->vn4_sock->sock->sk;
1079 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001080 ret = ip_mc_leave_group(sk, &mreq);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001081 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001082#if IS_ENABLED(CONFIG_IPV6)
1083 } else {
Jiri Bencb1be00a2015-09-24 13:50:02 +02001084 sk = vxlan->vn6_sock->sock->sk;
1085 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001086 ret = ipv6_stub->ipv6_sock_mc_drop(sk, ifindex,
1087 &ip->sin6.sin6_addr);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001088 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001089#endif
1090 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001091
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001092 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00001093}
1094
Jiri Bencf14eceb2016-02-16 21:59:01 +01001095static bool vxlan_remcsum(struct vxlanhdr *unparsed,
1096 struct sk_buff *skb, u32 vxflags)
Tom Herbertdfd86452015-01-12 17:00:38 -08001097{
Jiri Benc7d34fa72016-03-21 17:50:05 +01001098 size_t start, offset;
Tom Herbertdfd86452015-01-12 17:00:38 -08001099
Jiri Bencf14eceb2016-02-16 21:59:01 +01001100 if (!(unparsed->vx_flags & VXLAN_HF_RCO) || skb->remcsum_offload)
1101 goto out;
Tom Herbertb7fe10e2015-08-19 17:07:32 -07001102
Jiri Bencf14eceb2016-02-16 21:59:01 +01001103 start = vxlan_rco_start(unparsed->vx_vni);
1104 offset = start + vxlan_rco_offset(unparsed->vx_vni);
Tom Herbertdfd86452015-01-12 17:00:38 -08001105
Jiri Benc7d34fa72016-03-21 17:50:05 +01001106 if (!pskb_may_pull(skb, offset + sizeof(u16)))
Jiri Bencbe5cfea2016-02-16 21:58:59 +01001107 return false;
Tom Herbertdfd86452015-01-12 17:00:38 -08001108
Jiri Bencbe5cfea2016-02-16 21:58:59 +01001109 skb_remcsum_process(skb, (void *)(vxlan_hdr(skb) + 1), start, offset,
1110 !!(vxflags & VXLAN_F_REMCSUM_NOPARTIAL));
Jiri Bencf14eceb2016-02-16 21:59:01 +01001111out:
1112 unparsed->vx_flags &= ~VXLAN_HF_RCO;
1113 unparsed->vx_vni &= VXLAN_VNI_MASK;
Jiri Bencbe5cfea2016-02-16 21:58:59 +01001114 return true;
Tom Herbertdfd86452015-01-12 17:00:38 -08001115}
1116
Jiri Bencf14eceb2016-02-16 21:59:01 +01001117static void vxlan_parse_gbp_hdr(struct vxlanhdr *unparsed,
Jiri Benc64f87d32016-02-23 18:02:55 +01001118 struct sk_buff *skb, u32 vxflags,
Jiri Benc10a5af22016-02-23 18:02:59 +01001119 struct vxlan_metadata *md)
Jiri Benc3288af02016-02-16 21:59:00 +01001120{
Jiri Bencf14eceb2016-02-16 21:59:01 +01001121 struct vxlanhdr_gbp *gbp = (struct vxlanhdr_gbp *)unparsed;
Jiri Benc10a5af22016-02-23 18:02:59 +01001122 struct metadata_dst *tun_dst;
Jiri Benc3288af02016-02-16 21:59:00 +01001123
Jiri Bencf14eceb2016-02-16 21:59:01 +01001124 if (!(unparsed->vx_flags & VXLAN_HF_GBP))
1125 goto out;
1126
Jiri Benc3288af02016-02-16 21:59:00 +01001127 md->gbp = ntohs(gbp->policy_id);
1128
Jiri Benc10a5af22016-02-23 18:02:59 +01001129 tun_dst = (struct metadata_dst *)skb_dst(skb);
David S. Miller810813c2016-03-08 12:34:12 -05001130 if (tun_dst) {
Jiri Benc3288af02016-02-16 21:59:00 +01001131 tun_dst->u.tun_info.key.tun_flags |= TUNNEL_VXLAN_OPT;
David S. Miller810813c2016-03-08 12:34:12 -05001132 tun_dst->u.tun_info.options_len = sizeof(*md);
1133 }
Jiri Benc3288af02016-02-16 21:59:00 +01001134 if (gbp->dont_learn)
1135 md->gbp |= VXLAN_GBP_DONT_LEARN;
1136
1137 if (gbp->policy_applied)
1138 md->gbp |= VXLAN_GBP_POLICY_APPLIED;
Jiri Bencf14eceb2016-02-16 21:59:01 +01001139
Jiri Benc64f87d32016-02-23 18:02:55 +01001140 /* In flow-based mode, GBP is carried in dst_metadata */
1141 if (!(vxflags & VXLAN_F_COLLECT_METADATA))
1142 skb->mark = md->gbp;
Jiri Bencf14eceb2016-02-16 21:59:01 +01001143out:
1144 unparsed->vx_flags &= ~VXLAN_GBP_USED_BITS;
Jiri Benc3288af02016-02-16 21:59:00 +01001145}
1146
Jiri Bence1e53142016-04-05 14:47:13 +02001147static bool vxlan_parse_gpe_hdr(struct vxlanhdr *unparsed,
Jiri Benc61618ee2016-04-11 17:06:08 +02001148 __be16 *protocol,
Jiri Bence1e53142016-04-05 14:47:13 +02001149 struct sk_buff *skb, u32 vxflags)
1150{
1151 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)unparsed;
1152
1153 /* Need to have Next Protocol set for interfaces in GPE mode. */
1154 if (!gpe->np_applied)
1155 return false;
1156 /* "The initial version is 0. If a receiver does not support the
1157 * version indicated it MUST drop the packet.
1158 */
1159 if (gpe->version != 0)
1160 return false;
1161 /* "When the O bit is set to 1, the packet is an OAM packet and OAM
1162 * processing MUST occur." However, we don't implement OAM
1163 * processing, thus drop the packet.
1164 */
1165 if (gpe->oam_flag)
1166 return false;
1167
1168 switch (gpe->next_protocol) {
1169 case VXLAN_GPE_NP_IPV4:
1170 *protocol = htons(ETH_P_IP);
1171 break;
1172 case VXLAN_GPE_NP_IPV6:
1173 *protocol = htons(ETH_P_IPV6);
1174 break;
1175 case VXLAN_GPE_NP_ETHERNET:
1176 *protocol = htons(ETH_P_TEB);
1177 break;
1178 default:
1179 return false;
1180 }
1181
1182 unparsed->vx_flags &= ~VXLAN_GPE_USED_BITS;
1183 return true;
1184}
1185
Jiri Benc1ab016e2016-02-23 18:02:56 +01001186static bool vxlan_set_mac(struct vxlan_dev *vxlan,
1187 struct vxlan_sock *vs,
1188 struct sk_buff *skb)
Thomas Graf614732e2015-07-21 10:44:06 +02001189{
Thomas Graf614732e2015-07-21 10:44:06 +02001190 union vxlan_addr saddr;
Thomas Graf614732e2015-07-21 10:44:06 +02001191
Thomas Graf614732e2015-07-21 10:44:06 +02001192 skb_reset_mac_header(skb);
Thomas Graf614732e2015-07-21 10:44:06 +02001193 skb->protocol = eth_type_trans(skb, vxlan->dev);
1194 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1195
1196 /* Ignore packet loops (and multicast echo) */
1197 if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr))
Jiri Benc1ab016e2016-02-23 18:02:56 +01001198 return false;
Thomas Graf614732e2015-07-21 10:44:06 +02001199
Jiri Benc760c6802016-02-23 18:02:57 +01001200 /* Get address from the outer IP header */
Jiri Bencce212d02015-12-07 16:29:08 +01001201 if (vxlan_get_sk_family(vs) == AF_INET) {
Jiri Benc1ab016e2016-02-23 18:02:56 +01001202 saddr.sin.sin_addr.s_addr = ip_hdr(skb)->saddr;
Thomas Graf614732e2015-07-21 10:44:06 +02001203 saddr.sa.sa_family = AF_INET;
1204#if IS_ENABLED(CONFIG_IPV6)
1205 } else {
Jiri Benc1ab016e2016-02-23 18:02:56 +01001206 saddr.sin6.sin6_addr = ipv6_hdr(skb)->saddr;
Thomas Graf614732e2015-07-21 10:44:06 +02001207 saddr.sa.sa_family = AF_INET6;
1208#endif
1209 }
1210
Jiri Benc1ab016e2016-02-23 18:02:56 +01001211 if ((vxlan->flags & VXLAN_F_LEARN) &&
1212 vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source))
1213 return false;
1214
1215 return true;
1216}
1217
Jiri Benc760c6802016-02-23 18:02:57 +01001218static bool vxlan_ecn_decapsulate(struct vxlan_sock *vs, void *oiph,
1219 struct sk_buff *skb)
1220{
1221 int err = 0;
1222
1223 if (vxlan_get_sk_family(vs) == AF_INET)
1224 err = IP_ECN_decapsulate(oiph, skb);
1225#if IS_ENABLED(CONFIG_IPV6)
1226 else
1227 err = IP6_ECN_decapsulate(oiph, skb);
1228#endif
1229
1230 if (unlikely(err) && log_ecn_error) {
1231 if (vxlan_get_sk_family(vs) == AF_INET)
1232 net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n",
1233 &((struct iphdr *)oiph)->saddr,
1234 ((struct iphdr *)oiph)->tos);
1235 else
1236 net_info_ratelimited("non-ECT from %pI6\n",
1237 &((struct ipv6hdr *)oiph)->saddr);
1238 }
1239 return err <= 1;
1240}
1241
stephen hemmingerd3428942012-10-01 12:32:35 +00001242/* Callback from net/ipv4/udp.c to receive packets */
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001243static int vxlan_rcv(struct sock *sk, struct sk_buff *skb)
stephen hemmingerd3428942012-10-01 12:32:35 +00001244{
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001245 struct pcpu_sw_netstats *stats;
Jiri Bencc9e78ef2016-02-18 11:22:51 +01001246 struct vxlan_dev *vxlan;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001247 struct vxlan_sock *vs;
Jiri Bencf14eceb2016-02-16 21:59:01 +01001248 struct vxlanhdr unparsed;
Thomas Grafee122c72015-07-21 10:43:58 +02001249 struct vxlan_metadata _md;
1250 struct vxlan_metadata *md = &_md;
Jiri Benc61618ee2016-04-11 17:06:08 +02001251 __be16 protocol = htons(ETH_P_TEB);
Jiri Bence1e53142016-04-05 14:47:13 +02001252 bool raw_proto = false;
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001253 void *oiph;
stephen hemmingerd3428942012-10-01 12:32:35 +00001254
Jiri Bence1e53142016-04-05 14:47:13 +02001255 /* Need UDP and VXLAN header to be present */
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001256 if (!pskb_may_pull(skb, VXLAN_HLEN))
Hannes Frederic Sowae5aed002016-05-19 15:58:33 +02001257 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001258
Jiri Bencf14eceb2016-02-16 21:59:01 +01001259 unparsed = *vxlan_hdr(skb);
Jiri Benc288b01c2016-02-16 21:59:02 +01001260 /* VNI flag always required to be set */
1261 if (!(unparsed.vx_flags & VXLAN_HF_VNI)) {
1262 netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n",
1263 ntohl(vxlan_hdr(skb)->vx_flags),
1264 ntohl(vxlan_hdr(skb)->vx_vni));
1265 /* Return non vxlan pkt */
Hannes Frederic Sowae5aed002016-05-19 15:58:33 +02001266 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001267 }
Jiri Benc288b01c2016-02-16 21:59:02 +01001268 unparsed.vx_flags &= ~VXLAN_HF_VNI;
1269 unparsed.vx_vni &= ~VXLAN_VNI_MASK;
stephen hemmingerd3428942012-10-01 12:32:35 +00001270
Pravin B Shelar559835ea2013-09-24 10:25:40 -07001271 vs = rcu_dereference_sk_user_data(sk);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001272 if (!vs)
stephen hemmingerd3428942012-10-01 12:32:35 +00001273 goto drop;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001274
Jiri Bencc9e78ef2016-02-18 11:22:51 +01001275 vxlan = vxlan_vs_find_vni(vs, vxlan_vni(vxlan_hdr(skb)->vx_vni));
1276 if (!vxlan)
1277 goto drop;
1278
Jiri Bence1e53142016-04-05 14:47:13 +02001279 /* For backwards compatibility, only allow reserved fields to be
1280 * used by VXLAN extensions if explicitly requested.
1281 */
1282 if (vs->flags & VXLAN_F_GPE) {
1283 if (!vxlan_parse_gpe_hdr(&unparsed, &protocol, skb, vs->flags))
1284 goto drop;
1285 raw_proto = true;
1286 }
1287
1288 if (__iptunnel_pull_header(skb, VXLAN_HLEN, protocol, raw_proto,
1289 !net_eq(vxlan->net, dev_net(vxlan->dev))))
1290 goto drop;
Jiri Bencc9e78ef2016-02-18 11:22:51 +01001291
Thomas Grafee122c72015-07-21 10:43:58 +02001292 if (vxlan_collect_metadata(vs)) {
Jiri Benc07dabf22016-02-18 19:19:29 +01001293 __be32 vni = vxlan_vni(vxlan_hdr(skb)->vx_vni);
Jiri Benc10a5af22016-02-23 18:02:59 +01001294 struct metadata_dst *tun_dst;
Jiri Benc07dabf22016-02-18 19:19:29 +01001295
Pravin B Shelarc29a70d2015-08-26 23:46:50 -07001296 tun_dst = udp_tun_rx_dst(skb, vxlan_get_sk_family(vs), TUNNEL_KEY,
Jiri Benc07dabf22016-02-18 19:19:29 +01001297 vxlan_vni_to_tun_id(vni), sizeof(*md));
Pravin B Shelarc29a70d2015-08-26 23:46:50 -07001298
Thomas Grafee122c72015-07-21 10:43:58 +02001299 if (!tun_dst)
1300 goto drop;
1301
Geert Uytterhoeven0f1b7352015-09-04 12:49:32 +02001302 md = ip_tunnel_info_opts(&tun_dst->u.tun_info);
Jiri Benc10a5af22016-02-23 18:02:59 +01001303
1304 skb_dst_set(skb, (struct dst_entry *)tun_dst);
Thomas Grafee122c72015-07-21 10:43:58 +02001305 } else {
1306 memset(md, 0, sizeof(*md));
1307 }
1308
Jiri Bencf14eceb2016-02-16 21:59:01 +01001309 if (vs->flags & VXLAN_F_REMCSUM_RX)
1310 if (!vxlan_remcsum(&unparsed, skb, vs->flags))
1311 goto drop;
1312 if (vs->flags & VXLAN_F_GBP)
Jiri Benc10a5af22016-02-23 18:02:59 +01001313 vxlan_parse_gbp_hdr(&unparsed, skb, vs->flags, md);
Jiri Bence1e53142016-04-05 14:47:13 +02001314 /* Note that GBP and GPE can never be active together. This is
1315 * ensured in vxlan_dev_configure.
1316 */
Thomas Graf35114942015-01-15 03:53:55 +01001317
Jiri Bencf14eceb2016-02-16 21:59:01 +01001318 if (unparsed.vx_flags || unparsed.vx_vni) {
Tom Herbert3bf39472015-01-08 12:31:18 -08001319 /* If there are any unprocessed flags remaining treat
1320 * this as a malformed packet. This behavior diverges from
1321 * VXLAN RFC (RFC7348) which stipulates that bits in reserved
1322 * in reserved fields are to be ignored. The approach here
Simon Hormanc4b49512015-04-02 11:17:58 +09001323 * maintains compatibility with previous stack code, and also
Tom Herbert3bf39472015-01-08 12:31:18 -08001324 * is more robust and provides a little more security in
1325 * adding extensions to VXLAN.
1326 */
Jiri Benc288b01c2016-02-16 21:59:02 +01001327 goto drop;
Tom Herbert3bf39472015-01-08 12:31:18 -08001328 }
1329
Jiri Bence1e53142016-04-05 14:47:13 +02001330 if (!raw_proto) {
1331 if (!vxlan_set_mac(vxlan, vs, skb))
1332 goto drop;
1333 } else {
Jiri Benc8be0cfa2016-05-13 10:48:42 +02001334 skb_reset_mac_header(skb);
Jiri Bence1e53142016-04-05 14:47:13 +02001335 skb->dev = vxlan->dev;
1336 skb->pkt_type = PACKET_HOST;
1337 }
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001338
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001339 oiph = skb_network_header(skb);
1340 skb_reset_network_header(skb);
1341
1342 if (!vxlan_ecn_decapsulate(vs, oiph, skb)) {
1343 ++vxlan->dev->stats.rx_frame_errors;
1344 ++vxlan->dev->stats.rx_errors;
1345 goto drop;
1346 }
1347
1348 stats = this_cpu_ptr(vxlan->dev->tstats);
1349 u64_stats_update_begin(&stats->syncp);
1350 stats->rx_packets++;
1351 stats->rx_bytes += skb->len;
1352 u64_stats_update_end(&stats->syncp);
1353
1354 gro_cells_receive(&vxlan->gro_cells, skb);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001355 return 0;
1356
1357drop:
Jiri Benc288b01c2016-02-16 21:59:02 +01001358 /* Consume bad packet */
1359 kfree_skb(skb);
1360 return 0;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001361}
1362
David Stevense4f67ad2012-11-20 02:50:14 +00001363static int arp_reduce(struct net_device *dev, struct sk_buff *skb)
1364{
1365 struct vxlan_dev *vxlan = netdev_priv(dev);
1366 struct arphdr *parp;
1367 u8 *arpptr, *sha;
1368 __be32 sip, tip;
1369 struct neighbour *n;
1370
1371 if (dev->flags & IFF_NOARP)
1372 goto out;
1373
1374 if (!pskb_may_pull(skb, arp_hdr_len(dev))) {
1375 dev->stats.tx_dropped++;
1376 goto out;
1377 }
1378 parp = arp_hdr(skb);
1379
1380 if ((parp->ar_hrd != htons(ARPHRD_ETHER) &&
1381 parp->ar_hrd != htons(ARPHRD_IEEE802)) ||
1382 parp->ar_pro != htons(ETH_P_IP) ||
1383 parp->ar_op != htons(ARPOP_REQUEST) ||
1384 parp->ar_hln != dev->addr_len ||
1385 parp->ar_pln != 4)
1386 goto out;
1387 arpptr = (u8 *)parp + sizeof(struct arphdr);
1388 sha = arpptr;
1389 arpptr += dev->addr_len; /* sha */
1390 memcpy(&sip, arpptr, sizeof(sip));
1391 arpptr += sizeof(sip);
1392 arpptr += dev->addr_len; /* tha */
1393 memcpy(&tip, arpptr, sizeof(tip));
1394
1395 if (ipv4_is_loopback(tip) ||
1396 ipv4_is_multicast(tip))
1397 goto out;
1398
1399 n = neigh_lookup(&arp_tbl, &tip, dev);
1400
1401 if (n) {
David Stevense4f67ad2012-11-20 02:50:14 +00001402 struct vxlan_fdb *f;
1403 struct sk_buff *reply;
1404
1405 if (!(n->nud_state & NUD_CONNECTED)) {
1406 neigh_release(n);
1407 goto out;
1408 }
1409
1410 f = vxlan_find_mac(vxlan, n->ha);
Cong Wange4c7ed42013-08-31 13:44:33 +08001411 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
David Stevense4f67ad2012-11-20 02:50:14 +00001412 /* bridge-local neighbor */
1413 neigh_release(n);
1414 goto out;
1415 }
1416
1417 reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
1418 n->ha, sha);
1419
1420 neigh_release(n);
1421
David Stevens73461352014-03-18 12:32:29 -04001422 if (reply == NULL)
1423 goto out;
1424
David Stevense4f67ad2012-11-20 02:50:14 +00001425 skb_reset_mac_header(reply);
1426 __skb_pull(reply, skb_network_offset(reply));
1427 reply->ip_summed = CHECKSUM_UNNECESSARY;
1428 reply->pkt_type = PACKET_HOST;
1429
1430 if (netif_rx_ni(reply) == NET_RX_DROP)
1431 dev->stats.rx_dropped++;
Cong Wange4c7ed42013-08-31 13:44:33 +08001432 } else if (vxlan->flags & VXLAN_F_L3MISS) {
1433 union vxlan_addr ipa = {
1434 .sin.sin_addr.s_addr = tip,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001435 .sin.sin_family = AF_INET,
Cong Wange4c7ed42013-08-31 13:44:33 +08001436 };
1437
1438 vxlan_ip_miss(dev, &ipa);
1439 }
David Stevense4f67ad2012-11-20 02:50:14 +00001440out:
1441 consume_skb(skb);
1442 return NETDEV_TX_OK;
1443}
1444
Cong Wangf564f452013-08-31 13:44:36 +08001445#if IS_ENABLED(CONFIG_IPV6)
David Stevens4b29dba2014-03-24 10:39:58 -04001446static struct sk_buff *vxlan_na_create(struct sk_buff *request,
1447 struct neighbour *n, bool isrouter)
1448{
1449 struct net_device *dev = request->dev;
1450 struct sk_buff *reply;
1451 struct nd_msg *ns, *na;
1452 struct ipv6hdr *pip6;
1453 u8 *daddr;
1454 int na_olen = 8; /* opt hdr + ETH_ALEN for target */
1455 int ns_olen;
1456 int i, len;
1457
1458 if (dev == NULL)
1459 return NULL;
1460
1461 len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) +
1462 sizeof(*na) + na_olen + dev->needed_tailroom;
1463 reply = alloc_skb(len, GFP_ATOMIC);
1464 if (reply == NULL)
1465 return NULL;
1466
1467 reply->protocol = htons(ETH_P_IPV6);
1468 reply->dev = dev;
1469 skb_reserve(reply, LL_RESERVED_SPACE(request->dev));
1470 skb_push(reply, sizeof(struct ethhdr));
Zhang Shengju6297b912016-03-03 01:16:54 +00001471 skb_reset_mac_header(reply);
David Stevens4b29dba2014-03-24 10:39:58 -04001472
1473 ns = (struct nd_msg *)skb_transport_header(request);
1474
1475 daddr = eth_hdr(request)->h_source;
1476 ns_olen = request->len - skb_transport_offset(request) - sizeof(*ns);
1477 for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) {
1478 if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) {
1479 daddr = ns->opt + i + sizeof(struct nd_opt_hdr);
1480 break;
1481 }
1482 }
1483
1484 /* Ethernet header */
1485 ether_addr_copy(eth_hdr(reply)->h_dest, daddr);
1486 ether_addr_copy(eth_hdr(reply)->h_source, n->ha);
1487 eth_hdr(reply)->h_proto = htons(ETH_P_IPV6);
1488 reply->protocol = htons(ETH_P_IPV6);
1489
1490 skb_pull(reply, sizeof(struct ethhdr));
Zhang Shengju6297b912016-03-03 01:16:54 +00001491 skb_reset_network_header(reply);
David Stevens4b29dba2014-03-24 10:39:58 -04001492 skb_put(reply, sizeof(struct ipv6hdr));
1493
1494 /* IPv6 header */
1495
1496 pip6 = ipv6_hdr(reply);
1497 memset(pip6, 0, sizeof(struct ipv6hdr));
1498 pip6->version = 6;
1499 pip6->priority = ipv6_hdr(request)->priority;
1500 pip6->nexthdr = IPPROTO_ICMPV6;
1501 pip6->hop_limit = 255;
1502 pip6->daddr = ipv6_hdr(request)->saddr;
1503 pip6->saddr = *(struct in6_addr *)n->primary_key;
1504
1505 skb_pull(reply, sizeof(struct ipv6hdr));
Zhang Shengju6297b912016-03-03 01:16:54 +00001506 skb_reset_transport_header(reply);
David Stevens4b29dba2014-03-24 10:39:58 -04001507
1508 na = (struct nd_msg *)skb_put(reply, sizeof(*na) + na_olen);
1509
1510 /* Neighbor Advertisement */
1511 memset(na, 0, sizeof(*na)+na_olen);
1512 na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT;
1513 na->icmph.icmp6_router = isrouter;
1514 na->icmph.icmp6_override = 1;
1515 na->icmph.icmp6_solicited = 1;
1516 na->target = ns->target;
1517 ether_addr_copy(&na->opt[2], n->ha);
1518 na->opt[0] = ND_OPT_TARGET_LL_ADDR;
1519 na->opt[1] = na_olen >> 3;
1520
1521 na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr,
1522 &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6,
1523 csum_partial(na, sizeof(*na)+na_olen, 0));
1524
1525 pip6->payload_len = htons(sizeof(*na)+na_olen);
1526
1527 skb_push(reply, sizeof(struct ipv6hdr));
1528
1529 reply->ip_summed = CHECKSUM_UNNECESSARY;
1530
1531 return reply;
1532}
1533
Cong Wangf564f452013-08-31 13:44:36 +08001534static int neigh_reduce(struct net_device *dev, struct sk_buff *skb)
1535{
1536 struct vxlan_dev *vxlan = netdev_priv(dev);
David Stevens4b29dba2014-03-24 10:39:58 -04001537 struct nd_msg *msg;
Cong Wangf564f452013-08-31 13:44:36 +08001538 const struct ipv6hdr *iphdr;
1539 const struct in6_addr *saddr, *daddr;
David Stevens4b29dba2014-03-24 10:39:58 -04001540 struct neighbour *n;
1541 struct inet6_dev *in6_dev;
Cong Wangf564f452013-08-31 13:44:36 +08001542
1543 in6_dev = __in6_dev_get(dev);
1544 if (!in6_dev)
1545 goto out;
1546
Cong Wangf564f452013-08-31 13:44:36 +08001547 iphdr = ipv6_hdr(skb);
1548 saddr = &iphdr->saddr;
1549 daddr = &iphdr->daddr;
1550
Cong Wangf564f452013-08-31 13:44:36 +08001551 msg = (struct nd_msg *)skb_transport_header(skb);
1552 if (msg->icmph.icmp6_code != 0 ||
1553 msg->icmph.icmp6_type != NDISC_NEIGHBOUR_SOLICITATION)
1554 goto out;
1555
David Stevens4b29dba2014-03-24 10:39:58 -04001556 if (ipv6_addr_loopback(daddr) ||
1557 ipv6_addr_is_multicast(&msg->target))
1558 goto out;
1559
1560 n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev);
Cong Wangf564f452013-08-31 13:44:36 +08001561
1562 if (n) {
1563 struct vxlan_fdb *f;
David Stevens4b29dba2014-03-24 10:39:58 -04001564 struct sk_buff *reply;
Cong Wangf564f452013-08-31 13:44:36 +08001565
1566 if (!(n->nud_state & NUD_CONNECTED)) {
1567 neigh_release(n);
1568 goto out;
1569 }
1570
1571 f = vxlan_find_mac(vxlan, n->ha);
1572 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
1573 /* bridge-local neighbor */
1574 neigh_release(n);
1575 goto out;
1576 }
1577
David Stevens4b29dba2014-03-24 10:39:58 -04001578 reply = vxlan_na_create(skb, n,
1579 !!(f ? f->flags & NTF_ROUTER : 0));
1580
Cong Wangf564f452013-08-31 13:44:36 +08001581 neigh_release(n);
David Stevens4b29dba2014-03-24 10:39:58 -04001582
1583 if (reply == NULL)
1584 goto out;
1585
1586 if (netif_rx_ni(reply) == NET_RX_DROP)
1587 dev->stats.rx_dropped++;
1588
Cong Wangf564f452013-08-31 13:44:36 +08001589 } else if (vxlan->flags & VXLAN_F_L3MISS) {
David Stevens4b29dba2014-03-24 10:39:58 -04001590 union vxlan_addr ipa = {
1591 .sin6.sin6_addr = msg->target,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001592 .sin6.sin6_family = AF_INET6,
David Stevens4b29dba2014-03-24 10:39:58 -04001593 };
1594
Cong Wangf564f452013-08-31 13:44:36 +08001595 vxlan_ip_miss(dev, &ipa);
1596 }
1597
1598out:
1599 consume_skb(skb);
1600 return NETDEV_TX_OK;
1601}
1602#endif
1603
David Stevense4f67ad2012-11-20 02:50:14 +00001604static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb)
1605{
1606 struct vxlan_dev *vxlan = netdev_priv(dev);
1607 struct neighbour *n;
David Stevense4f67ad2012-11-20 02:50:14 +00001608
1609 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
1610 return false;
1611
1612 n = NULL;
1613 switch (ntohs(eth_hdr(skb)->h_proto)) {
1614 case ETH_P_IP:
Cong Wange15a00a2013-08-31 13:44:34 +08001615 {
1616 struct iphdr *pip;
1617
David Stevense4f67ad2012-11-20 02:50:14 +00001618 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
1619 return false;
1620 pip = ip_hdr(skb);
1621 n = neigh_lookup(&arp_tbl, &pip->daddr, dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08001622 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1623 union vxlan_addr ipa = {
1624 .sin.sin_addr.s_addr = pip->daddr,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001625 .sin.sin_family = AF_INET,
Cong Wange4c7ed42013-08-31 13:44:33 +08001626 };
1627
1628 vxlan_ip_miss(dev, &ipa);
1629 return false;
1630 }
1631
David Stevense4f67ad2012-11-20 02:50:14 +00001632 break;
Cong Wange15a00a2013-08-31 13:44:34 +08001633 }
1634#if IS_ENABLED(CONFIG_IPV6)
1635 case ETH_P_IPV6:
1636 {
1637 struct ipv6hdr *pip6;
1638
1639 if (!pskb_may_pull(skb, sizeof(struct ipv6hdr)))
1640 return false;
1641 pip6 = ipv6_hdr(skb);
1642 n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev);
1643 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1644 union vxlan_addr ipa = {
1645 .sin6.sin6_addr = pip6->daddr,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001646 .sin6.sin6_family = AF_INET6,
Cong Wange15a00a2013-08-31 13:44:34 +08001647 };
1648
1649 vxlan_ip_miss(dev, &ipa);
1650 return false;
1651 }
1652
1653 break;
1654 }
1655#endif
David Stevense4f67ad2012-11-20 02:50:14 +00001656 default:
1657 return false;
1658 }
1659
1660 if (n) {
1661 bool diff;
1662
Joe Perches7367d0b2013-09-01 11:51:23 -07001663 diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha);
David Stevense4f67ad2012-11-20 02:50:14 +00001664 if (diff) {
1665 memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest,
1666 dev->addr_len);
1667 memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len);
1668 }
1669 neigh_release(n);
1670 return diff;
Cong Wange4c7ed42013-08-31 13:44:33 +08001671 }
1672
David Stevense4f67ad2012-11-20 02:50:14 +00001673 return false;
1674}
1675
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001676static void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, u32 vxflags,
Thomas Graf35114942015-01-15 03:53:55 +01001677 struct vxlan_metadata *md)
1678{
1679 struct vxlanhdr_gbp *gbp;
1680
Thomas Grafdb79a622015-02-04 17:00:04 +01001681 if (!md->gbp)
1682 return;
1683
Thomas Graf35114942015-01-15 03:53:55 +01001684 gbp = (struct vxlanhdr_gbp *)vxh;
Jiri Benc54bfd872016-02-16 21:58:58 +01001685 vxh->vx_flags |= VXLAN_HF_GBP;
Thomas Graf35114942015-01-15 03:53:55 +01001686
1687 if (md->gbp & VXLAN_GBP_DONT_LEARN)
1688 gbp->dont_learn = 1;
1689
1690 if (md->gbp & VXLAN_GBP_POLICY_APPLIED)
1691 gbp->policy_applied = 1;
1692
1693 gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK);
1694}
1695
Jiri Bence1e53142016-04-05 14:47:13 +02001696static int vxlan_build_gpe_hdr(struct vxlanhdr *vxh, u32 vxflags,
1697 __be16 protocol)
1698{
1699 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)vxh;
1700
1701 gpe->np_applied = 1;
1702
1703 switch (protocol) {
1704 case htons(ETH_P_IP):
1705 gpe->next_protocol = VXLAN_GPE_NP_IPV4;
1706 return 0;
1707 case htons(ETH_P_IPV6):
1708 gpe->next_protocol = VXLAN_GPE_NP_IPV6;
1709 return 0;
1710 case htons(ETH_P_TEB):
1711 gpe->next_protocol = VXLAN_GPE_NP_ETHERNET;
1712 return 0;
1713 }
1714 return -EPFNOSUPPORT;
1715}
1716
Jiri Bencf491e562016-02-02 18:09:16 +01001717static int vxlan_build_skb(struct sk_buff *skb, struct dst_entry *dst,
1718 int iphdr_len, __be32 vni,
1719 struct vxlan_metadata *md, u32 vxflags,
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01001720 bool udp_sum)
Cong Wange4c7ed42013-08-31 13:44:33 +08001721{
Cong Wange4c7ed42013-08-31 13:44:33 +08001722 struct vxlanhdr *vxh;
Cong Wange4c7ed42013-08-31 13:44:33 +08001723 int min_headroom;
1724 int err;
Tom Herbertdfd86452015-01-12 17:00:38 -08001725 int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL;
Jiri Bence1e53142016-04-05 14:47:13 +02001726 __be16 inner_protocol = htons(ETH_P_TEB);
Cong Wange4c7ed42013-08-31 13:44:33 +08001727
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001728 if ((vxflags & VXLAN_F_REMCSUM_TX) &&
Tom Herbertdfd86452015-01-12 17:00:38 -08001729 skb->ip_summed == CHECKSUM_PARTIAL) {
1730 int csum_start = skb_checksum_start_offset(skb);
1731
1732 if (csum_start <= VXLAN_MAX_REMCSUM_START &&
1733 !(csum_start & VXLAN_RCO_SHIFT_MASK) &&
1734 (skb->csum_offset == offsetof(struct udphdr, check) ||
Edward Creeb5708502016-02-11 20:57:17 +00001735 skb->csum_offset == offsetof(struct tcphdr, check)))
Tom Herbertdfd86452015-01-12 17:00:38 -08001736 type |= SKB_GSO_TUNNEL_REMCSUM;
Tom Herbertdfd86452015-01-12 17:00:38 -08001737 }
1738
Cong Wange4c7ed42013-08-31 13:44:33 +08001739 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
Jiri Bencf491e562016-02-02 18:09:16 +01001740 + VXLAN_HLEN + iphdr_len
Jiri Pirkodf8a39d2015-01-13 17:13:44 +01001741 + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001742
1743 /* Need space for new headers (invalidates iph ptr) */
1744 err = skb_cow_head(skb, min_headroom);
Jiri Bence1e53142016-04-05 14:47:13 +02001745 if (unlikely(err))
1746 goto out_free;
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001747
Jiri Pirko59682502014-11-19 14:04:59 +01001748 skb = vlan_hwaccel_push_inside(skb);
1749 if (WARN_ON(!skb))
1750 return -ENOMEM;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07001751
Alexander Duyckaed069d2016-04-14 15:33:37 -04001752 err = iptunnel_handle_offloads(skb, type);
1753 if (err)
1754 goto out_free;
Jesse Grossb736a622015-04-09 11:19:14 -07001755
Pravin B Shelar49560532013-08-19 11:23:17 -07001756 vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh));
Jiri Benc54bfd872016-02-16 21:58:58 +01001757 vxh->vx_flags = VXLAN_HF_VNI;
1758 vxh->vx_vni = vxlan_vni_field(vni);
Pravin B Shelar49560532013-08-19 11:23:17 -07001759
Tom Herbertdfd86452015-01-12 17:00:38 -08001760 if (type & SKB_GSO_TUNNEL_REMCSUM) {
Jiri Benc54bfd872016-02-16 21:58:58 +01001761 unsigned int start;
Tom Herbertdfd86452015-01-12 17:00:38 -08001762
Jiri Benc54bfd872016-02-16 21:58:58 +01001763 start = skb_checksum_start_offset(skb) - sizeof(struct vxlanhdr);
1764 vxh->vx_vni |= vxlan_compute_rco(start, skb->csum_offset);
1765 vxh->vx_flags |= VXLAN_HF_RCO;
Tom Herbertdfd86452015-01-12 17:00:38 -08001766
1767 if (!skb_is_gso(skb)) {
1768 skb->ip_summed = CHECKSUM_NONE;
1769 skb->encapsulation = 0;
1770 }
1771 }
1772
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001773 if (vxflags & VXLAN_F_GBP)
1774 vxlan_build_gbp_hdr(vxh, vxflags, md);
Jiri Bence1e53142016-04-05 14:47:13 +02001775 if (vxflags & VXLAN_F_GPE) {
1776 err = vxlan_build_gpe_hdr(vxh, vxflags, skb->protocol);
1777 if (err < 0)
1778 goto out_free;
1779 inner_protocol = skb->protocol;
1780 }
Thomas Graf35114942015-01-15 03:53:55 +01001781
Jiri Bence1e53142016-04-05 14:47:13 +02001782 skb_set_inner_protocol(skb, inner_protocol);
Pravin B Shelar039f5062015-12-24 14:34:54 -08001783 return 0;
Jiri Bence1e53142016-04-05 14:47:13 +02001784
1785out_free:
1786 kfree_skb(skb);
1787 return err;
Pravin B Shelar49560532013-08-19 11:23:17 -07001788}
Pravin B Shelar49560532013-08-19 11:23:17 -07001789
Jiri Benc1a8496b2016-02-02 18:09:14 +01001790static struct rtable *vxlan_get_route(struct vxlan_dev *vxlan,
1791 struct sk_buff *skb, int oif, u8 tos,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001792 __be32 daddr, __be32 *saddr,
1793 struct dst_cache *dst_cache,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001794 const struct ip_tunnel_info *info)
Jiri Benc1a8496b2016-02-02 18:09:14 +01001795{
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001796 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
Jiri Benc1a8496b2016-02-02 18:09:14 +01001797 struct rtable *rt = NULL;
1798 struct flowi4 fl4;
1799
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001800 if (tos && !info)
1801 use_cache = false;
1802 if (use_cache) {
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001803 rt = dst_cache_get_ip4(dst_cache, saddr);
1804 if (rt)
1805 return rt;
1806 }
1807
Jiri Benc1a8496b2016-02-02 18:09:14 +01001808 memset(&fl4, 0, sizeof(fl4));
1809 fl4.flowi4_oif = oif;
1810 fl4.flowi4_tos = RT_TOS(tos);
1811 fl4.flowi4_mark = skb->mark;
1812 fl4.flowi4_proto = IPPROTO_UDP;
1813 fl4.daddr = daddr;
1814 fl4.saddr = vxlan->cfg.saddr.sin.sin_addr.s_addr;
1815
1816 rt = ip_route_output_key(vxlan->net, &fl4);
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001817 if (!IS_ERR(rt)) {
Jiri Benc1a8496b2016-02-02 18:09:14 +01001818 *saddr = fl4.saddr;
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001819 if (use_cache)
1820 dst_cache_set_ip4(dst_cache, &rt->dst, fl4.saddr);
1821 }
Jiri Benc1a8496b2016-02-02 18:09:14 +01001822 return rt;
1823}
1824
Jiri Bence5d4b292015-12-07 13:04:30 +01001825#if IS_ENABLED(CONFIG_IPV6)
1826static struct dst_entry *vxlan6_get_route(struct vxlan_dev *vxlan,
Daniel Borkmann14006152016-03-04 15:15:08 +01001827 struct sk_buff *skb, int oif, u8 tos,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01001828 __be32 label,
Jiri Bence5d4b292015-12-07 13:04:30 +01001829 const struct in6_addr *daddr,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001830 struct in6_addr *saddr,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001831 struct dst_cache *dst_cache,
1832 const struct ip_tunnel_info *info)
Jiri Bence5d4b292015-12-07 13:04:30 +01001833{
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001834 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
Jiri Bence5d4b292015-12-07 13:04:30 +01001835 struct dst_entry *ndst;
1836 struct flowi6 fl6;
1837 int err;
1838
Daniel Borkmann14006152016-03-04 15:15:08 +01001839 if (tos && !info)
1840 use_cache = false;
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001841 if (use_cache) {
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001842 ndst = dst_cache_get_ip6(dst_cache, saddr);
1843 if (ndst)
1844 return ndst;
1845 }
1846
Jiri Bence5d4b292015-12-07 13:04:30 +01001847 memset(&fl6, 0, sizeof(fl6));
1848 fl6.flowi6_oif = oif;
1849 fl6.daddr = *daddr;
1850 fl6.saddr = vxlan->cfg.saddr.sin6.sin6_addr;
Daniel Borkmanneaa93bf2016-03-18 18:37:57 +01001851 fl6.flowlabel = ip6_make_flowinfo(RT_TOS(tos), label);
Jiri Bence5d4b292015-12-07 13:04:30 +01001852 fl6.flowi6_mark = skb->mark;
1853 fl6.flowi6_proto = IPPROTO_UDP;
1854
1855 err = ipv6_stub->ipv6_dst_lookup(vxlan->net,
1856 vxlan->vn6_sock->sock->sk,
1857 &ndst, &fl6);
1858 if (err < 0)
1859 return ERR_PTR(err);
1860
1861 *saddr = fl6.saddr;
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001862 if (use_cache)
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001863 dst_cache_set_ip6(dst_cache, ndst, saddr);
Jiri Bence5d4b292015-12-07 13:04:30 +01001864 return ndst;
1865}
1866#endif
1867
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001868/* Bypass encapsulation if the destination is local */
1869static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan,
1870 struct vxlan_dev *dst_vxlan)
1871{
Li RongQing8f849852014-01-04 13:57:59 +08001872 struct pcpu_sw_netstats *tx_stats, *rx_stats;
Cong Wange4c7ed42013-08-31 13:44:33 +08001873 union vxlan_addr loopback;
1874 union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip;
Li RongQingce6502a2014-10-16 08:49:41 +08001875 struct net_device *dev = skb->dev;
1876 int len = skb->len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001877
Li RongQing8f849852014-01-04 13:57:59 +08001878 tx_stats = this_cpu_ptr(src_vxlan->dev->tstats);
1879 rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001880 skb->pkt_type = PACKET_HOST;
1881 skb->encapsulation = 0;
1882 skb->dev = dst_vxlan->dev;
1883 __skb_pull(skb, skb_network_offset(skb));
1884
Cong Wange4c7ed42013-08-31 13:44:33 +08001885 if (remote_ip->sa.sa_family == AF_INET) {
1886 loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1887 loopback.sa.sa_family = AF_INET;
1888#if IS_ENABLED(CONFIG_IPV6)
1889 } else {
1890 loopback.sin6.sin6_addr = in6addr_loopback;
1891 loopback.sa.sa_family = AF_INET6;
1892#endif
1893 }
1894
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001895 if (dst_vxlan->flags & VXLAN_F_LEARN)
Cong Wange4c7ed42013-08-31 13:44:33 +08001896 vxlan_snoop(skb->dev, &loopback, eth_hdr(skb)->h_source);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001897
1898 u64_stats_update_begin(&tx_stats->syncp);
1899 tx_stats->tx_packets++;
Li RongQingce6502a2014-10-16 08:49:41 +08001900 tx_stats->tx_bytes += len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001901 u64_stats_update_end(&tx_stats->syncp);
1902
1903 if (netif_rx(skb) == NET_RX_SUCCESS) {
1904 u64_stats_update_begin(&rx_stats->syncp);
1905 rx_stats->rx_packets++;
Li RongQingce6502a2014-10-16 08:49:41 +08001906 rx_stats->rx_bytes += len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001907 u64_stats_update_end(&rx_stats->syncp);
1908 } else {
Li RongQingce6502a2014-10-16 08:49:41 +08001909 dev->stats.rx_dropped++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001910 }
1911}
1912
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001913static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev,
1914 struct vxlan_rdst *rdst, bool did_rsc)
stephen hemmingerd3428942012-10-01 12:32:35 +00001915{
Paolo Abenid71785f2016-02-12 15:43:57 +01001916 struct dst_cache *dst_cache;
Thomas Graf3093fbe2015-07-21 10:44:00 +02001917 struct ip_tunnel_info *info;
stephen hemmingerd3428942012-10-01 12:32:35 +00001918 struct vxlan_dev *vxlan = netdev_priv(dev);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001919 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001920 struct rtable *rt = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001921 const struct iphdr *old_iph;
Cong Wange4c7ed42013-08-31 13:44:33 +08001922 union vxlan_addr *dst;
Thomas Grafee122c72015-07-21 10:43:58 +02001923 union vxlan_addr remote_ip;
1924 struct vxlan_metadata _md;
1925 struct vxlan_metadata *md = &_md;
Cong Wange4c7ed42013-08-31 13:44:33 +08001926 __be16 src_port = 0, dst_port;
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01001927 __be32 vni, label;
stephen hemmingerd3428942012-10-01 12:32:35 +00001928 __be16 df = 0;
1929 __u8 tos, ttl;
Pravin B Shelar0e6fbc52013-06-17 17:49:56 -07001930 int err;
Thomas Grafee122c72015-07-21 10:43:58 +02001931 u32 flags = vxlan->flags;
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01001932 bool udp_sum = false;
Jiri Bencf491e562016-02-02 18:09:16 +01001933 bool xnet = !net_eq(vxlan->net, dev_net(vxlan->dev));
stephen hemmingerd3428942012-10-01 12:32:35 +00001934
Jiri Benc61adedf2015-08-20 13:56:25 +02001935 info = skb_tunnel_info(skb);
Thomas Graf3093fbe2015-07-21 10:44:00 +02001936
Thomas Grafee122c72015-07-21 10:43:58 +02001937 if (rdst) {
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001938 dst_port = rdst->remote_port ? rdst->remote_port : vxlan->cfg.dst_port;
Thomas Grafee122c72015-07-21 10:43:58 +02001939 vni = rdst->remote_vni;
1940 dst = &rdst->remote_ip;
Paolo Abenid71785f2016-02-12 15:43:57 +01001941 dst_cache = &rdst->dst_cache;
Thomas Grafee122c72015-07-21 10:43:58 +02001942 } else {
1943 if (!info) {
1944 WARN_ONCE(1, "%s: Missing encapsulation instructions\n",
1945 dev->name);
1946 goto drop;
1947 }
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001948 dst_port = info->key.tp_dst ? : vxlan->cfg.dst_port;
Jiri Benc54bfd872016-02-16 21:58:58 +01001949 vni = vxlan_tun_id_to_vni(info->key.tun_id);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001950 remote_ip.sa.sa_family = ip_tunnel_info_af(info);
1951 if (remote_ip.sa.sa_family == AF_INET)
Jiri Benca725e512015-08-20 13:56:30 +02001952 remote_ip.sin.sin_addr.s_addr = info->key.u.ipv4.dst;
1953 else
1954 remote_ip.sin6.sin6_addr = info->key.u.ipv6.dst;
Thomas Grafee122c72015-07-21 10:43:58 +02001955 dst = &remote_ip;
Paolo Abenid71785f2016-02-12 15:43:57 +01001956 dst_cache = &info->dst_cache;
Thomas Grafee122c72015-07-21 10:43:58 +02001957 }
David Stevense4f67ad2012-11-20 02:50:14 +00001958
Cong Wange4c7ed42013-08-31 13:44:33 +08001959 if (vxlan_addr_any(dst)) {
David Stevense4f67ad2012-11-20 02:50:14 +00001960 if (did_rsc) {
David Stevense4f67ad2012-11-20 02:50:14 +00001961 /* short-circuited back to local bridge */
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001962 vxlan_encap_bypass(skb, vxlan, vxlan);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001963 return;
David Stevense4f67ad2012-11-20 02:50:14 +00001964 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001965 goto drop;
David Stevense4f67ad2012-11-20 02:50:14 +00001966 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001967
stephen hemmingerd3428942012-10-01 12:32:35 +00001968 old_iph = ip_hdr(skb);
1969
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001970 ttl = vxlan->cfg.ttl;
Cong Wange4c7ed42013-08-31 13:44:33 +08001971 if (!ttl && vxlan_addr_multicast(dst))
stephen hemmingerd3428942012-10-01 12:32:35 +00001972 ttl = 1;
1973
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001974 tos = vxlan->cfg.tos;
stephen hemmingerd3428942012-10-01 12:32:35 +00001975 if (tos == 1)
Pravin B Shelar206aaaf2013-03-25 14:49:53 +00001976 tos = ip_tunnel_get_dsfield(old_iph, skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001977
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01001978 label = vxlan->cfg.label;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001979 src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
1980 vxlan->cfg.port_max, true);
stephen hemmingerd3428942012-10-01 12:32:35 +00001981
Jiri Benca725e512015-08-20 13:56:30 +02001982 if (info) {
Jiri Benca725e512015-08-20 13:56:30 +02001983 ttl = info->key.ttl;
1984 tos = info->key.tos;
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01001985 label = info->key.label;
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01001986 udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
Jiri Benca725e512015-08-20 13:56:30 +02001987
1988 if (info->options_len)
Pravin B Shelar4c222792015-08-30 18:09:38 -07001989 md = ip_tunnel_info_opts(info);
Jiri Benca725e512015-08-20 13:56:30 +02001990 } else {
1991 md->gbp = skb->mark;
1992 }
1993
Cong Wange4c7ed42013-08-31 13:44:33 +08001994 if (dst->sa.sa_family == AF_INET) {
Jiri Benc1a8496b2016-02-02 18:09:14 +01001995 __be32 saddr;
1996
Jiri Bencb1be00a2015-09-24 13:50:02 +02001997 if (!vxlan->vn4_sock)
1998 goto drop;
1999 sk = vxlan->vn4_sock->sock->sk;
2000
Jiri Benc1a8496b2016-02-02 18:09:14 +01002001 rt = vxlan_get_route(vxlan, skb,
2002 rdst ? rdst->remote_ifindex : 0, tos,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01002003 dst->sin.sin_addr.s_addr, &saddr,
Paolo Abenid71785f2016-02-12 15:43:57 +01002004 dst_cache, info);
Cong Wange4c7ed42013-08-31 13:44:33 +08002005 if (IS_ERR(rt)) {
2006 netdev_dbg(dev, "no route to %pI4\n",
2007 &dst->sin.sin_addr.s_addr);
2008 dev->stats.tx_carrier_errors++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00002009 goto tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08002010 }
2011
2012 if (rt->dst.dev == dev) {
2013 netdev_dbg(dev, "circular route to %pI4\n",
2014 &dst->sin.sin_addr.s_addr);
2015 dev->stats.collisions++;
Fan Dufffc15a2013-12-09 10:33:53 +08002016 goto rt_tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08002017 }
2018
2019 /* Bypass encapsulation if the destination is local */
2020 if (rt->rt_flags & RTCF_LOCAL &&
2021 !(rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
2022 struct vxlan_dev *dst_vxlan;
2023
2024 ip_rt_put(rt);
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02002025 dst_vxlan = vxlan_find_vni(vxlan->net, vni,
Thomas Grafac5132d2015-01-15 03:53:56 +01002026 dst->sa.sa_family, dst_port,
2027 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08002028 if (!dst_vxlan)
2029 goto tx_error;
2030 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
2031 return;
2032 }
2033
Alexander Duyck6ceb31c2016-02-19 11:26:31 -08002034 if (!info)
2035 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM_TX);
2036 else if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT)
2037 df = htons(IP_DF);
2038
Cong Wange4c7ed42013-08-31 13:44:33 +08002039 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
2040 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
Jiri Bencf491e562016-02-02 18:09:16 +01002041 err = vxlan_build_skb(skb, &rt->dst, sizeof(struct iphdr),
Jiri Benc54bfd872016-02-16 21:58:58 +01002042 vni, md, flags, udp_sum);
Jiri Bencf491e562016-02-02 18:09:16 +01002043 if (err < 0)
2044 goto xmit_tx_error;
2045
2046 udp_tunnel_xmit_skb(rt, sk, skb, saddr,
2047 dst->sin.sin_addr.s_addr, tos, ttl, df,
2048 src_port, dst_port, xnet, !udp_sum);
Cong Wange4c7ed42013-08-31 13:44:33 +08002049#if IS_ENABLED(CONFIG_IPV6)
2050 } else {
Cong Wange4c7ed42013-08-31 13:44:33 +08002051 struct dst_entry *ndst;
Jiri Bence5d4b292015-12-07 13:04:30 +01002052 struct in6_addr saddr;
Jiri Benc6f264e42015-08-20 13:56:29 +02002053 u32 rt6i_flags;
Cong Wange4c7ed42013-08-31 13:44:33 +08002054
Jiri Bencb1be00a2015-09-24 13:50:02 +02002055 if (!vxlan->vn6_sock)
2056 goto drop;
2057 sk = vxlan->vn6_sock->sock->sk;
2058
Jiri Bence5d4b292015-12-07 13:04:30 +01002059 ndst = vxlan6_get_route(vxlan, skb,
Daniel Borkmann14006152016-03-04 15:15:08 +01002060 rdst ? rdst->remote_ifindex : 0, tos,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002061 label, &dst->sin6.sin6_addr, &saddr,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01002062 dst_cache, info);
Jiri Bence5d4b292015-12-07 13:04:30 +01002063 if (IS_ERR(ndst)) {
Cong Wange4c7ed42013-08-31 13:44:33 +08002064 netdev_dbg(dev, "no route to %pI6\n",
2065 &dst->sin6.sin6_addr);
2066 dev->stats.tx_carrier_errors++;
2067 goto tx_error;
2068 }
2069
2070 if (ndst->dev == dev) {
2071 netdev_dbg(dev, "circular route to %pI6\n",
2072 &dst->sin6.sin6_addr);
2073 dst_release(ndst);
2074 dev->stats.collisions++;
2075 goto tx_error;
2076 }
2077
2078 /* Bypass encapsulation if the destination is local */
Jiri Benc6f264e42015-08-20 13:56:29 +02002079 rt6i_flags = ((struct rt6_info *)ndst)->rt6i_flags;
2080 if (rt6i_flags & RTF_LOCAL &&
2081 !(rt6i_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
Cong Wange4c7ed42013-08-31 13:44:33 +08002082 struct vxlan_dev *dst_vxlan;
2083
2084 dst_release(ndst);
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02002085 dst_vxlan = vxlan_find_vni(vxlan->net, vni,
Thomas Grafac5132d2015-01-15 03:53:56 +01002086 dst->sa.sa_family, dst_port,
2087 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08002088 if (!dst_vxlan)
2089 goto tx_error;
2090 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
2091 return;
2092 }
2093
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01002094 if (!info)
2095 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM6_TX);
Jesse Gross35e2d112016-01-20 16:22:47 -08002096
Daniel Borkmann14006152016-03-04 15:15:08 +01002097 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
Cong Wange4c7ed42013-08-31 13:44:33 +08002098 ttl = ttl ? : ip6_dst_hoplimit(ndst);
Jiri Bencf491e562016-02-02 18:09:16 +01002099 skb_scrub_packet(skb, xnet);
2100 err = vxlan_build_skb(skb, ndst, sizeof(struct ipv6hdr),
Jiri Benc54bfd872016-02-16 21:58:58 +01002101 vni, md, flags, udp_sum);
Jiri Bencf491e562016-02-02 18:09:16 +01002102 if (err < 0) {
2103 dst_release(ndst);
2104 return;
2105 }
2106 udp_tunnel6_xmit_skb(ndst, sk, skb, dev,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002107 &saddr, &dst->sin6.sin6_addr, tos, ttl,
2108 label, src_port, dst_port, !udp_sum);
Cong Wange4c7ed42013-08-31 13:44:33 +08002109#endif
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00002110 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002111
Stephen Hemminger4ad16932013-06-17 14:16:11 -07002112 return;
stephen hemmingerd3428942012-10-01 12:32:35 +00002113
2114drop:
2115 dev->stats.tx_dropped++;
2116 goto tx_free;
2117
Jiri Bencf491e562016-02-02 18:09:16 +01002118xmit_tx_error:
2119 /* skb is already freed. */
2120 skb = NULL;
Pravin B Shelar49560532013-08-19 11:23:17 -07002121rt_tx_error:
2122 ip_rt_put(rt);
stephen hemmingerd3428942012-10-01 12:32:35 +00002123tx_error:
2124 dev->stats.tx_errors++;
2125tx_free:
2126 dev_kfree_skb(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00002127}
2128
David Stevens66817122013-03-15 04:35:51 +00002129/* Transmit local packets over Vxlan
2130 *
2131 * Outer IP header inherits ECN and DF from inner header.
2132 * Outer UDP destination is the VXLAN assigned port.
2133 * source port is based on hash of flow
2134 */
2135static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev)
2136{
2137 struct vxlan_dev *vxlan = netdev_priv(dev);
Thomas Graf3093fbe2015-07-21 10:44:00 +02002138 const struct ip_tunnel_info *info;
David Stevens66817122013-03-15 04:35:51 +00002139 struct ethhdr *eth;
2140 bool did_rsc = false;
Eric Dumazet8f646c92014-01-06 09:54:31 -08002141 struct vxlan_rdst *rdst, *fdst = NULL;
David Stevens66817122013-03-15 04:35:51 +00002142 struct vxlan_fdb *f;
David Stevens66817122013-03-15 04:35:51 +00002143
Jiri Benc61adedf2015-08-20 13:56:25 +02002144 info = skb_tunnel_info(skb);
Thomas Graf3093fbe2015-07-21 10:44:00 +02002145
David Stevens66817122013-03-15 04:35:51 +00002146 skb_reset_mac_header(skb);
David Stevens66817122013-03-15 04:35:51 +00002147
Jiri Benc47e5d1b2016-04-05 14:47:11 +02002148 if (vxlan->flags & VXLAN_F_COLLECT_METADATA) {
2149 if (info && info->mode & IP_TUNNEL_INFO_TX)
2150 vxlan_xmit_one(skb, dev, NULL, false);
2151 else
2152 kfree_skb(skb);
2153 return NETDEV_TX_OK;
2154 }
2155
2156 if (vxlan->flags & VXLAN_F_PROXY) {
2157 eth = eth_hdr(skb);
Cong Wangf564f452013-08-31 13:44:36 +08002158 if (ntohs(eth->h_proto) == ETH_P_ARP)
2159 return arp_reduce(dev, skb);
2160#if IS_ENABLED(CONFIG_IPV6)
2161 else if (ntohs(eth->h_proto) == ETH_P_IPV6 &&
Li RongQing91269e32014-10-16 09:17:18 +08002162 pskb_may_pull(skb, sizeof(struct ipv6hdr)
2163 + sizeof(struct nd_msg)) &&
Cong Wangf564f452013-08-31 13:44:36 +08002164 ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) {
2165 struct nd_msg *msg;
2166
2167 msg = (struct nd_msg *)skb_transport_header(skb);
2168 if (msg->icmph.icmp6_code == 0 &&
2169 msg->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION)
2170 return neigh_reduce(dev, skb);
2171 }
2172#endif
2173 }
David Stevens66817122013-03-15 04:35:51 +00002174
Jiri Benc47e5d1b2016-04-05 14:47:11 +02002175 eth = eth_hdr(skb);
David Stevens66817122013-03-15 04:35:51 +00002176 f = vxlan_find_mac(vxlan, eth->h_dest);
David Stevensae884082013-04-19 00:36:26 +00002177 did_rsc = false;
2178
2179 if (f && (f->flags & NTF_ROUTER) && (vxlan->flags & VXLAN_F_RSC) &&
Cong Wange15a00a2013-08-31 13:44:34 +08002180 (ntohs(eth->h_proto) == ETH_P_IP ||
2181 ntohs(eth->h_proto) == ETH_P_IPV6)) {
David Stevensae884082013-04-19 00:36:26 +00002182 did_rsc = route_shortcircuit(dev, skb);
2183 if (did_rsc)
2184 f = vxlan_find_mac(vxlan, eth->h_dest);
2185 }
2186
David Stevens66817122013-03-15 04:35:51 +00002187 if (f == NULL) {
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002188 f = vxlan_find_mac(vxlan, all_zeros_mac);
2189 if (f == NULL) {
2190 if ((vxlan->flags & VXLAN_F_L2MISS) &&
2191 !is_multicast_ether_addr(eth->h_dest))
2192 vxlan_fdb_miss(vxlan, eth->h_dest);
David Stevens66817122013-03-15 04:35:51 +00002193
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002194 dev->stats.tx_dropped++;
Eric Dumazet8f646c92014-01-06 09:54:31 -08002195 kfree_skb(skb);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002196 return NETDEV_TX_OK;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -07002197 }
David Stevens66817122013-03-15 04:35:51 +00002198 }
2199
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002200 list_for_each_entry_rcu(rdst, &f->remotes, list) {
2201 struct sk_buff *skb1;
2202
Eric Dumazet8f646c92014-01-06 09:54:31 -08002203 if (!fdst) {
2204 fdst = rdst;
2205 continue;
2206 }
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002207 skb1 = skb_clone(skb, GFP_ATOMIC);
2208 if (skb1)
2209 vxlan_xmit_one(skb1, dev, rdst, did_rsc);
2210 }
2211
Eric Dumazet8f646c92014-01-06 09:54:31 -08002212 if (fdst)
2213 vxlan_xmit_one(skb, dev, fdst, did_rsc);
2214 else
2215 kfree_skb(skb);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07002216 return NETDEV_TX_OK;
David Stevens66817122013-03-15 04:35:51 +00002217}
2218
stephen hemmingerd3428942012-10-01 12:32:35 +00002219/* Walk the forwarding table and purge stale entries */
2220static void vxlan_cleanup(unsigned long arg)
2221{
2222 struct vxlan_dev *vxlan = (struct vxlan_dev *) arg;
2223 unsigned long next_timer = jiffies + FDB_AGE_INTERVAL;
2224 unsigned int h;
2225
2226 if (!netif_running(vxlan->dev))
2227 return;
2228
stephen hemmingerd3428942012-10-01 12:32:35 +00002229 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2230 struct hlist_node *p, *n;
Sorin Dumitru14e1d0f2015-05-26 10:42:04 +03002231
2232 spin_lock_bh(&vxlan->hash_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002233 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2234 struct vxlan_fdb *f
2235 = container_of(p, struct vxlan_fdb, hlist);
2236 unsigned long timeout;
2237
stephen hemminger3c172862012-10-26 06:24:34 +00002238 if (f->state & NUD_PERMANENT)
stephen hemmingerd3428942012-10-01 12:32:35 +00002239 continue;
2240
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002241 timeout = f->used + vxlan->cfg.age_interval * HZ;
stephen hemmingerd3428942012-10-01 12:32:35 +00002242 if (time_before_eq(timeout, jiffies)) {
2243 netdev_dbg(vxlan->dev,
2244 "garbage collect %pM\n",
2245 f->eth_addr);
2246 f->state = NUD_STALE;
2247 vxlan_fdb_destroy(vxlan, f);
2248 } else if (time_before(timeout, next_timer))
2249 next_timer = timeout;
2250 }
Sorin Dumitru14e1d0f2015-05-26 10:42:04 +03002251 spin_unlock_bh(&vxlan->hash_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002252 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002253
2254 mod_timer(&vxlan->age_timer, next_timer);
2255}
2256
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002257static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan)
2258{
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002259 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
Jiri Benc54bfd872016-02-16 21:58:58 +01002260 __be32 vni = vxlan->default_dst.remote_vni;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002261
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002262 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002263 hlist_add_head_rcu(&vxlan->hlist, vni_head(vs, vni));
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002264 spin_unlock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002265}
2266
stephen hemmingerd3428942012-10-01 12:32:35 +00002267/* Setup stats when device is created */
2268static int vxlan_init(struct net_device *dev)
2269{
WANG Cong1c213bd2014-02-13 11:46:28 -08002270 dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
Pravin B Shelare8171042013-03-25 14:49:46 +00002271 if (!dev->tstats)
stephen hemmingerd3428942012-10-01 12:32:35 +00002272 return -ENOMEM;
2273
2274 return 0;
2275}
2276
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002277static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan)
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002278{
2279 struct vxlan_fdb *f;
2280
2281 spin_lock_bh(&vxlan->hash_lock);
2282 f = __vxlan_find_mac(vxlan, all_zeros_mac);
2283 if (f)
2284 vxlan_fdb_destroy(vxlan, f);
2285 spin_unlock_bh(&vxlan->hash_lock);
2286}
2287
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002288static void vxlan_uninit(struct net_device *dev)
2289{
2290 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002291
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002292 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002293
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002294 free_percpu(dev->tstats);
2295}
2296
stephen hemmingerd3428942012-10-01 12:32:35 +00002297/* Start ageing timer and join group when device is brought up */
2298static int vxlan_open(struct net_device *dev)
2299{
2300 struct vxlan_dev *vxlan = netdev_priv(dev);
Jiri Benc205f3562015-09-24 13:50:01 +02002301 int ret;
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002302
Jiri Benc205f3562015-09-24 13:50:01 +02002303 ret = vxlan_sock_add(vxlan);
2304 if (ret < 0)
2305 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002306
Gao feng79d4a942013-12-10 16:37:32 +08002307 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002308 ret = vxlan_igmp_join(vxlan);
Marcelo Ricardo Leitnerbef00572015-08-25 20:22:35 -03002309 if (ret == -EADDRINUSE)
2310 ret = 0;
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002311 if (ret) {
Jiri Benc205f3562015-09-24 13:50:01 +02002312 vxlan_sock_release(vxlan);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002313 return ret;
2314 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002315 }
2316
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002317 if (vxlan->cfg.age_interval)
stephen hemmingerd3428942012-10-01 12:32:35 +00002318 mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL);
2319
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002320 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002321}
2322
2323/* Purge the forwarding table */
2324static void vxlan_flush(struct vxlan_dev *vxlan)
2325{
Cong Wang31fec5a2013-05-27 22:35:52 +00002326 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002327
2328 spin_lock_bh(&vxlan->hash_lock);
2329 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2330 struct hlist_node *p, *n;
2331 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2332 struct vxlan_fdb *f
2333 = container_of(p, struct vxlan_fdb, hlist);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002334 /* the all_zeros_mac entry is deleted at vxlan_uninit */
2335 if (!is_zero_ether_addr(f->eth_addr))
2336 vxlan_fdb_destroy(vxlan, f);
stephen hemmingerd3428942012-10-01 12:32:35 +00002337 }
2338 }
2339 spin_unlock_bh(&vxlan->hash_lock);
2340}
2341
2342/* Cleanup timer and forwarding table on shutdown */
2343static int vxlan_stop(struct net_device *dev)
2344{
2345 struct vxlan_dev *vxlan = netdev_priv(dev);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002346 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002347 int ret = 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002348
Marcelo Ricardo Leitner24c0e682015-03-23 16:23:12 -03002349 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip) &&
WANG Congf13b1682015-04-08 14:48:30 -07002350 !vxlan_group_used(vn, vxlan))
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002351 ret = vxlan_igmp_leave(vxlan);
stephen hemmingerd3428942012-10-01 12:32:35 +00002352
2353 del_timer_sync(&vxlan->age_timer);
2354
2355 vxlan_flush(vxlan);
Jiri Benc205f3562015-09-24 13:50:01 +02002356 vxlan_sock_release(vxlan);
stephen hemmingerd3428942012-10-01 12:32:35 +00002357
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002358 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002359}
2360
stephen hemmingerd3428942012-10-01 12:32:35 +00002361/* Stub, nothing needs to be done. */
2362static void vxlan_set_multicast_list(struct net_device *dev)
2363{
2364}
2365
David Wragg72564b52016-02-10 00:05:55 +00002366static int __vxlan_change_mtu(struct net_device *dev,
2367 struct net_device *lowerdev,
2368 struct vxlan_rdst *dst, int new_mtu, bool strict)
2369{
2370 int max_mtu = IP_MAX_MTU;
2371
2372 if (lowerdev)
2373 max_mtu = lowerdev->mtu;
2374
2375 if (dst->remote_ip.sa.sa_family == AF_INET6)
2376 max_mtu -= VXLAN6_HEADROOM;
2377 else
2378 max_mtu -= VXLAN_HEADROOM;
2379
2380 if (new_mtu < 68)
2381 return -EINVAL;
2382
2383 if (new_mtu > max_mtu) {
2384 if (strict)
2385 return -EINVAL;
2386
2387 new_mtu = max_mtu;
2388 }
2389
2390 dev->mtu = new_mtu;
2391 return 0;
2392}
2393
Daniel Borkmann345010b2013-12-18 00:21:08 +01002394static int vxlan_change_mtu(struct net_device *dev, int new_mtu)
2395{
2396 struct vxlan_dev *vxlan = netdev_priv(dev);
2397 struct vxlan_rdst *dst = &vxlan->default_dst;
David Wragg72564b52016-02-10 00:05:55 +00002398 struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
2399 dst->remote_ifindex);
2400 return __vxlan_change_mtu(dev, lowerdev, dst, new_mtu, true);
Daniel Borkmann345010b2013-12-18 00:21:08 +01002401}
2402
Pravin B Shelarfc4099f2015-10-22 18:17:16 -07002403static int vxlan_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
2404{
2405 struct vxlan_dev *vxlan = netdev_priv(dev);
2406 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2407 __be16 sport, dport;
2408
2409 sport = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
2410 vxlan->cfg.port_max, true);
2411 dport = info->key.tp_dst ? : vxlan->cfg.dst_port;
2412
Jiri Benc239e9442015-12-07 13:04:31 +01002413 if (ip_tunnel_info_af(info) == AF_INET) {
Jiri Benc1a8496b2016-02-02 18:09:14 +01002414 struct rtable *rt;
2415
Jiri Benc239e9442015-12-07 13:04:31 +01002416 if (!vxlan->vn4_sock)
2417 return -EINVAL;
Jiri Benc1a8496b2016-02-02 18:09:14 +01002418 rt = vxlan_get_route(vxlan, skb, 0, info->key.tos,
2419 info->key.u.ipv4.dst,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01002420 &info->key.u.ipv4.src, NULL, info);
Jiri Benc1a8496b2016-02-02 18:09:14 +01002421 if (IS_ERR(rt))
2422 return PTR_ERR(rt);
2423 ip_rt_put(rt);
Jiri Benc239e9442015-12-07 13:04:31 +01002424 } else {
2425#if IS_ENABLED(CONFIG_IPV6)
2426 struct dst_entry *ndst;
2427
2428 if (!vxlan->vn6_sock)
2429 return -EINVAL;
Daniel Borkmann14006152016-03-04 15:15:08 +01002430 ndst = vxlan6_get_route(vxlan, skb, 0, info->key.tos,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002431 info->key.label, &info->key.u.ipv6.dst,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01002432 &info->key.u.ipv6.src, NULL, info);
Jiri Benc239e9442015-12-07 13:04:31 +01002433 if (IS_ERR(ndst))
2434 return PTR_ERR(ndst);
2435 dst_release(ndst);
Jiri Benc239e9442015-12-07 13:04:31 +01002436#else /* !CONFIG_IPV6 */
2437 return -EPFNOSUPPORT;
2438#endif
2439 }
Jiri Benc1a8496b2016-02-02 18:09:14 +01002440 info->key.tp_src = sport;
2441 info->key.tp_dst = dport;
Jiri Benc239e9442015-12-07 13:04:31 +01002442 return 0;
Pravin B Shelarfc4099f2015-10-22 18:17:16 -07002443}
2444
Jiri Benc0c867c92016-04-05 14:47:10 +02002445static const struct net_device_ops vxlan_netdev_ether_ops = {
stephen hemmingerd3428942012-10-01 12:32:35 +00002446 .ndo_init = vxlan_init,
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002447 .ndo_uninit = vxlan_uninit,
stephen hemmingerd3428942012-10-01 12:32:35 +00002448 .ndo_open = vxlan_open,
2449 .ndo_stop = vxlan_stop,
2450 .ndo_start_xmit = vxlan_xmit,
Pravin B Shelare8171042013-03-25 14:49:46 +00002451 .ndo_get_stats64 = ip_tunnel_get_stats64,
stephen hemmingerd3428942012-10-01 12:32:35 +00002452 .ndo_set_rx_mode = vxlan_set_multicast_list,
Daniel Borkmann345010b2013-12-18 00:21:08 +01002453 .ndo_change_mtu = vxlan_change_mtu,
stephen hemmingerd3428942012-10-01 12:32:35 +00002454 .ndo_validate_addr = eth_validate_addr,
2455 .ndo_set_mac_address = eth_mac_addr,
2456 .ndo_fdb_add = vxlan_fdb_add,
2457 .ndo_fdb_del = vxlan_fdb_delete,
2458 .ndo_fdb_dump = vxlan_fdb_dump,
Pravin B Shelarfc4099f2015-10-22 18:17:16 -07002459 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
stephen hemmingerd3428942012-10-01 12:32:35 +00002460};
2461
Jiri Bence1e53142016-04-05 14:47:13 +02002462static const struct net_device_ops vxlan_netdev_raw_ops = {
2463 .ndo_init = vxlan_init,
2464 .ndo_uninit = vxlan_uninit,
2465 .ndo_open = vxlan_open,
2466 .ndo_stop = vxlan_stop,
2467 .ndo_start_xmit = vxlan_xmit,
2468 .ndo_get_stats64 = ip_tunnel_get_stats64,
2469 .ndo_change_mtu = vxlan_change_mtu,
2470 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
2471};
2472
stephen hemmingerd3428942012-10-01 12:32:35 +00002473/* Info for udev, that this is a virtual tunnel endpoint */
2474static struct device_type vxlan_type = {
2475 .name = "vxlan",
2476};
2477
Sabrina Dubrocae5de25d2016-07-11 13:12:28 +02002478/* Calls the ndo_udp_tunnel_add of the caller in order to
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002479 * supply the listening VXLAN udp ports. Callers are expected
Sabrina Dubrocae5de25d2016-07-11 13:12:28 +02002480 * to implement the ndo_udp_tunnel_add.
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002481 */
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02002482static void vxlan_push_rx_ports(struct net_device *dev)
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002483{
2484 struct vxlan_sock *vs;
2485 struct net *net = dev_net(dev);
2486 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002487 unsigned int i;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002488
2489 spin_lock(&vn->sock_lock);
2490 for (i = 0; i < PORT_HASH_SIZE; ++i) {
Alexander Duycke7b3db52016-06-16 12:20:52 -07002491 hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist)
2492 udp_tunnel_push_rx_port(dev, vs->sock,
Alexander Duyckb9adcd692016-06-16 12:23:19 -07002493 (vs->flags & VXLAN_F_GPE) ?
2494 UDP_TUNNEL_TYPE_VXLAN_GPE :
Alexander Duycke7b3db52016-06-16 12:20:52 -07002495 UDP_TUNNEL_TYPE_VXLAN);
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002496 }
2497 spin_unlock(&vn->sock_lock);
2498}
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002499
stephen hemmingerd3428942012-10-01 12:32:35 +00002500/* Initialize the device structure. */
2501static void vxlan_setup(struct net_device *dev)
2502{
2503 struct vxlan_dev *vxlan = netdev_priv(dev);
Cong Wang31fec5a2013-05-27 22:35:52 +00002504 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002505
Jiri Benc65226ef2016-04-28 16:36:30 +02002506 eth_hw_addr_random(dev);
2507 ether_setup(dev);
2508
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002509 dev->destructor = free_netdev;
stephen hemmingerd3428942012-10-01 12:32:35 +00002510 SET_NETDEV_DEVTYPE(dev, &vxlan_type);
2511
stephen hemmingerd3428942012-10-01 12:32:35 +00002512 dev->features |= NETIF_F_LLTX;
Joseph Gasparakisd6727fe2012-12-07 14:14:16 +00002513 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002514 dev->features |= NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002515 dev->features |= NETIF_F_GSO_SOFTWARE;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002516
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002517 dev->vlan_features = dev->features;
2518 dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002519 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002520 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002521 dev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Eric Dumazet02875872014-10-05 18:38:35 -07002522 netif_keep_dst(dev);
Jiri Benc0c867c92016-04-05 14:47:10 +02002523 dev->priv_flags |= IFF_NO_QUEUE;
stephen hemmingerd3428942012-10-01 12:32:35 +00002524
stephen hemminger553675f2013-05-16 11:35:20 +00002525 INIT_LIST_HEAD(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00002526 spin_lock_init(&vxlan->hash_lock);
2527
2528 init_timer_deferrable(&vxlan->age_timer);
2529 vxlan->age_timer.function = vxlan_cleanup;
2530 vxlan->age_timer.data = (unsigned long) vxlan;
2531
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002532 vxlan->cfg.dst_port = htons(vxlan_port);
stephen hemminger05f47d62012-10-09 20:35:50 +00002533
stephen hemmingerd3428942012-10-01 12:32:35 +00002534 vxlan->dev = dev;
2535
Tom Herbert58ce31c2015-08-19 17:07:33 -07002536 gro_cells_init(&vxlan->gro_cells, dev);
2537
stephen hemmingerd3428942012-10-01 12:32:35 +00002538 for (h = 0; h < FDB_HASH_SIZE; ++h)
2539 INIT_HLIST_HEAD(&vxlan->fdb_head[h]);
2540}
2541
Jiri Benc0c867c92016-04-05 14:47:10 +02002542static void vxlan_ether_setup(struct net_device *dev)
2543{
Jiri Benc0c867c92016-04-05 14:47:10 +02002544 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
2545 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
2546 dev->netdev_ops = &vxlan_netdev_ether_ops;
2547}
2548
Jiri Bence1e53142016-04-05 14:47:13 +02002549static void vxlan_raw_setup(struct net_device *dev)
2550{
Jiri Benc65226ef2016-04-28 16:36:30 +02002551 dev->header_ops = NULL;
Jiri Bence1e53142016-04-05 14:47:13 +02002552 dev->type = ARPHRD_NONE;
2553 dev->hard_header_len = 0;
2554 dev->addr_len = 0;
Jiri Bence1e53142016-04-05 14:47:13 +02002555 dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
2556 dev->netdev_ops = &vxlan_netdev_raw_ops;
2557}
2558
stephen hemmingerd3428942012-10-01 12:32:35 +00002559static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = {
2560 [IFLA_VXLAN_ID] = { .type = NLA_U32 },
stephen hemminger5d174dd2013-04-27 11:31:55 +00002561 [IFLA_VXLAN_GROUP] = { .len = FIELD_SIZEOF(struct iphdr, daddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002562 [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002563 [IFLA_VXLAN_LINK] = { .type = NLA_U32 },
2564 [IFLA_VXLAN_LOCAL] = { .len = FIELD_SIZEOF(struct iphdr, saddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002565 [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002566 [IFLA_VXLAN_TOS] = { .type = NLA_U8 },
2567 [IFLA_VXLAN_TTL] = { .type = NLA_U8 },
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002568 [IFLA_VXLAN_LABEL] = { .type = NLA_U32 },
stephen hemmingerd3428942012-10-01 12:32:35 +00002569 [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 },
2570 [IFLA_VXLAN_AGEING] = { .type = NLA_U32 },
2571 [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 },
stephen hemminger05f47d62012-10-09 20:35:50 +00002572 [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) },
David Stevense4f67ad2012-11-20 02:50:14 +00002573 [IFLA_VXLAN_PROXY] = { .type = NLA_U8 },
2574 [IFLA_VXLAN_RSC] = { .type = NLA_U8 },
2575 [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 },
2576 [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 },
Alexei Starovoitovf8a9b1b2015-07-30 20:10:22 -07002577 [IFLA_VXLAN_COLLECT_METADATA] = { .type = NLA_U8 },
stephen hemminger823aa872013-04-27 11:31:57 +00002578 [IFLA_VXLAN_PORT] = { .type = NLA_U16 },
Tom Herbert5c91ae02014-11-06 18:06:01 -08002579 [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 },
2580 [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 },
2581 [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 },
Tom Herbertdfd86452015-01-12 17:00:38 -08002582 [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 },
2583 [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 },
Thomas Graf35114942015-01-15 03:53:55 +01002584 [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, },
Jiri Bence1e53142016-04-05 14:47:13 +02002585 [IFLA_VXLAN_GPE] = { .type = NLA_FLAG, },
Tom Herbert0ace2ca2015-02-10 16:30:32 -08002586 [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG },
stephen hemmingerd3428942012-10-01 12:32:35 +00002587};
2588
2589static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[])
2590{
2591 if (tb[IFLA_ADDRESS]) {
2592 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) {
2593 pr_debug("invalid link address (not ethernet)\n");
2594 return -EINVAL;
2595 }
2596
2597 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) {
2598 pr_debug("invalid all zero ethernet address\n");
2599 return -EADDRNOTAVAIL;
2600 }
2601 }
2602
2603 if (!data)
2604 return -EINVAL;
2605
2606 if (data[IFLA_VXLAN_ID]) {
2607 __u32 id = nla_get_u32(data[IFLA_VXLAN_ID]);
2608 if (id >= VXLAN_VID_MASK)
2609 return -ERANGE;
2610 }
2611
stephen hemminger05f47d62012-10-09 20:35:50 +00002612 if (data[IFLA_VXLAN_PORT_RANGE]) {
2613 const struct ifla_vxlan_port_range *p
2614 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2615
2616 if (ntohs(p->high) < ntohs(p->low)) {
2617 pr_debug("port range %u .. %u not valid\n",
2618 ntohs(p->low), ntohs(p->high));
2619 return -EINVAL;
2620 }
2621 }
2622
stephen hemmingerd3428942012-10-01 12:32:35 +00002623 return 0;
2624}
2625
Yan Burman1b13c972013-01-29 23:43:07 +00002626static void vxlan_get_drvinfo(struct net_device *netdev,
2627 struct ethtool_drvinfo *drvinfo)
2628{
2629 strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version));
2630 strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver));
2631}
2632
2633static const struct ethtool_ops vxlan_ethtool_ops = {
2634 .get_drvinfo = vxlan_get_drvinfo,
2635 .get_link = ethtool_op_get_link,
2636};
2637
Tom Herbert3ee64f32014-07-13 19:49:42 -07002638static struct socket *vxlan_create_sock(struct net *net, bool ipv6,
2639 __be16 port, u32 flags)
stephen hemminger553675f2013-05-16 11:35:20 +00002640{
Cong Wange4c7ed42013-08-31 13:44:33 +08002641 struct socket *sock;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002642 struct udp_port_cfg udp_conf;
2643 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002644
Tom Herbert3ee64f32014-07-13 19:49:42 -07002645 memset(&udp_conf, 0, sizeof(udp_conf));
2646
2647 if (ipv6) {
2648 udp_conf.family = AF_INET6;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002649 udp_conf.use_udp6_rx_checksums =
Alexander Duyck3dc2b6a2014-11-24 20:08:38 -08002650 !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX);
Jiri Benca43a9ef2015-08-28 20:48:22 +02002651 udp_conf.ipv6_v6only = 1;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002652 } else {
2653 udp_conf.family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +08002654 }
2655
Tom Herbert3ee64f32014-07-13 19:49:42 -07002656 udp_conf.local_udp_port = port;
Cong Wange4c7ed42013-08-31 13:44:33 +08002657
Tom Herbert3ee64f32014-07-13 19:49:42 -07002658 /* Open UDP socket */
2659 err = udp_sock_create(net, &udp_conf, &sock);
2660 if (err < 0)
2661 return ERR_PTR(err);
Cong Wange4c7ed42013-08-31 13:44:33 +08002662
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002663 return sock;
Cong Wange4c7ed42013-08-31 13:44:33 +08002664}
2665
2666/* Create new listen socket if needed */
Jiri Bencb1be00a2015-09-24 13:50:02 +02002667static struct vxlan_sock *vxlan_socket_create(struct net *net, bool ipv6,
2668 __be16 port, u32 flags)
Cong Wange4c7ed42013-08-31 13:44:33 +08002669{
2670 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2671 struct vxlan_sock *vs;
2672 struct socket *sock;
Cong Wang31fec5a2013-05-27 22:35:52 +00002673 unsigned int h;
Andy Zhouacbf74a2014-09-16 17:31:18 -07002674 struct udp_tunnel_sock_cfg tunnel_cfg;
stephen hemminger553675f2013-05-16 11:35:20 +00002675
Or Gerlitzdc01e7d2014-01-20 13:59:21 +02002676 vs = kzalloc(sizeof(*vs), GFP_KERNEL);
Cong Wange4c7ed42013-08-31 13:44:33 +08002677 if (!vs)
stephen hemminger553675f2013-05-16 11:35:20 +00002678 return ERR_PTR(-ENOMEM);
2679
2680 for (h = 0; h < VNI_HASH_SIZE; ++h)
2681 INIT_HLIST_HEAD(&vs->vni_list[h]);
2682
Tom Herbert3ee64f32014-07-13 19:49:42 -07002683 sock = vxlan_create_sock(net, ipv6, port, flags);
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002684 if (IS_ERR(sock)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002685 pr_info("Cannot bind port %d, err=%ld\n", ntohs(port),
2686 PTR_ERR(sock));
stephen hemminger553675f2013-05-16 11:35:20 +00002687 kfree(vs);
Duan Jionge50fddc2013-11-01 13:09:43 +08002688 return ERR_CAST(sock);
stephen hemminger553675f2013-05-16 11:35:20 +00002689 }
2690
Cong Wange4c7ed42013-08-31 13:44:33 +08002691 vs->sock = sock;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002692 atomic_set(&vs->refcnt, 1);
Tom Herbertaf33c1a2015-01-20 11:23:05 -08002693 vs->flags = (flags & VXLAN_F_RCV_FLAGS);
stephen hemminger553675f2013-05-16 11:35:20 +00002694
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002695 spin_lock(&vn->sock_lock);
2696 hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
Alexander Duycke7b3db52016-06-16 12:20:52 -07002697 udp_tunnel_notify_add_rx_port(sock,
Alexander Duyckb9adcd692016-06-16 12:23:19 -07002698 (vs->flags & VXLAN_F_GPE) ?
2699 UDP_TUNNEL_TYPE_VXLAN_GPE :
Alexander Duycke7b3db52016-06-16 12:20:52 -07002700 UDP_TUNNEL_TYPE_VXLAN);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002701 spin_unlock(&vn->sock_lock);
stephen hemminger553675f2013-05-16 11:35:20 +00002702
2703 /* Mark socket as an encapsulation socket. */
Tom Herbert5602c482016-04-05 08:22:53 -07002704 memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
Andy Zhouacbf74a2014-09-16 17:31:18 -07002705 tunnel_cfg.sk_user_data = vs;
2706 tunnel_cfg.encap_type = 1;
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01002707 tunnel_cfg.encap_rcv = vxlan_rcv;
Andy Zhouacbf74a2014-09-16 17:31:18 -07002708 tunnel_cfg.encap_destroy = NULL;
Tom Herbert5602c482016-04-05 08:22:53 -07002709 tunnel_cfg.gro_receive = vxlan_gro_receive;
2710 tunnel_cfg.gro_complete = vxlan_gro_complete;
Andy Zhouacbf74a2014-09-16 17:31:18 -07002711
2712 setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
Cong Wange4c7ed42013-08-31 13:44:33 +08002713
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002714 return vs;
2715}
stephen hemminger553675f2013-05-16 11:35:20 +00002716
Jiri Bencb1be00a2015-09-24 13:50:02 +02002717static int __vxlan_sock_add(struct vxlan_dev *vxlan, bool ipv6)
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002718{
Jiri Benc205f3562015-09-24 13:50:01 +02002719 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
2720 struct vxlan_sock *vs = NULL;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002721
Jiri Benc205f3562015-09-24 13:50:01 +02002722 if (!vxlan->cfg.no_share) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002723 spin_lock(&vn->sock_lock);
Jiri Benc205f3562015-09-24 13:50:01 +02002724 vs = vxlan_find_sock(vxlan->net, ipv6 ? AF_INET6 : AF_INET,
2725 vxlan->cfg.dst_port, vxlan->flags);
2726 if (vs && !atomic_add_unless(&vs->refcnt, 1, 0)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002727 spin_unlock(&vn->sock_lock);
Jiri Benc205f3562015-09-24 13:50:01 +02002728 return -EBUSY;
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002729 }
2730 spin_unlock(&vn->sock_lock);
2731 }
Jiri Benc205f3562015-09-24 13:50:01 +02002732 if (!vs)
Jiri Bencb1be00a2015-09-24 13:50:02 +02002733 vs = vxlan_socket_create(vxlan->net, ipv6,
2734 vxlan->cfg.dst_port, vxlan->flags);
Jiri Benc205f3562015-09-24 13:50:01 +02002735 if (IS_ERR(vs))
2736 return PTR_ERR(vs);
Jiri Bencb1be00a2015-09-24 13:50:02 +02002737#if IS_ENABLED(CONFIG_IPV6)
2738 if (ipv6)
2739 vxlan->vn6_sock = vs;
2740 else
2741#endif
2742 vxlan->vn4_sock = vs;
Jiri Benc205f3562015-09-24 13:50:01 +02002743 vxlan_vs_add_dev(vs, vxlan);
2744 return 0;
stephen hemminger553675f2013-05-16 11:35:20 +00002745}
2746
Jiri Bencb1be00a2015-09-24 13:50:02 +02002747static int vxlan_sock_add(struct vxlan_dev *vxlan)
2748{
2749 bool ipv6 = vxlan->flags & VXLAN_F_IPV6;
2750 bool metadata = vxlan->flags & VXLAN_F_COLLECT_METADATA;
2751 int ret = 0;
2752
2753 vxlan->vn4_sock = NULL;
2754#if IS_ENABLED(CONFIG_IPV6)
2755 vxlan->vn6_sock = NULL;
2756 if (ipv6 || metadata)
2757 ret = __vxlan_sock_add(vxlan, true);
2758#endif
2759 if (!ret && (!ipv6 || metadata))
2760 ret = __vxlan_sock_add(vxlan, false);
2761 if (ret < 0)
2762 vxlan_sock_release(vxlan);
2763 return ret;
2764}
2765
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002766static int vxlan_dev_configure(struct net *src_net, struct net_device *dev,
2767 struct vxlan_config *conf)
stephen hemmingerd3428942012-10-01 12:32:35 +00002768{
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002769 struct vxlan_net *vn = net_generic(src_net, vxlan_net_id);
Nicolas Dichtel07b9b372016-01-07 11:26:53 +01002770 struct vxlan_dev *vxlan = netdev_priv(dev), *tmp;
Atzm Watanabec7995c42013-04-16 02:50:52 +00002771 struct vxlan_rdst *dst = &vxlan->default_dst;
Jiri Bencb1be00a2015-09-24 13:50:02 +02002772 unsigned short needed_headroom = ETH_HLEN;
stephen hemmingerd3428942012-10-01 12:32:35 +00002773 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002774 bool use_ipv6 = false;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002775 __be16 default_port = vxlan->cfg.dst_port;
David Wragg7e059152016-02-10 00:05:58 +00002776 struct net_device *lowerdev = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +00002777
Jiri Bence1e53142016-04-05 14:47:13 +02002778 if (conf->flags & VXLAN_F_GPE) {
2779 if (conf->flags & ~VXLAN_F_ALLOWED_GPE)
2780 return -EINVAL;
2781 /* For now, allow GPE only together with COLLECT_METADATA.
2782 * This can be relaxed later; in such case, the other side
2783 * of the PtP link will have to be provided.
2784 */
2785 if (!(conf->flags & VXLAN_F_COLLECT_METADATA))
2786 return -EINVAL;
2787
2788 vxlan_raw_setup(dev);
2789 } else {
2790 vxlan_ether_setup(dev);
2791 }
Jiri Benc0c867c92016-04-05 14:47:10 +02002792
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002793 vxlan->net = src_net;
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002794
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002795 dst->remote_vni = conf->vni;
2796
2797 memcpy(&dst->remote_ip, &conf->remote_ip, sizeof(conf->remote_ip));
stephen hemmingerd3428942012-10-01 12:32:35 +00002798
Mike Rapoport5933a7b2014-04-01 09:23:01 +03002799 /* Unless IPv6 is explicitly requested, assume IPv4 */
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002800 if (!dst->remote_ip.sa.sa_family)
2801 dst->remote_ip.sa.sa_family = AF_INET;
stephen hemmingerd3428942012-10-01 12:32:35 +00002802
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002803 if (dst->remote_ip.sa.sa_family == AF_INET6 ||
Jiri Benc057ba292015-09-17 16:11:11 +02002804 vxlan->cfg.saddr.sa.sa_family == AF_INET6) {
2805 if (!IS_ENABLED(CONFIG_IPV6))
2806 return -EPFNOSUPPORT;
Cong Wange4c7ed42013-08-31 13:44:33 +08002807 use_ipv6 = true;
Jiri Bencb1be00a2015-09-24 13:50:02 +02002808 vxlan->flags |= VXLAN_F_IPV6;
Jiri Benc057ba292015-09-17 16:11:11 +02002809 }
Cong Wange4c7ed42013-08-31 13:44:33 +08002810
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002811 if (conf->label && !use_ipv6) {
2812 pr_info("label only supported in use with IPv6\n");
2813 return -EINVAL;
2814 }
2815
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002816 if (conf->remote_ifindex) {
David Wragg7e059152016-02-10 00:05:58 +00002817 lowerdev = __dev_get_by_index(src_net, conf->remote_ifindex);
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002818 dst->remote_ifindex = conf->remote_ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +00002819
stephen hemminger34e02aa2012-10-09 20:35:53 +00002820 if (!lowerdev) {
Atzm Watanabec7995c42013-04-16 02:50:52 +00002821 pr_info("ifindex %d does not exist\n", dst->remote_ifindex);
stephen hemminger34e02aa2012-10-09 20:35:53 +00002822 return -ENODEV;
stephen hemmingerd3428942012-10-01 12:32:35 +00002823 }
stephen hemminger34e02aa2012-10-09 20:35:53 +00002824
Cong Wange4c7ed42013-08-31 13:44:33 +08002825#if IS_ENABLED(CONFIG_IPV6)
2826 if (use_ipv6) {
2827 struct inet6_dev *idev = __in6_dev_get(lowerdev);
2828 if (idev && idev->cnf.disable_ipv6) {
2829 pr_info("IPv6 is disabled via sysctl\n");
2830 return -EPERM;
2831 }
Cong Wange4c7ed42013-08-31 13:44:33 +08002832 }
2833#endif
2834
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002835 if (!conf->mtu)
Cong Wange4c7ed42013-08-31 13:44:33 +08002836 dev->mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
Alexander Duyck1ba56fb2012-11-13 13:10:59 +00002837
Jiri Bencb1be00a2015-09-24 13:50:02 +02002838 needed_headroom = lowerdev->hard_header_len;
Jiri Benc9dc2ad12015-09-17 16:11:10 +02002839 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002840
David Wragg7e059152016-02-10 00:05:58 +00002841 if (conf->mtu) {
2842 err = __vxlan_change_mtu(dev, lowerdev, dst, conf->mtu, false);
2843 if (err)
2844 return err;
2845 }
2846
Jiri Bencb1be00a2015-09-24 13:50:02 +02002847 if (use_ipv6 || conf->flags & VXLAN_F_COLLECT_METADATA)
2848 needed_headroom += VXLAN6_HEADROOM;
2849 else
2850 needed_headroom += VXLAN_HEADROOM;
2851 dev->needed_headroom = needed_headroom;
2852
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002853 memcpy(&vxlan->cfg, conf, sizeof(*conf));
Jiri Bence1e53142016-04-05 14:47:13 +02002854 if (!vxlan->cfg.dst_port) {
2855 if (conf->flags & VXLAN_F_GPE)
2856 vxlan->cfg.dst_port = 4790; /* IANA assigned VXLAN-GPE port */
2857 else
2858 vxlan->cfg.dst_port = default_port;
2859 }
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002860 vxlan->flags |= conf->flags;
stephen hemmingerd3428942012-10-01 12:32:35 +00002861
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002862 if (!vxlan->cfg.age_interval)
2863 vxlan->cfg.age_interval = FDB_AGE_DEFAULT;
Vincent Bernatafb97182012-10-30 10:27:16 +00002864
Nicolas Dichtel07b9b372016-01-07 11:26:53 +01002865 list_for_each_entry(tmp, &vn->vxlan_list, next) {
2866 if (tmp->cfg.vni == conf->vni &&
2867 (tmp->default_dst.remote_ip.sa.sa_family == AF_INET6 ||
2868 tmp->cfg.saddr.sa.sa_family == AF_INET6) == use_ipv6 &&
2869 tmp->cfg.dst_port == vxlan->cfg.dst_port &&
2870 (tmp->flags & VXLAN_F_RCV_FLAGS) ==
2871 (vxlan->flags & VXLAN_F_RCV_FLAGS))
stephen hemminger553675f2013-05-16 11:35:20 +00002872 return -EEXIST;
Nicolas Dichtel07b9b372016-01-07 11:26:53 +01002873 }
stephen hemminger553675f2013-05-16 11:35:20 +00002874
Wilfried Klaebe7ad24ea2014-05-11 00:12:32 +00002875 dev->ethtool_ops = &vxlan_ethtool_ops;
Yan Burman1b13c972013-01-29 23:43:07 +00002876
Sridhar Samudrala2936b6a2013-09-17 12:12:40 -07002877 /* create an fdb entry for a valid default destination */
2878 if (!vxlan_addr_any(&vxlan->default_dst.remote_ip)) {
2879 err = vxlan_fdb_create(vxlan, all_zeros_mac,
2880 &vxlan->default_dst.remote_ip,
2881 NUD_REACHABLE|NUD_PERMANENT,
2882 NLM_F_EXCL|NLM_F_CREATE,
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002883 vxlan->cfg.dst_port,
Sridhar Samudrala2936b6a2013-09-17 12:12:40 -07002884 vxlan->default_dst.remote_vni,
2885 vxlan->default_dst.remote_ifindex,
2886 NTF_SELF);
2887 if (err)
2888 return err;
2889 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002890
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002891 err = register_netdevice(dev);
2892 if (err) {
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002893 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002894 return err;
2895 }
2896
stephen hemminger553675f2013-05-16 11:35:20 +00002897 list_add(&vxlan->next, &vn->vxlan_list);
stephen hemminger553675f2013-05-16 11:35:20 +00002898
2899 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002900}
2901
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002902static int vxlan_newlink(struct net *src_net, struct net_device *dev,
2903 struct nlattr *tb[], struct nlattr *data[])
2904{
2905 struct vxlan_config conf;
2906 int err;
2907
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002908 memset(&conf, 0, sizeof(conf));
Jesse Grosse277de52015-10-16 16:36:00 -07002909
2910 if (data[IFLA_VXLAN_ID])
Jiri Benc54bfd872016-02-16 21:58:58 +01002911 conf.vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID]));
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002912
2913 if (data[IFLA_VXLAN_GROUP]) {
2914 conf.remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]);
2915 } else if (data[IFLA_VXLAN_GROUP6]) {
2916 if (!IS_ENABLED(CONFIG_IPV6))
2917 return -EPFNOSUPPORT;
2918
2919 conf.remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]);
2920 conf.remote_ip.sa.sa_family = AF_INET6;
2921 }
2922
2923 if (data[IFLA_VXLAN_LOCAL]) {
2924 conf.saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]);
2925 conf.saddr.sa.sa_family = AF_INET;
2926 } else if (data[IFLA_VXLAN_LOCAL6]) {
2927 if (!IS_ENABLED(CONFIG_IPV6))
2928 return -EPFNOSUPPORT;
2929
2930 /* TODO: respect scope id */
2931 conf.saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]);
2932 conf.saddr.sa.sa_family = AF_INET6;
2933 }
2934
2935 if (data[IFLA_VXLAN_LINK])
2936 conf.remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]);
2937
2938 if (data[IFLA_VXLAN_TOS])
2939 conf.tos = nla_get_u8(data[IFLA_VXLAN_TOS]);
2940
2941 if (data[IFLA_VXLAN_TTL])
2942 conf.ttl = nla_get_u8(data[IFLA_VXLAN_TTL]);
2943
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002944 if (data[IFLA_VXLAN_LABEL])
2945 conf.label = nla_get_be32(data[IFLA_VXLAN_LABEL]) &
2946 IPV6_FLOWLABEL_MASK;
2947
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002948 if (!data[IFLA_VXLAN_LEARNING] || nla_get_u8(data[IFLA_VXLAN_LEARNING]))
2949 conf.flags |= VXLAN_F_LEARN;
2950
2951 if (data[IFLA_VXLAN_AGEING])
2952 conf.age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]);
2953
2954 if (data[IFLA_VXLAN_PROXY] && nla_get_u8(data[IFLA_VXLAN_PROXY]))
2955 conf.flags |= VXLAN_F_PROXY;
2956
2957 if (data[IFLA_VXLAN_RSC] && nla_get_u8(data[IFLA_VXLAN_RSC]))
2958 conf.flags |= VXLAN_F_RSC;
2959
2960 if (data[IFLA_VXLAN_L2MISS] && nla_get_u8(data[IFLA_VXLAN_L2MISS]))
2961 conf.flags |= VXLAN_F_L2MISS;
2962
2963 if (data[IFLA_VXLAN_L3MISS] && nla_get_u8(data[IFLA_VXLAN_L3MISS]))
2964 conf.flags |= VXLAN_F_L3MISS;
2965
2966 if (data[IFLA_VXLAN_LIMIT])
2967 conf.addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]);
2968
Alexei Starovoitovf8a9b1b2015-07-30 20:10:22 -07002969 if (data[IFLA_VXLAN_COLLECT_METADATA] &&
2970 nla_get_u8(data[IFLA_VXLAN_COLLECT_METADATA]))
2971 conf.flags |= VXLAN_F_COLLECT_METADATA;
2972
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002973 if (data[IFLA_VXLAN_PORT_RANGE]) {
2974 const struct ifla_vxlan_port_range *p
2975 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2976 conf.port_min = ntohs(p->low);
2977 conf.port_max = ntohs(p->high);
2978 }
2979
2980 if (data[IFLA_VXLAN_PORT])
2981 conf.dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]);
2982
Alexander Duyck6ceb31c2016-02-19 11:26:31 -08002983 if (data[IFLA_VXLAN_UDP_CSUM] &&
2984 !nla_get_u8(data[IFLA_VXLAN_UDP_CSUM]))
2985 conf.flags |= VXLAN_F_UDP_ZERO_CSUM_TX;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002986
2987 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX] &&
2988 nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX]))
2989 conf.flags |= VXLAN_F_UDP_ZERO_CSUM6_TX;
2990
2991 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX] &&
2992 nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX]))
2993 conf.flags |= VXLAN_F_UDP_ZERO_CSUM6_RX;
2994
2995 if (data[IFLA_VXLAN_REMCSUM_TX] &&
2996 nla_get_u8(data[IFLA_VXLAN_REMCSUM_TX]))
2997 conf.flags |= VXLAN_F_REMCSUM_TX;
2998
2999 if (data[IFLA_VXLAN_REMCSUM_RX] &&
3000 nla_get_u8(data[IFLA_VXLAN_REMCSUM_RX]))
3001 conf.flags |= VXLAN_F_REMCSUM_RX;
3002
3003 if (data[IFLA_VXLAN_GBP])
3004 conf.flags |= VXLAN_F_GBP;
3005
Jiri Bence1e53142016-04-05 14:47:13 +02003006 if (data[IFLA_VXLAN_GPE])
3007 conf.flags |= VXLAN_F_GPE;
3008
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003009 if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL])
3010 conf.flags |= VXLAN_F_REMCSUM_NOPARTIAL;
3011
Chen Haiquance577662016-05-27 10:49:11 +08003012 if (tb[IFLA_MTU])
3013 conf.mtu = nla_get_u32(tb[IFLA_MTU]);
3014
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003015 err = vxlan_dev_configure(src_net, dev, &conf);
3016 switch (err) {
3017 case -ENODEV:
3018 pr_info("ifindex %d does not exist\n", conf.remote_ifindex);
3019 break;
3020
3021 case -EPERM:
3022 pr_info("IPv6 is disabled via sysctl\n");
3023 break;
3024
3025 case -EEXIST:
Jiri Benc54bfd872016-02-16 21:58:58 +01003026 pr_info("duplicate VNI %u\n", be32_to_cpu(conf.vni));
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003027 break;
Jiri Bence1e53142016-04-05 14:47:13 +02003028
3029 case -EINVAL:
3030 pr_info("unsupported combination of extensions\n");
3031 break;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003032 }
3033
3034 return err;
3035}
3036
stephen hemmingerd3428942012-10-01 12:32:35 +00003037static void vxlan_dellink(struct net_device *dev, struct list_head *head)
3038{
3039 struct vxlan_dev *vxlan = netdev_priv(dev);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003040 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00003041
stephen hemmingerfe5c3562013-07-13 10:18:18 -07003042 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07003043 if (!hlist_unhashed(&vxlan->hlist))
3044 hlist_del_rcu(&vxlan->hlist);
stephen hemmingerfe5c3562013-07-13 10:18:18 -07003045 spin_unlock(&vn->sock_lock);
3046
Tom Herbert58ce31c2015-08-19 17:07:33 -07003047 gro_cells_destroy(&vxlan->gro_cells);
stephen hemminger553675f2013-05-16 11:35:20 +00003048 list_del(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00003049 unregister_netdevice_queue(dev, head);
3050}
3051
3052static size_t vxlan_get_size(const struct net_device *dev)
3053{
3054
3055 return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */
Cong Wange4c7ed42013-08-31 13:44:33 +08003056 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00003057 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */
Cong Wange4c7ed42013-08-31 13:44:33 +08003058 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00003059 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */
3060 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01003061 nla_total_size(sizeof(__be32)) + /* IFLA_VXLAN_LABEL */
stephen hemmingerd3428942012-10-01 12:32:35 +00003062 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */
David Stevense4f67ad2012-11-20 02:50:14 +00003063 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */
3064 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */
3065 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */
3066 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */
Alexei Starovoitovda8b43c2015-08-04 22:51:07 -07003067 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_COLLECT_METADATA */
stephen hemmingerd3428942012-10-01 12:32:35 +00003068 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */
3069 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */
stephen hemminger05f47d62012-10-09 20:35:50 +00003070 nla_total_size(sizeof(struct ifla_vxlan_port_range)) +
Tom Herbert359a0ea2014-06-04 17:20:29 -07003071 nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */
3072 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */
3073 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */
3074 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */
Tom Herbertdfd86452015-01-12 17:00:38 -08003075 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */
3076 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */
stephen hemmingerd3428942012-10-01 12:32:35 +00003077 0;
3078}
3079
3080static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev)
3081{
3082 const struct vxlan_dev *vxlan = netdev_priv(dev);
Atzm Watanabec7995c42013-04-16 02:50:52 +00003083 const struct vxlan_rdst *dst = &vxlan->default_dst;
stephen hemminger05f47d62012-10-09 20:35:50 +00003084 struct ifla_vxlan_port_range ports = {
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003085 .low = htons(vxlan->cfg.port_min),
3086 .high = htons(vxlan->cfg.port_max),
stephen hemminger05f47d62012-10-09 20:35:50 +00003087 };
stephen hemmingerd3428942012-10-01 12:32:35 +00003088
Jiri Benc54bfd872016-02-16 21:58:58 +01003089 if (nla_put_u32(skb, IFLA_VXLAN_ID, be32_to_cpu(dst->remote_vni)))
stephen hemmingerd3428942012-10-01 12:32:35 +00003090 goto nla_put_failure;
3091
Cong Wange4c7ed42013-08-31 13:44:33 +08003092 if (!vxlan_addr_any(&dst->remote_ip)) {
3093 if (dst->remote_ip.sa.sa_family == AF_INET) {
Jiri Benc930345e2015-03-29 16:59:25 +02003094 if (nla_put_in_addr(skb, IFLA_VXLAN_GROUP,
3095 dst->remote_ip.sin.sin_addr.s_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003096 goto nla_put_failure;
3097#if IS_ENABLED(CONFIG_IPV6)
3098 } else {
Jiri Benc930345e2015-03-29 16:59:25 +02003099 if (nla_put_in6_addr(skb, IFLA_VXLAN_GROUP6,
3100 &dst->remote_ip.sin6.sin6_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003101 goto nla_put_failure;
3102#endif
3103 }
3104 }
stephen hemmingerd3428942012-10-01 12:32:35 +00003105
Atzm Watanabec7995c42013-04-16 02:50:52 +00003106 if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +00003107 goto nla_put_failure;
3108
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003109 if (!vxlan_addr_any(&vxlan->cfg.saddr)) {
3110 if (vxlan->cfg.saddr.sa.sa_family == AF_INET) {
Jiri Benc930345e2015-03-29 16:59:25 +02003111 if (nla_put_in_addr(skb, IFLA_VXLAN_LOCAL,
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003112 vxlan->cfg.saddr.sin.sin_addr.s_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003113 goto nla_put_failure;
3114#if IS_ENABLED(CONFIG_IPV6)
3115 } else {
Jiri Benc930345e2015-03-29 16:59:25 +02003116 if (nla_put_in6_addr(skb, IFLA_VXLAN_LOCAL6,
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003117 &vxlan->cfg.saddr.sin6.sin6_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003118 goto nla_put_failure;
3119#endif
3120 }
3121 }
stephen hemmingerd3428942012-10-01 12:32:35 +00003122
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003123 if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->cfg.ttl) ||
3124 nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->cfg.tos) ||
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01003125 nla_put_be32(skb, IFLA_VXLAN_LABEL, vxlan->cfg.label) ||
David Stevense4f67ad2012-11-20 02:50:14 +00003126 nla_put_u8(skb, IFLA_VXLAN_LEARNING,
3127 !!(vxlan->flags & VXLAN_F_LEARN)) ||
3128 nla_put_u8(skb, IFLA_VXLAN_PROXY,
3129 !!(vxlan->flags & VXLAN_F_PROXY)) ||
3130 nla_put_u8(skb, IFLA_VXLAN_RSC, !!(vxlan->flags & VXLAN_F_RSC)) ||
3131 nla_put_u8(skb, IFLA_VXLAN_L2MISS,
3132 !!(vxlan->flags & VXLAN_F_L2MISS)) ||
3133 nla_put_u8(skb, IFLA_VXLAN_L3MISS,
3134 !!(vxlan->flags & VXLAN_F_L3MISS)) ||
Alexei Starovoitovda8b43c2015-08-04 22:51:07 -07003135 nla_put_u8(skb, IFLA_VXLAN_COLLECT_METADATA,
3136 !!(vxlan->flags & VXLAN_F_COLLECT_METADATA)) ||
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003137 nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->cfg.age_interval) ||
3138 nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->cfg.addrmax) ||
3139 nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->cfg.dst_port) ||
Tom Herbert359a0ea2014-06-04 17:20:29 -07003140 nla_put_u8(skb, IFLA_VXLAN_UDP_CSUM,
Alexander Duyck6ceb31c2016-02-19 11:26:31 -08003141 !(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM_TX)) ||
Tom Herbert359a0ea2014-06-04 17:20:29 -07003142 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_TX,
3143 !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_TX)) ||
3144 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_RX,
Tom Herbertdfd86452015-01-12 17:00:38 -08003145 !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_RX)) ||
3146 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_TX,
3147 !!(vxlan->flags & VXLAN_F_REMCSUM_TX)) ||
3148 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_RX,
3149 !!(vxlan->flags & VXLAN_F_REMCSUM_RX)))
stephen hemmingerd3428942012-10-01 12:32:35 +00003150 goto nla_put_failure;
3151
stephen hemminger05f47d62012-10-09 20:35:50 +00003152 if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports))
3153 goto nla_put_failure;
3154
Thomas Graf35114942015-01-15 03:53:55 +01003155 if (vxlan->flags & VXLAN_F_GBP &&
3156 nla_put_flag(skb, IFLA_VXLAN_GBP))
3157 goto nla_put_failure;
3158
Jiri Bence1e53142016-04-05 14:47:13 +02003159 if (vxlan->flags & VXLAN_F_GPE &&
3160 nla_put_flag(skb, IFLA_VXLAN_GPE))
3161 goto nla_put_failure;
3162
Tom Herbert0ace2ca2015-02-10 16:30:32 -08003163 if (vxlan->flags & VXLAN_F_REMCSUM_NOPARTIAL &&
3164 nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL))
3165 goto nla_put_failure;
3166
stephen hemmingerd3428942012-10-01 12:32:35 +00003167 return 0;
3168
3169nla_put_failure:
3170 return -EMSGSIZE;
3171}
3172
Nicolas Dichtel1728d4f2015-01-15 15:11:17 +01003173static struct net *vxlan_get_link_net(const struct net_device *dev)
3174{
3175 struct vxlan_dev *vxlan = netdev_priv(dev);
3176
3177 return vxlan->net;
3178}
3179
stephen hemmingerd3428942012-10-01 12:32:35 +00003180static struct rtnl_link_ops vxlan_link_ops __read_mostly = {
3181 .kind = "vxlan",
3182 .maxtype = IFLA_VXLAN_MAX,
3183 .policy = vxlan_policy,
3184 .priv_size = sizeof(struct vxlan_dev),
3185 .setup = vxlan_setup,
3186 .validate = vxlan_validate,
3187 .newlink = vxlan_newlink,
3188 .dellink = vxlan_dellink,
3189 .get_size = vxlan_get_size,
3190 .fill_info = vxlan_fill_info,
Nicolas Dichtel1728d4f2015-01-15 15:11:17 +01003191 .get_link_net = vxlan_get_link_net,
stephen hemmingerd3428942012-10-01 12:32:35 +00003192};
3193
Nicolas Dichtelcf5da332016-06-13 10:31:05 +02003194struct net_device *vxlan_dev_create(struct net *net, const char *name,
3195 u8 name_assign_type,
3196 struct vxlan_config *conf)
3197{
3198 struct nlattr *tb[IFLA_MAX + 1];
3199 struct net_device *dev;
3200 int err;
3201
3202 memset(&tb, 0, sizeof(tb));
3203
3204 dev = rtnl_create_link(net, name, name_assign_type,
3205 &vxlan_link_ops, tb);
3206 if (IS_ERR(dev))
3207 return dev;
3208
3209 err = vxlan_dev_configure(net, dev, conf);
3210 if (err < 0) {
3211 free_netdev(dev);
3212 return ERR_PTR(err);
3213 }
3214
3215 err = rtnl_configure_link(dev, NULL);
3216 if (err < 0) {
3217 LIST_HEAD(list_kill);
3218
3219 vxlan_dellink(dev, &list_kill);
3220 unregister_netdevice_many(&list_kill);
3221 return ERR_PTR(err);
3222 }
3223
3224 return dev;
3225}
3226EXPORT_SYMBOL_GPL(vxlan_dev_create);
3227
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003228static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn,
3229 struct net_device *dev)
3230{
3231 struct vxlan_dev *vxlan, *next;
3232 LIST_HEAD(list_kill);
3233
3234 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
3235 struct vxlan_rdst *dst = &vxlan->default_dst;
3236
3237 /* In case we created vxlan device with carrier
3238 * and we loose the carrier due to module unload
3239 * we also need to remove vxlan device. In other
3240 * cases, it's not necessary and remote_ifindex
3241 * is 0 here, so no matches.
3242 */
3243 if (dst->remote_ifindex == dev->ifindex)
3244 vxlan_dellink(vxlan->dev, &list_kill);
3245 }
3246
3247 unregister_netdevice_many(&list_kill);
3248}
3249
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02003250static int vxlan_netdevice_event(struct notifier_block *unused,
3251 unsigned long event, void *ptr)
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003252{
3253 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
Daniel Borkmann783c1462014-01-22 21:07:53 +01003254 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003255
Daniel Borkmann783c1462014-01-22 21:07:53 +01003256 if (event == NETDEV_UNREGISTER)
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003257 vxlan_handle_lowerdev_unregister(vn, dev);
Alexander Duyck7c46a642016-06-16 12:21:00 -07003258 else if (event == NETDEV_UDP_TUNNEL_PUSH_INFO)
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02003259 vxlan_push_rx_ports(dev);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003260
3261 return NOTIFY_DONE;
3262}
3263
3264static struct notifier_block vxlan_notifier_block __read_mostly = {
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02003265 .notifier_call = vxlan_netdevice_event,
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003266};
3267
stephen hemmingerd3428942012-10-01 12:32:35 +00003268static __net_init int vxlan_init_net(struct net *net)
3269{
3270 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Cong Wang31fec5a2013-05-27 22:35:52 +00003271 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00003272
stephen hemminger553675f2013-05-16 11:35:20 +00003273 INIT_LIST_HEAD(&vn->vxlan_list);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07003274 spin_lock_init(&vn->sock_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00003275
stephen hemminger553675f2013-05-16 11:35:20 +00003276 for (h = 0; h < PORT_HASH_SIZE; ++h)
3277 INIT_HLIST_HEAD(&vn->sock_list[h]);
stephen hemmingerd3428942012-10-01 12:32:35 +00003278
3279 return 0;
3280}
3281
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003282static void __net_exit vxlan_exit_net(struct net *net)
3283{
3284 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
3285 struct vxlan_dev *vxlan, *next;
3286 struct net_device *dev, *aux;
3287 LIST_HEAD(list);
3288
3289 rtnl_lock();
3290 for_each_netdev_safe(net, dev, aux)
3291 if (dev->rtnl_link_ops == &vxlan_link_ops)
3292 unregister_netdevice_queue(dev, &list);
3293
3294 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
3295 /* If vxlan->dev is in the same netns, it has already been added
3296 * to the list by the previous loop.
3297 */
Tom Herbert58ce31c2015-08-19 17:07:33 -07003298 if (!net_eq(dev_net(vxlan->dev), net)) {
3299 gro_cells_destroy(&vxlan->gro_cells);
John W. Linville13c3ed62015-05-18 13:51:24 -04003300 unregister_netdevice_queue(vxlan->dev, &list);
Tom Herbert58ce31c2015-08-19 17:07:33 -07003301 }
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003302 }
3303
3304 unregister_netdevice_many(&list);
3305 rtnl_unlock();
3306}
3307
stephen hemmingerd3428942012-10-01 12:32:35 +00003308static struct pernet_operations vxlan_net_ops = {
3309 .init = vxlan_init_net,
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003310 .exit = vxlan_exit_net,
stephen hemmingerd3428942012-10-01 12:32:35 +00003311 .id = &vxlan_net_id,
3312 .size = sizeof(struct vxlan_net),
3313};
3314
3315static int __init vxlan_init_module(void)
3316{
3317 int rc;
3318
3319 get_random_bytes(&vxlan_salt, sizeof(vxlan_salt));
3320
Daniel Borkmann783c1462014-01-22 21:07:53 +01003321 rc = register_pernet_subsys(&vxlan_net_ops);
stephen hemmingerd3428942012-10-01 12:32:35 +00003322 if (rc)
3323 goto out1;
3324
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003325 rc = register_netdevice_notifier(&vxlan_notifier_block);
stephen hemmingerd3428942012-10-01 12:32:35 +00003326 if (rc)
3327 goto out2;
3328
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003329 rc = rtnl_link_register(&vxlan_link_ops);
3330 if (rc)
3331 goto out3;
stephen hemmingerd3428942012-10-01 12:32:35 +00003332
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003333 return 0;
3334out3:
3335 unregister_netdevice_notifier(&vxlan_notifier_block);
stephen hemmingerd3428942012-10-01 12:32:35 +00003336out2:
Daniel Borkmann783c1462014-01-22 21:07:53 +01003337 unregister_pernet_subsys(&vxlan_net_ops);
stephen hemmingerd3428942012-10-01 12:32:35 +00003338out1:
3339 return rc;
3340}
Cong Wang7332a132013-05-27 22:35:53 +00003341late_initcall(vxlan_init_module);
stephen hemmingerd3428942012-10-01 12:32:35 +00003342
3343static void __exit vxlan_cleanup_module(void)
3344{
Stephen Hemmingerb7153982013-06-17 14:16:09 -07003345 rtnl_link_unregister(&vxlan_link_ops);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003346 unregister_netdevice_notifier(&vxlan_notifier_block);
Daniel Borkmann783c1462014-01-22 21:07:53 +01003347 unregister_pernet_subsys(&vxlan_net_ops);
3348 /* rcu_barrier() is called by netns */
stephen hemmingerd3428942012-10-01 12:32:35 +00003349}
3350module_exit(vxlan_cleanup_module);
3351
3352MODULE_LICENSE("GPL");
3353MODULE_VERSION(VXLAN_VERSION);
stephen hemminger3b8df3c2013-04-27 11:31:52 +00003354MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>");
Jesse Brandeburgead51392014-01-17 11:00:33 -08003355MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic");
stephen hemmingerd3428942012-10-01 12:32:35 +00003356MODULE_ALIAS_RTNL_LINK("vxlan");