blob: 34c519eb1db5092a6e1bd17e02b9a3e53a5c5cb2 [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>
14#include <linux/types.h>
15#include <linux/module.h>
16#include <linux/errno.h>
17#include <linux/slab.h>
18#include <linux/skbuff.h>
19#include <linux/rculist.h>
20#include <linux/netdevice.h>
21#include <linux/in.h>
22#include <linux/ip.h>
23#include <linux/udp.h>
24#include <linux/igmp.h>
25#include <linux/etherdevice.h>
26#include <linux/if_ether.h>
Pravin B Shelar1eaa8172013-08-19 11:23:29 -070027#include <linux/if_vlan.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000028#include <linux/hash.h>
Yan Burman1b13c972013-01-29 23:43:07 +000029#include <linux/ethtool.h>
David Stevense4f67ad2012-11-20 02:50:14 +000030#include <net/arp.h>
31#include <net/ndisc.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000032#include <net/ip.h>
Pravin B Shelarc5441932013-03-25 14:49:35 +000033#include <net/ip_tunnels.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000034#include <net/icmp.h>
35#include <net/udp.h>
Tom Herbert3ee64f32014-07-13 19:49:42 -070036#include <net/udp_tunnel.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000037#include <net/rtnetlink.h>
38#include <net/route.h>
39#include <net/dsfield.h>
40#include <net/inet_ecn.h>
41#include <net/net_namespace.h>
42#include <net/netns/generic.h>
Pravin B Shelar012a5722013-08-19 11:23:07 -070043#include <net/vxlan.h>
Or Gerlitzdc01e7d2014-01-20 13:59:21 +020044#include <net/protocol.h>
Andy Zhouacbf74a2014-09-16 17:31:18 -070045#include <net/udp_tunnel.h>
Cong Wange4c7ed42013-08-31 13:44:33 +080046#if IS_ENABLED(CONFIG_IPV6)
47#include <net/ipv6.h>
48#include <net/addrconf.h>
49#include <net/ip6_tunnel.h>
Cong Wang660d98c2013-09-02 10:06:52 +080050#include <net/ip6_checksum.h>
Cong Wange4c7ed42013-08-31 13:44:33 +080051#endif
stephen hemmingerd3428942012-10-01 12:32:35 +000052
53#define VXLAN_VERSION "0.1"
54
stephen hemminger553675f2013-05-16 11:35:20 +000055#define PORT_HASH_BITS 8
56#define PORT_HASH_SIZE (1<<PORT_HASH_BITS)
stephen hemmingerd3428942012-10-01 12:32:35 +000057#define VNI_HASH_BITS 10
58#define VNI_HASH_SIZE (1<<VNI_HASH_BITS)
59#define FDB_HASH_BITS 8
60#define FDB_HASH_SIZE (1<<FDB_HASH_BITS)
61#define FDB_AGE_DEFAULT 300 /* 5 min */
62#define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */
63
stephen hemminger23c578b2013-04-27 11:31:53 +000064/* UDP port for VXLAN traffic.
65 * The IANA assigned port is 4789, but the Linux default is 8472
Stephen Hemminger234f5b72013-06-17 14:16:41 -070066 * for compatibility with early adopters.
stephen hemminger23c578b2013-04-27 11:31:53 +000067 */
Stephen Hemminger9daaa392013-06-17 14:16:12 -070068static unsigned short vxlan_port __read_mostly = 8472;
69module_param_named(udp_port, vxlan_port, ushort, 0444);
stephen hemmingerd3428942012-10-01 12:32:35 +000070MODULE_PARM_DESC(udp_port, "Destination UDP port");
71
72static bool log_ecn_error = true;
73module_param(log_ecn_error, bool, 0644);
74MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
75
Pravin B Shelar60d9d4c2013-06-20 00:26:31 -070076static int vxlan_net_id;
stephen hemminger553675f2013-05-16 11:35:20 +000077
Mike Rapoportafbd8ba2013-06-25 16:01:51 +030078static const u8 all_zeros_mac[ETH_ALEN];
79
stephen hemminger553675f2013-05-16 11:35:20 +000080/* per-network namespace private data for this module */
81struct vxlan_net {
82 struct list_head vxlan_list;
83 struct hlist_head sock_list[PORT_HASH_SIZE];
Stephen Hemminger1c51a912013-06-17 14:16:11 -070084 spinlock_t sock_lock;
stephen hemminger553675f2013-05-16 11:35:20 +000085};
86
Cong Wange4c7ed42013-08-31 13:44:33 +080087union vxlan_addr {
88 struct sockaddr_in sin;
89 struct sockaddr_in6 sin6;
90 struct sockaddr sa;
91};
92
David Stevens66817122013-03-15 04:35:51 +000093struct vxlan_rdst {
Cong Wange4c7ed42013-08-31 13:44:33 +080094 union vxlan_addr remote_ip;
David Stevens66817122013-03-15 04:35:51 +000095 __be16 remote_port;
96 u32 remote_vni;
97 u32 remote_ifindex;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -070098 struct list_head list;
Mike Rapoportbc7892b2013-06-25 16:01:54 +030099 struct rcu_head rcu;
David Stevens66817122013-03-15 04:35:51 +0000100};
101
stephen hemmingerd3428942012-10-01 12:32:35 +0000102/* Forwarding table entry */
103struct vxlan_fdb {
104 struct hlist_node hlist; /* linked list of entries */
105 struct rcu_head rcu;
106 unsigned long updated; /* jiffies */
107 unsigned long used;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700108 struct list_head remotes;
stephen hemmingerd3428942012-10-01 12:32:35 +0000109 u16 state; /* see ndm_state */
David Stevensae884082013-04-19 00:36:26 +0000110 u8 flags; /* see ndm_flags */
stephen hemmingerd3428942012-10-01 12:32:35 +0000111 u8 eth_addr[ETH_ALEN];
112};
113
stephen hemmingerd3428942012-10-01 12:32:35 +0000114/* Pseudo network device */
115struct vxlan_dev {
stephen hemminger553675f2013-05-16 11:35:20 +0000116 struct hlist_node hlist; /* vni hash table */
117 struct list_head next; /* vxlan's per namespace list */
118 struct vxlan_sock *vn_sock; /* listening socket */
stephen hemmingerd3428942012-10-01 12:32:35 +0000119 struct net_device *dev;
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +0200120 struct net *net; /* netns for packet i/o */
Atzm Watanabec7995c42013-04-16 02:50:52 +0000121 struct vxlan_rdst default_dst; /* default destination */
Cong Wange4c7ed42013-08-31 13:44:33 +0800122 union vxlan_addr saddr; /* source address */
stephen hemminger823aa872013-04-27 11:31:57 +0000123 __be16 dst_port;
stephen hemminger05f47d62012-10-09 20:35:50 +0000124 __u16 port_min; /* source port range */
125 __u16 port_max;
stephen hemmingerd3428942012-10-01 12:32:35 +0000126 __u8 tos; /* TOS override */
127 __u8 ttl;
Tom Herbert359a0ea2014-06-04 17:20:29 -0700128 u32 flags; /* VXLAN_F_* in vxlan.h */
stephen hemmingerd3428942012-10-01 12:32:35 +0000129
130 unsigned long age_interval;
131 struct timer_list age_timer;
132 spinlock_t hash_lock;
133 unsigned int addrcnt;
134 unsigned int addrmax;
stephen hemmingerd3428942012-10-01 12:32:35 +0000135
136 struct hlist_head fdb_head[FDB_HASH_SIZE];
137};
138
139/* salt for hash table */
140static u32 vxlan_salt __read_mostly;
Stephen Hemminger758c57d2013-06-17 14:16:09 -0700141static struct workqueue_struct *vxlan_wq;
stephen hemmingerd3428942012-10-01 12:32:35 +0000142
Cong Wange4c7ed42013-08-31 13:44:33 +0800143#if IS_ENABLED(CONFIG_IPV6)
144static inline
145bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
146{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200147 if (a->sa.sa_family != b->sa.sa_family)
148 return false;
149 if (a->sa.sa_family == AF_INET6)
150 return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr);
151 else
152 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
Cong Wange4c7ed42013-08-31 13:44:33 +0800153}
154
155static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
156{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200157 if (ipa->sa.sa_family == AF_INET6)
158 return ipv6_addr_any(&ipa->sin6.sin6_addr);
159 else
160 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
Cong Wange4c7ed42013-08-31 13:44:33 +0800161}
162
163static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
164{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200165 if (ipa->sa.sa_family == AF_INET6)
166 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
167 else
168 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
Cong Wange4c7ed42013-08-31 13:44:33 +0800169}
170
171static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
172{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200173 if (nla_len(nla) >= sizeof(struct in6_addr)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200174 ip->sin6.sin6_addr = nla_get_in6_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200175 ip->sa.sa_family = AF_INET6;
176 return 0;
177 } else if (nla_len(nla) >= sizeof(__be32)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200178 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200179 ip->sa.sa_family = AF_INET;
180 return 0;
181 } else {
182 return -EAFNOSUPPORT;
183 }
Cong Wange4c7ed42013-08-31 13:44:33 +0800184}
185
186static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
Jiri Bencf0ef3122015-03-29 16:17:37 +0200187 const union vxlan_addr *ip)
Cong Wange4c7ed42013-08-31 13:44:33 +0800188{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200189 if (ip->sa.sa_family == AF_INET6)
Jiri Benc930345e2015-03-29 16:59:25 +0200190 return nla_put_in6_addr(skb, attr, &ip->sin6.sin6_addr);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200191 else
Jiri Benc930345e2015-03-29 16:59:25 +0200192 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +0800193}
194
195#else /* !CONFIG_IPV6 */
196
197static inline
198bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
199{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200200 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
Cong Wange4c7ed42013-08-31 13:44:33 +0800201}
202
203static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
204{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200205 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
Cong Wange4c7ed42013-08-31 13:44:33 +0800206}
207
208static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
209{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200210 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
Cong Wange4c7ed42013-08-31 13:44:33 +0800211}
212
213static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
214{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200215 if (nla_len(nla) >= sizeof(struct in6_addr)) {
216 return -EAFNOSUPPORT;
217 } else if (nla_len(nla) >= sizeof(__be32)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200218 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200219 ip->sa.sa_family = AF_INET;
220 return 0;
221 } else {
222 return -EAFNOSUPPORT;
223 }
Cong Wange4c7ed42013-08-31 13:44:33 +0800224}
225
226static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
Jiri Bencf0ef3122015-03-29 16:17:37 +0200227 const union vxlan_addr *ip)
Cong Wange4c7ed42013-08-31 13:44:33 +0800228{
Jiri Benc930345e2015-03-29 16:59:25 +0200229 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +0800230}
231#endif
232
stephen hemminger553675f2013-05-16 11:35:20 +0000233/* Virtual Network hash table head */
234static inline struct hlist_head *vni_head(struct vxlan_sock *vs, u32 id)
235{
236 return &vs->vni_list[hash_32(id, VNI_HASH_BITS)];
237}
238
239/* Socket hash table head */
240static inline struct hlist_head *vs_head(struct net *net, __be16 port)
stephen hemmingerd3428942012-10-01 12:32:35 +0000241{
242 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
243
stephen hemminger553675f2013-05-16 11:35:20 +0000244 return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)];
245}
246
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700247/* First remote destination for a forwarding entry.
248 * Guaranteed to be non-NULL because remotes are never deleted.
249 */
stephen hemminger5ca54612013-08-04 17:17:39 -0700250static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700251{
stephen hemminger5ca54612013-08-04 17:17:39 -0700252 return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list);
253}
254
255static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb)
256{
257 return list_first_entry(&fdb->remotes, struct vxlan_rdst, list);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700258}
259
Thomas Grafac5132d2015-01-15 03:53:56 +0100260/* Find VXLAN socket based on network namespace, address family and UDP port
261 * and enabled unshareable flags.
262 */
263static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family,
264 __be16 port, u32 flags)
stephen hemminger553675f2013-05-16 11:35:20 +0000265{
266 struct vxlan_sock *vs;
Tom Herbertaf33c1a2015-01-20 11:23:05 -0800267
268 flags &= VXLAN_F_RCV_FLAGS;
stephen hemminger553675f2013-05-16 11:35:20 +0000269
270 hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) {
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -0200271 if (inet_sk(vs->sock->sk)->inet_sport == port &&
Thomas Grafac5132d2015-01-15 03:53:56 +0100272 inet_sk(vs->sock->sk)->sk.sk_family == family &&
Tom Herbertaf33c1a2015-01-20 11:23:05 -0800273 vs->flags == flags)
stephen hemminger553675f2013-05-16 11:35:20 +0000274 return vs;
275 }
276 return NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000277}
278
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700279static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, u32 id)
stephen hemmingerd3428942012-10-01 12:32:35 +0000280{
281 struct vxlan_dev *vxlan;
stephen hemmingerd3428942012-10-01 12:32:35 +0000282
stephen hemminger553675f2013-05-16 11:35:20 +0000283 hlist_for_each_entry_rcu(vxlan, vni_head(vs, id), hlist) {
Atzm Watanabec7995c42013-04-16 02:50:52 +0000284 if (vxlan->default_dst.remote_vni == id)
stephen hemmingerd3428942012-10-01 12:32:35 +0000285 return vxlan;
286 }
287
288 return NULL;
289}
290
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700291/* Look up VNI in a per net namespace table */
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -0200292static struct vxlan_dev *vxlan_find_vni(struct net *net, u32 id,
Thomas Grafac5132d2015-01-15 03:53:56 +0100293 sa_family_t family, __be16 port,
294 u32 flags)
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700295{
296 struct vxlan_sock *vs;
297
Thomas Grafac5132d2015-01-15 03:53:56 +0100298 vs = vxlan_find_sock(net, family, port, flags);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700299 if (!vs)
300 return NULL;
301
302 return vxlan_vs_find_vni(vs, id);
303}
304
stephen hemmingerd3428942012-10-01 12:32:35 +0000305/* Fill in neighbour message in skbuff. */
306static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan,
Stephen Hemminger234f5b72013-06-17 14:16:41 -0700307 const struct vxlan_fdb *fdb,
308 u32 portid, u32 seq, int type, unsigned int flags,
309 const struct vxlan_rdst *rdst)
stephen hemmingerd3428942012-10-01 12:32:35 +0000310{
311 unsigned long now = jiffies;
312 struct nda_cacheinfo ci;
313 struct nlmsghdr *nlh;
314 struct ndmsg *ndm;
David Stevense4f67ad2012-11-20 02:50:14 +0000315 bool send_ip, send_eth;
stephen hemmingerd3428942012-10-01 12:32:35 +0000316
317 nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags);
318 if (nlh == NULL)
319 return -EMSGSIZE;
320
321 ndm = nlmsg_data(nlh);
322 memset(ndm, 0, sizeof(*ndm));
David Stevense4f67ad2012-11-20 02:50:14 +0000323
324 send_eth = send_ip = true;
325
326 if (type == RTM_GETNEIGH) {
327 ndm->ndm_family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +0800328 send_ip = !vxlan_addr_any(&rdst->remote_ip);
David Stevense4f67ad2012-11-20 02:50:14 +0000329 send_eth = !is_zero_ether_addr(fdb->eth_addr);
330 } else
331 ndm->ndm_family = AF_BRIDGE;
stephen hemmingerd3428942012-10-01 12:32:35 +0000332 ndm->ndm_state = fdb->state;
333 ndm->ndm_ifindex = vxlan->dev->ifindex;
David Stevensae884082013-04-19 00:36:26 +0000334 ndm->ndm_flags = fdb->flags;
Jun Zhao545469f2014-07-26 00:38:59 +0800335 ndm->ndm_type = RTN_UNICAST;
stephen hemmingerd3428942012-10-01 12:32:35 +0000336
Nicolas Dichtel193523b2015-01-20 15:15:47 +0100337 if (!net_eq(dev_net(vxlan->dev), vxlan->net) &&
Nicolas Dichtel49670822015-01-26 14:10:53 +0100338 nla_put_s32(skb, NDA_LINK_NETNSID,
Nicolas Dichtel7a0877d2015-05-07 11:02:49 +0200339 peernet2id_alloc(dev_net(vxlan->dev), vxlan->net)))
Nicolas Dichtel193523b2015-01-20 15:15:47 +0100340 goto nla_put_failure;
341
David Stevense4f67ad2012-11-20 02:50:14 +0000342 if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000343 goto nla_put_failure;
344
Cong Wange4c7ed42013-08-31 13:44:33 +0800345 if (send_ip && vxlan_nla_put_addr(skb, NDA_DST, &rdst->remote_ip))
David Stevens66817122013-03-15 04:35:51 +0000346 goto nla_put_failure;
347
stephen hemminger823aa872013-04-27 11:31:57 +0000348 if (rdst->remote_port && rdst->remote_port != vxlan->dst_port &&
David Stevens66817122013-03-15 04:35:51 +0000349 nla_put_be16(skb, NDA_PORT, rdst->remote_port))
350 goto nla_put_failure;
Atzm Watanabec7995c42013-04-16 02:50:52 +0000351 if (rdst->remote_vni != vxlan->default_dst.remote_vni &&
Pravin B Shelar60d9d4c2013-06-20 00:26:31 -0700352 nla_put_u32(skb, NDA_VNI, rdst->remote_vni))
David Stevens66817122013-03-15 04:35:51 +0000353 goto nla_put_failure;
354 if (rdst->remote_ifindex &&
355 nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +0000356 goto nla_put_failure;
357
358 ci.ndm_used = jiffies_to_clock_t(now - fdb->used);
359 ci.ndm_confirmed = 0;
360 ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated);
361 ci.ndm_refcnt = 0;
362
363 if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci))
364 goto nla_put_failure;
365
Johannes Berg053c0952015-01-16 22:09:00 +0100366 nlmsg_end(skb, nlh);
367 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +0000368
369nla_put_failure:
370 nlmsg_cancel(skb, nlh);
371 return -EMSGSIZE;
372}
373
374static inline size_t vxlan_nlmsg_size(void)
375{
376 return NLMSG_ALIGN(sizeof(struct ndmsg))
377 + nla_total_size(ETH_ALEN) /* NDA_LLADDR */
Cong Wange4c7ed42013-08-31 13:44:33 +0800378 + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */
stephen hemminger73cf3312013-04-27 11:31:54 +0000379 + nla_total_size(sizeof(__be16)) /* NDA_PORT */
David Stevens66817122013-03-15 04:35:51 +0000380 + nla_total_size(sizeof(__be32)) /* NDA_VNI */
381 + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */
Nicolas Dichtel49670822015-01-26 14:10:53 +0100382 + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */
stephen hemmingerd3428942012-10-01 12:32:35 +0000383 + nla_total_size(sizeof(struct nda_cacheinfo));
384}
385
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200386static void vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
387 struct vxlan_rdst *rd, int type)
stephen hemmingerd3428942012-10-01 12:32:35 +0000388{
389 struct net *net = dev_net(vxlan->dev);
390 struct sk_buff *skb;
391 int err = -ENOBUFS;
392
393 skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC);
394 if (skb == NULL)
395 goto errout;
396
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200397 err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd);
stephen hemmingerd3428942012-10-01 12:32:35 +0000398 if (err < 0) {
399 /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */
400 WARN_ON(err == -EMSGSIZE);
401 kfree_skb(skb);
402 goto errout;
403 }
404
405 rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC);
406 return;
407errout:
408 if (err < 0)
409 rtnl_set_sk_err(net, RTNLGRP_NEIGH, err);
410}
411
Cong Wange4c7ed42013-08-31 13:44:33 +0800412static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa)
David Stevense4f67ad2012-11-20 02:50:14 +0000413{
414 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700415 struct vxlan_fdb f = {
416 .state = NUD_STALE,
417 };
418 struct vxlan_rdst remote = {
Cong Wange4c7ed42013-08-31 13:44:33 +0800419 .remote_ip = *ipa, /* goes to NDA_DST */
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700420 .remote_vni = VXLAN_N_VID,
421 };
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700422
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200423 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH);
David Stevense4f67ad2012-11-20 02:50:14 +0000424}
425
426static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN])
427{
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700428 struct vxlan_fdb f = {
429 .state = NUD_STALE,
430 };
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200431 struct vxlan_rdst remote = { };
David Stevense4f67ad2012-11-20 02:50:14 +0000432
David Stevense4f67ad2012-11-20 02:50:14 +0000433 memcpy(f.eth_addr, eth_addr, ETH_ALEN);
434
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200435 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH);
David Stevense4f67ad2012-11-20 02:50:14 +0000436}
437
stephen hemmingerd3428942012-10-01 12:32:35 +0000438/* Hash Ethernet address */
439static u32 eth_hash(const unsigned char *addr)
440{
441 u64 value = get_unaligned((u64 *)addr);
442
443 /* only want 6 bytes */
444#ifdef __BIG_ENDIAN
stephen hemmingerd3428942012-10-01 12:32:35 +0000445 value >>= 16;
stephen hemminger321fb992012-10-09 20:35:47 +0000446#else
447 value <<= 16;
stephen hemmingerd3428942012-10-01 12:32:35 +0000448#endif
449 return hash_64(value, FDB_HASH_BITS);
450}
451
452/* Hash chain to use given mac address */
453static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan,
454 const u8 *mac)
455{
456 return &vxlan->fdb_head[eth_hash(mac)];
457}
458
459/* Look up Ethernet address in forwarding table */
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000460static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan,
stephen hemmingerd3428942012-10-01 12:32:35 +0000461 const u8 *mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000462{
463 struct hlist_head *head = vxlan_fdb_head(vxlan, mac);
464 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000465
Sasha Levinb67bfe02013-02-27 17:06:00 -0800466 hlist_for_each_entry_rcu(f, head, hlist) {
Joe Perches7367d0b2013-09-01 11:51:23 -0700467 if (ether_addr_equal(mac, f->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000468 return f;
469 }
470
471 return NULL;
472}
473
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000474static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan,
475 const u8 *mac)
476{
477 struct vxlan_fdb *f;
478
479 f = __vxlan_find_mac(vxlan, mac);
480 if (f)
481 f->used = jiffies;
482
483 return f;
484}
485
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300486/* caller should hold vxlan->hash_lock */
487static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800488 union vxlan_addr *ip, __be16 port,
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300489 __u32 vni, __u32 ifindex)
490{
491 struct vxlan_rdst *rd;
492
493 list_for_each_entry(rd, &f->remotes, list) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800494 if (vxlan_addr_equal(&rd->remote_ip, ip) &&
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300495 rd->remote_port == port &&
496 rd->remote_vni == vni &&
497 rd->remote_ifindex == ifindex)
498 return rd;
499 }
500
501 return NULL;
502}
503
Thomas Richter906dc182013-07-19 17:20:07 +0200504/* Replace destination of unicast mac */
505static int vxlan_fdb_replace(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800506 union vxlan_addr *ip, __be16 port, __u32 vni, __u32 ifindex)
Thomas Richter906dc182013-07-19 17:20:07 +0200507{
508 struct vxlan_rdst *rd;
509
510 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
511 if (rd)
512 return 0;
513
514 rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list);
515 if (!rd)
516 return 0;
Cong Wange4c7ed42013-08-31 13:44:33 +0800517 rd->remote_ip = *ip;
Thomas Richter906dc182013-07-19 17:20:07 +0200518 rd->remote_port = port;
519 rd->remote_vni = vni;
520 rd->remote_ifindex = ifindex;
521 return 1;
522}
523
David Stevens66817122013-03-15 04:35:51 +0000524/* Add/update destinations for multicast */
525static int vxlan_fdb_append(struct vxlan_fdb *f,
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200526 union vxlan_addr *ip, __be16 port, __u32 vni,
527 __u32 ifindex, struct vxlan_rdst **rdp)
David Stevens66817122013-03-15 04:35:51 +0000528{
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700529 struct vxlan_rdst *rd;
David Stevens66817122013-03-15 04:35:51 +0000530
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300531 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
532 if (rd)
533 return 0;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700534
David Stevens66817122013-03-15 04:35:51 +0000535 rd = kmalloc(sizeof(*rd), GFP_ATOMIC);
536 if (rd == NULL)
537 return -ENOBUFS;
Cong Wange4c7ed42013-08-31 13:44:33 +0800538 rd->remote_ip = *ip;
David Stevens66817122013-03-15 04:35:51 +0000539 rd->remote_port = port;
540 rd->remote_vni = vni;
541 rd->remote_ifindex = ifindex;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700542
543 list_add_tail_rcu(&rd->list, &f->remotes);
544
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200545 *rdp = rd;
David Stevens66817122013-03-15 04:35:51 +0000546 return 1;
547}
548
Tom Herbertdfd86452015-01-12 17:00:38 -0800549static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb,
550 unsigned int off,
551 struct vxlanhdr *vh, size_t hdrlen,
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800552 u32 data, struct gro_remcsum *grc,
553 bool nopartial)
Tom Herbertdfd86452015-01-12 17:00:38 -0800554{
555 size_t start, offset, plen;
Tom Herbertdfd86452015-01-12 17:00:38 -0800556
557 if (skb->remcsum_offload)
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800558 return NULL;
Tom Herbertdfd86452015-01-12 17:00:38 -0800559
560 if (!NAPI_GRO_CB(skb)->csum_valid)
561 return NULL;
562
563 start = (data & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT;
564 offset = start + ((data & VXLAN_RCO_UDP) ?
565 offsetof(struct udphdr, check) :
566 offsetof(struct tcphdr, check));
567
568 plen = hdrlen + offset + sizeof(u16);
569
570 /* Pull checksum that will be written */
571 if (skb_gro_header_hard(skb, off + plen)) {
572 vh = skb_gro_header_slow(skb, off + plen, off);
573 if (!vh)
574 return NULL;
575 }
576
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800577 skb_gro_remcsum_process(skb, (void *)vh + hdrlen,
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800578 start, offset, grc, nopartial);
Tom Herbertdfd86452015-01-12 17:00:38 -0800579
580 skb->remcsum_offload = 1;
581
582 return vh;
583}
584
Tom Herberta2b12f32015-01-12 17:00:37 -0800585static struct sk_buff **vxlan_gro_receive(struct sk_buff **head,
586 struct sk_buff *skb,
587 struct udp_offload *uoff)
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200588{
589 struct sk_buff *p, **pp = NULL;
590 struct vxlanhdr *vh, *vh2;
Jesse Gross9b174d82014-12-30 19:10:15 -0800591 unsigned int hlen, off_vx;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200592 int flush = 1;
Tom Herbertdfd86452015-01-12 17:00:38 -0800593 struct vxlan_sock *vs = container_of(uoff, struct vxlan_sock,
594 udp_offloads);
595 u32 flags;
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800596 struct gro_remcsum grc;
597
598 skb_gro_remcsum_init(&grc);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200599
600 off_vx = skb_gro_offset(skb);
601 hlen = off_vx + sizeof(*vh);
602 vh = skb_gro_header_fast(skb, off_vx);
603 if (skb_gro_header_hard(skb, hlen)) {
604 vh = skb_gro_header_slow(skb, hlen, off_vx);
605 if (unlikely(!vh))
606 goto out;
607 }
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200608
Tom Herbertdfd86452015-01-12 17:00:38 -0800609 skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */
610 skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr));
611
612 flags = ntohl(vh->vx_flags);
613
614 if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) {
615 vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr),
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800616 ntohl(vh->vx_vni), &grc,
617 !!(vs->flags &
618 VXLAN_F_REMCSUM_NOPARTIAL));
Tom Herbertdfd86452015-01-12 17:00:38 -0800619
620 if (!vh)
621 goto out;
622 }
623
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200624 flush = 0;
625
626 for (p = *head; p; p = p->next) {
627 if (!NAPI_GRO_CB(p)->same_flow)
628 continue;
629
630 vh2 = (struct vxlanhdr *)(p->data + off_vx);
Thomas Graf35114942015-01-15 03:53:55 +0100631 if (vh->vx_flags != vh2->vx_flags ||
632 vh->vx_vni != vh2->vx_vni) {
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200633 NAPI_GRO_CB(p)->same_flow = 0;
634 continue;
635 }
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200636 }
637
Jesse Gross9b174d82014-12-30 19:10:15 -0800638 pp = eth_gro_receive(head, skb);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200639
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200640out:
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800641 skb_gro_remcsum_cleanup(skb, &grc);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200642 NAPI_GRO_CB(skb)->flush |= flush;
643
644 return pp;
645}
646
Tom Herberta2b12f32015-01-12 17:00:37 -0800647static int vxlan_gro_complete(struct sk_buff *skb, int nhoff,
648 struct udp_offload *uoff)
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200649{
Jesse Grosscfdf1e12014-11-10 11:45:13 -0800650 udp_tunnel_gro_complete(skb, nhoff);
651
Jesse Gross9b174d82014-12-30 19:10:15 -0800652 return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr));
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200653}
654
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700655/* Notify netdevs that UDP port started listening */
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200656static void vxlan_notify_add_rx_port(struct vxlan_sock *vs)
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700657{
658 struct net_device *dev;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200659 struct sock *sk = vs->sock->sk;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700660 struct net *net = sock_net(sk);
661 sa_family_t sa_family = sk->sk_family;
Joseph Gasparakis35e42372013-09-13 07:34:13 -0700662 __be16 port = inet_sk(sk)->inet_sport;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200663 int err;
664
665 if (sa_family == AF_INET) {
666 err = udp_add_offload(&vs->udp_offloads);
667 if (err)
668 pr_warn("vxlan: udp_add_offload failed with status %d\n", err);
669 }
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700670
671 rcu_read_lock();
672 for_each_netdev_rcu(net, dev) {
673 if (dev->netdev_ops->ndo_add_vxlan_port)
674 dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
675 port);
676 }
677 rcu_read_unlock();
678}
679
680/* Notify netdevs that UDP port is no more listening */
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200681static void vxlan_notify_del_rx_port(struct vxlan_sock *vs)
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700682{
683 struct net_device *dev;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200684 struct sock *sk = vs->sock->sk;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700685 struct net *net = sock_net(sk);
686 sa_family_t sa_family = sk->sk_family;
Joseph Gasparakis35e42372013-09-13 07:34:13 -0700687 __be16 port = inet_sk(sk)->inet_sport;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700688
689 rcu_read_lock();
690 for_each_netdev_rcu(net, dev) {
691 if (dev->netdev_ops->ndo_del_vxlan_port)
692 dev->netdev_ops->ndo_del_vxlan_port(dev, sa_family,
693 port);
694 }
695 rcu_read_unlock();
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200696
697 if (sa_family == AF_INET)
698 udp_del_offload(&vs->udp_offloads);
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700699}
700
stephen hemmingerd3428942012-10-01 12:32:35 +0000701/* Add new entry to forwarding table -- assumes lock held */
702static int vxlan_fdb_create(struct vxlan_dev *vxlan,
Cong Wange4c7ed42013-08-31 13:44:33 +0800703 const u8 *mac, union vxlan_addr *ip,
David Stevens66817122013-03-15 04:35:51 +0000704 __u16 state, __u16 flags,
stephen hemminger73cf3312013-04-27 11:31:54 +0000705 __be16 port, __u32 vni, __u32 ifindex,
David Stevensae884082013-04-19 00:36:26 +0000706 __u8 ndm_flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000707{
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200708 struct vxlan_rdst *rd = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000709 struct vxlan_fdb *f;
710 int notify = 0;
711
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000712 f = __vxlan_find_mac(vxlan, mac);
stephen hemmingerd3428942012-10-01 12:32:35 +0000713 if (f) {
714 if (flags & NLM_F_EXCL) {
715 netdev_dbg(vxlan->dev,
716 "lost race to create %pM\n", mac);
717 return -EEXIST;
718 }
719 if (f->state != state) {
720 f->state = state;
721 f->updated = jiffies;
722 notify = 1;
723 }
David Stevensae884082013-04-19 00:36:26 +0000724 if (f->flags != ndm_flags) {
725 f->flags = ndm_flags;
726 f->updated = jiffies;
727 notify = 1;
728 }
Thomas Richter906dc182013-07-19 17:20:07 +0200729 if ((flags & NLM_F_REPLACE)) {
730 /* Only change unicasts */
731 if (!(is_multicast_ether_addr(f->eth_addr) ||
732 is_zero_ether_addr(f->eth_addr))) {
Li RongQing60840422015-04-22 15:49:10 +0800733 notify |= vxlan_fdb_replace(f, ip, port, vni,
Thomas Richter906dc182013-07-19 17:20:07 +0200734 ifindex);
Thomas Richter906dc182013-07-19 17:20:07 +0200735 } else
736 return -EOPNOTSUPP;
737 }
David Stevens66817122013-03-15 04:35:51 +0000738 if ((flags & NLM_F_APPEND) &&
Mike Rapoport58e4c762013-06-25 16:01:56 +0300739 (is_multicast_ether_addr(f->eth_addr) ||
740 is_zero_ether_addr(f->eth_addr))) {
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200741 int rc = vxlan_fdb_append(f, ip, port, vni, ifindex,
742 &rd);
David Stevens66817122013-03-15 04:35:51 +0000743
744 if (rc < 0)
745 return rc;
746 notify |= rc;
747 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000748 } else {
749 if (!(flags & NLM_F_CREATE))
750 return -ENOENT;
751
752 if (vxlan->addrmax && vxlan->addrcnt >= vxlan->addrmax)
753 return -ENOSPC;
754
Thomas Richter906dc182013-07-19 17:20:07 +0200755 /* Disallow replace to add a multicast entry */
756 if ((flags & NLM_F_REPLACE) &&
757 (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac)))
758 return -EOPNOTSUPP;
759
Cong Wange4c7ed42013-08-31 13:44:33 +0800760 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
stephen hemmingerd3428942012-10-01 12:32:35 +0000761 f = kmalloc(sizeof(*f), GFP_ATOMIC);
762 if (!f)
763 return -ENOMEM;
764
765 notify = 1;
stephen hemmingerd3428942012-10-01 12:32:35 +0000766 f->state = state;
David Stevensae884082013-04-19 00:36:26 +0000767 f->flags = ndm_flags;
stephen hemmingerd3428942012-10-01 12:32:35 +0000768 f->updated = f->used = jiffies;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700769 INIT_LIST_HEAD(&f->remotes);
stephen hemmingerd3428942012-10-01 12:32:35 +0000770 memcpy(f->eth_addr, mac, ETH_ALEN);
771
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200772 vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700773
stephen hemmingerd3428942012-10-01 12:32:35 +0000774 ++vxlan->addrcnt;
775 hlist_add_head_rcu(&f->hlist,
776 vxlan_fdb_head(vxlan, mac));
777 }
778
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200779 if (notify) {
780 if (rd == NULL)
781 rd = first_remote_rtnl(f);
782 vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH);
783 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000784
785 return 0;
786}
787
Wei Yongjun6706c822013-04-11 19:00:35 +0000788static void vxlan_fdb_free(struct rcu_head *head)
David Stevens66817122013-03-15 04:35:51 +0000789{
790 struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700791 struct vxlan_rdst *rd, *nd;
David Stevens66817122013-03-15 04:35:51 +0000792
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700793 list_for_each_entry_safe(rd, nd, &f->remotes, list)
David Stevens66817122013-03-15 04:35:51 +0000794 kfree(rd);
David Stevens66817122013-03-15 04:35:51 +0000795 kfree(f);
796}
797
stephen hemmingerd3428942012-10-01 12:32:35 +0000798static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f)
799{
800 netdev_dbg(vxlan->dev,
801 "delete %pM\n", f->eth_addr);
802
803 --vxlan->addrcnt;
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200804 vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_DELNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +0000805
806 hlist_del_rcu(&f->hlist);
David Stevens66817122013-03-15 04:35:51 +0000807 call_rcu(&f->rcu, vxlan_fdb_free);
stephen hemmingerd3428942012-10-01 12:32:35 +0000808}
809
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300810static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan,
Cong Wange4c7ed42013-08-31 13:44:33 +0800811 union vxlan_addr *ip, __be16 *port, u32 *vni, u32 *ifindex)
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300812{
813 struct net *net = dev_net(vxlan->dev);
Cong Wange4c7ed42013-08-31 13:44:33 +0800814 int err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300815
816 if (tb[NDA_DST]) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800817 err = vxlan_nla_get_addr(ip, tb[NDA_DST]);
818 if (err)
819 return err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300820 } else {
Cong Wange4c7ed42013-08-31 13:44:33 +0800821 union vxlan_addr *remote = &vxlan->default_dst.remote_ip;
822 if (remote->sa.sa_family == AF_INET) {
823 ip->sin.sin_addr.s_addr = htonl(INADDR_ANY);
824 ip->sa.sa_family = AF_INET;
825#if IS_ENABLED(CONFIG_IPV6)
826 } else {
827 ip->sin6.sin6_addr = in6addr_any;
828 ip->sa.sa_family = AF_INET6;
829#endif
830 }
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300831 }
832
833 if (tb[NDA_PORT]) {
834 if (nla_len(tb[NDA_PORT]) != sizeof(__be16))
835 return -EINVAL;
836 *port = nla_get_be16(tb[NDA_PORT]);
837 } else {
838 *port = vxlan->dst_port;
839 }
840
841 if (tb[NDA_VNI]) {
842 if (nla_len(tb[NDA_VNI]) != sizeof(u32))
843 return -EINVAL;
844 *vni = nla_get_u32(tb[NDA_VNI]);
845 } else {
846 *vni = vxlan->default_dst.remote_vni;
847 }
848
849 if (tb[NDA_IFINDEX]) {
850 struct net_device *tdev;
851
852 if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32))
853 return -EINVAL;
854 *ifindex = nla_get_u32(tb[NDA_IFINDEX]);
Ying Xue73763942014-01-15 10:23:41 +0800855 tdev = __dev_get_by_index(net, *ifindex);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300856 if (!tdev)
857 return -EADDRNOTAVAIL;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300858 } else {
859 *ifindex = 0;
860 }
861
862 return 0;
863}
864
stephen hemmingerd3428942012-10-01 12:32:35 +0000865/* Add static entry (via netlink) */
866static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[],
867 struct net_device *dev,
Jiri Pirkof6f64242014-11-28 14:34:15 +0100868 const unsigned char *addr, u16 vid, u16 flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000869{
870 struct vxlan_dev *vxlan = netdev_priv(dev);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300871 /* struct net *net = dev_net(vxlan->dev); */
Cong Wange4c7ed42013-08-31 13:44:33 +0800872 union vxlan_addr ip;
stephen hemminger73cf3312013-04-27 11:31:54 +0000873 __be16 port;
874 u32 vni, ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +0000875 int err;
876
877 if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) {
878 pr_info("RTM_NEWNEIGH with invalid state %#x\n",
879 ndm->ndm_state);
880 return -EINVAL;
881 }
882
883 if (tb[NDA_DST] == NULL)
884 return -EINVAL;
885
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300886 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
887 if (err)
888 return err;
David Stevens66817122013-03-15 04:35:51 +0000889
Mike Rapoport5933a7b2014-04-01 09:23:01 +0300890 if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family)
891 return -EAFNOSUPPORT;
892
stephen hemmingerd3428942012-10-01 12:32:35 +0000893 spin_lock_bh(&vxlan->hash_lock);
Cong Wange4c7ed42013-08-31 13:44:33 +0800894 err = vxlan_fdb_create(vxlan, addr, &ip, ndm->ndm_state, flags,
stephen hemminger73cf3312013-04-27 11:31:54 +0000895 port, vni, ifindex, ndm->ndm_flags);
stephen hemmingerd3428942012-10-01 12:32:35 +0000896 spin_unlock_bh(&vxlan->hash_lock);
897
898 return err;
899}
900
901/* Delete entry (via netlink) */
Vlad Yasevich1690be62013-02-13 12:00:18 +0000902static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[],
903 struct net_device *dev,
Jiri Pirkof6f64242014-11-28 14:34:15 +0100904 const unsigned char *addr, u16 vid)
stephen hemmingerd3428942012-10-01 12:32:35 +0000905{
906 struct vxlan_dev *vxlan = netdev_priv(dev);
907 struct vxlan_fdb *f;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300908 struct vxlan_rdst *rd = NULL;
Cong Wange4c7ed42013-08-31 13:44:33 +0800909 union vxlan_addr ip;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300910 __be16 port;
911 u32 vni, ifindex;
912 int err;
913
914 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
915 if (err)
916 return err;
917
918 err = -ENOENT;
stephen hemmingerd3428942012-10-01 12:32:35 +0000919
920 spin_lock_bh(&vxlan->hash_lock);
921 f = vxlan_find_mac(vxlan, addr);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300922 if (!f)
923 goto out;
924
Cong Wange4c7ed42013-08-31 13:44:33 +0800925 if (!vxlan_addr_any(&ip)) {
926 rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300927 if (!rd)
928 goto out;
stephen hemmingerd3428942012-10-01 12:32:35 +0000929 }
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300930
931 err = 0;
932
933 /* remove a destination if it's not the only one on the list,
934 * otherwise destroy the fdb entry
935 */
936 if (rd && !list_is_singular(&f->remotes)) {
937 list_del_rcu(&rd->list);
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200938 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH);
Wei Yongjundcdc7a52013-08-17 07:32:09 +0800939 kfree_rcu(rd, rcu);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300940 goto out;
941 }
942
943 vxlan_fdb_destroy(vxlan, f);
944
945out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000946 spin_unlock_bh(&vxlan->hash_lock);
947
948 return err;
949}
950
951/* Dump forwarding table */
952static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb,
Jamal Hadi Salim5d5eacb2014-07-10 07:01:58 -0400953 struct net_device *dev,
954 struct net_device *filter_dev, int idx)
stephen hemmingerd3428942012-10-01 12:32:35 +0000955{
956 struct vxlan_dev *vxlan = netdev_priv(dev);
957 unsigned int h;
958
959 for (h = 0; h < FDB_HASH_SIZE; ++h) {
960 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000961 int err;
962
Sasha Levinb67bfe02013-02-27 17:06:00 -0800963 hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) {
David Stevens66817122013-03-15 04:35:51 +0000964 struct vxlan_rdst *rd;
stephen hemmingerd3428942012-10-01 12:32:35 +0000965
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700966 if (idx < cb->args[0])
967 goto skip;
968
969 list_for_each_entry_rcu(rd, &f->remotes, list) {
David Stevens66817122013-03-15 04:35:51 +0000970 err = vxlan_fdb_info(skb, vxlan, f,
971 NETLINK_CB(cb->skb).portid,
972 cb->nlh->nlmsg_seq,
973 RTM_NEWNEIGH,
974 NLM_F_MULTI, rd);
975 if (err < 0)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700976 goto out;
David Stevens66817122013-03-15 04:35:51 +0000977 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700978skip:
979 ++idx;
stephen hemmingerd3428942012-10-01 12:32:35 +0000980 }
981 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700982out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000983 return idx;
984}
985
986/* Watch incoming packets to learn mapping between Ethernet address
987 * and Tunnel endpoint.
Simon Hormanc4b49512015-04-02 11:17:58 +0900988 * Return true if packet is bogus and should be dropped.
stephen hemmingerd3428942012-10-01 12:32:35 +0000989 */
stephen hemminger26a41ae62013-06-17 12:09:58 -0700990static bool vxlan_snoop(struct net_device *dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800991 union vxlan_addr *src_ip, const u8 *src_mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000992{
993 struct vxlan_dev *vxlan = netdev_priv(dev);
994 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000995
996 f = vxlan_find_mac(vxlan, src_mac);
997 if (likely(f)) {
stephen hemminger5ca54612013-08-04 17:17:39 -0700998 struct vxlan_rdst *rdst = first_remote_rcu(f);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700999
Cong Wange4c7ed42013-08-31 13:44:33 +08001000 if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip)))
stephen hemminger26a41ae62013-06-17 12:09:58 -07001001 return false;
1002
1003 /* Don't migrate static entries, drop packets */
stephen hemmingereb064c32013-06-18 14:27:01 -07001004 if (f->state & NUD_NOARP)
stephen hemminger26a41ae62013-06-17 12:09:58 -07001005 return true;
stephen hemmingerd3428942012-10-01 12:32:35 +00001006
1007 if (net_ratelimit())
1008 netdev_info(dev,
Cong Wange4c7ed42013-08-31 13:44:33 +08001009 "%pM migrated from %pIS to %pIS\n",
Rasmus Villemoesa4870f72015-02-07 03:17:31 +01001010 src_mac, &rdst->remote_ip.sa, &src_ip->sa);
stephen hemmingerd3428942012-10-01 12:32:35 +00001011
Cong Wange4c7ed42013-08-31 13:44:33 +08001012 rdst->remote_ip = *src_ip;
stephen hemmingerd3428942012-10-01 12:32:35 +00001013 f->updated = jiffies;
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +02001014 vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +00001015 } else {
1016 /* learned new entry */
1017 spin_lock(&vxlan->hash_lock);
stephen hemminger3bf74b12013-06-17 12:09:57 -07001018
1019 /* close off race between vxlan_flush and incoming packets */
1020 if (netif_running(dev))
1021 vxlan_fdb_create(vxlan, src_mac, src_ip,
1022 NUD_REACHABLE,
1023 NLM_F_EXCL|NLM_F_CREATE,
1024 vxlan->dst_port,
1025 vxlan->default_dst.remote_vni,
1026 0, NTF_SELF);
stephen hemmingerd3428942012-10-01 12:32:35 +00001027 spin_unlock(&vxlan->hash_lock);
1028 }
stephen hemminger26a41ae62013-06-17 12:09:58 -07001029
1030 return false;
stephen hemmingerd3428942012-10-01 12:32:35 +00001031}
1032
stephen hemmingerd3428942012-10-01 12:32:35 +00001033/* See if multicast group is already in use by other ID */
Gao feng95ab0992013-12-10 16:37:33 +08001034static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev)
stephen hemmingerd3428942012-10-01 12:32:35 +00001035{
stephen hemminger553675f2013-05-16 11:35:20 +00001036 struct vxlan_dev *vxlan;
stephen hemmingerd3428942012-10-01 12:32:35 +00001037
Gao feng95ab0992013-12-10 16:37:33 +08001038 /* The vxlan_sock is only used by dev, leaving group has
1039 * no effect on other vxlan devices.
1040 */
1041 if (atomic_read(&dev->vn_sock->refcnt) == 1)
1042 return false;
1043
stephen hemminger553675f2013-05-16 11:35:20 +00001044 list_for_each_entry(vxlan, &vn->vxlan_list, next) {
Gao feng95ab0992013-12-10 16:37:33 +08001045 if (!netif_running(vxlan->dev) || vxlan == dev)
stephen hemminger553675f2013-05-16 11:35:20 +00001046 continue;
stephen hemmingerd3428942012-10-01 12:32:35 +00001047
Gao feng95ab0992013-12-10 16:37:33 +08001048 if (vxlan->vn_sock != dev->vn_sock)
1049 continue;
1050
1051 if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip,
1052 &dev->default_dst.remote_ip))
1053 continue;
1054
1055 if (vxlan->default_dst.remote_ifindex !=
1056 dev->default_dst.remote_ifindex)
1057 continue;
1058
1059 return true;
stephen hemminger553675f2013-05-16 11:35:20 +00001060 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001061
1062 return false;
1063}
1064
Pravin B Shelar012a5722013-08-19 11:23:07 -07001065void vxlan_sock_release(struct vxlan_sock *vs)
stephen hemmingerd3428942012-10-01 12:32:35 +00001066{
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07001067 struct sock *sk = vs->sock->sk;
1068 struct net *net = sock_net(sk);
1069 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Pravin B Shelar012a5722013-08-19 11:23:07 -07001070
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001071 if (!atomic_dec_and_test(&vs->refcnt))
1072 return;
1073
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001074 spin_lock(&vn->sock_lock);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001075 hlist_del_rcu(&vs->hlist);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +02001076 vxlan_notify_del_rx_port(vs);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001077 spin_unlock(&vn->sock_lock);
1078
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001079 queue_work(vxlan_wq, &vs->del_work);
1080}
Pravin B Shelar012a5722013-08-19 11:23:07 -07001081EXPORT_SYMBOL_GPL(vxlan_sock_release);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001082
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001083/* Update multicast group membership when first VNI on
Simon Hormanc4b49512015-04-02 11:17:58 +09001084 * multicast address is brought up
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001085 */
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001086static int vxlan_igmp_join(struct vxlan_dev *vxlan)
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001087{
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001088 struct vxlan_sock *vs = vxlan->vn_sock;
1089 struct sock *sk = vs->sock->sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001090 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1091 int ifindex = vxlan->default_dst.remote_ifindex;
Marcelo Ricardo Leitner149d7542015-03-20 10:26:21 -03001092 int ret = -EINVAL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001093
stephen hemmingerd3428942012-10-01 12:32:35 +00001094 lock_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001095 if (ip->sa.sa_family == AF_INET) {
1096 struct ip_mreqn mreq = {
1097 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1098 .imr_ifindex = ifindex,
1099 };
1100
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001101 ret = ip_mc_join_group(sk, &mreq);
Cong Wange4c7ed42013-08-31 13:44:33 +08001102#if IS_ENABLED(CONFIG_IPV6)
1103 } else {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001104 ret = ipv6_stub->ipv6_sock_mc_join(sk, ifindex,
1105 &ip->sin6.sin6_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +08001106#endif
1107 }
stephen hemminger3fc2de22013-07-18 08:40:15 -07001108 release_sock(sk);
1109
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001110 return ret;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001111}
1112
1113/* Inverse of vxlan_igmp_join when last VNI is brought down */
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001114static int vxlan_igmp_leave(struct vxlan_dev *vxlan)
stephen hemminger3fc2de22013-07-18 08:40:15 -07001115{
stephen hemminger3fc2de22013-07-18 08:40:15 -07001116 struct vxlan_sock *vs = vxlan->vn_sock;
1117 struct sock *sk = vs->sock->sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001118 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1119 int ifindex = vxlan->default_dst.remote_ifindex;
Marcelo Ricardo Leitner149d7542015-03-20 10:26:21 -03001120 int ret = -EINVAL;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001121
1122 lock_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001123 if (ip->sa.sa_family == AF_INET) {
1124 struct ip_mreqn mreq = {
1125 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1126 .imr_ifindex = ifindex,
1127 };
1128
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001129 ret = ip_mc_leave_group(sk, &mreq);
Cong Wange4c7ed42013-08-31 13:44:33 +08001130#if IS_ENABLED(CONFIG_IPV6)
1131 } else {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001132 ret = ipv6_stub->ipv6_sock_mc_drop(sk, ifindex,
1133 &ip->sin6.sin6_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +08001134#endif
1135 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001136 release_sock(sk);
stephen hemmingerd3428942012-10-01 12:32:35 +00001137
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001138 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00001139}
1140
Tom Herbertdfd86452015-01-12 17:00:38 -08001141static struct vxlanhdr *vxlan_remcsum(struct sk_buff *skb, struct vxlanhdr *vh,
Tom Herbert0ace2ca2015-02-10 16:30:32 -08001142 size_t hdrlen, u32 data, bool nopartial)
Tom Herbertdfd86452015-01-12 17:00:38 -08001143{
1144 size_t start, offset, plen;
Tom Herbertdfd86452015-01-12 17:00:38 -08001145
Tom Herbertdfd86452015-01-12 17:00:38 -08001146 start = (data & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT;
1147 offset = start + ((data & VXLAN_RCO_UDP) ?
1148 offsetof(struct udphdr, check) :
1149 offsetof(struct tcphdr, check));
1150
1151 plen = hdrlen + offset + sizeof(u16);
1152
1153 if (!pskb_may_pull(skb, plen))
1154 return NULL;
1155
1156 vh = (struct vxlanhdr *)(udp_hdr(skb) + 1);
1157
Tom Herbert0ace2ca2015-02-10 16:30:32 -08001158 skb_remcsum_process(skb, (void *)vh + hdrlen, start, offset,
1159 nopartial);
Tom Herbertdfd86452015-01-12 17:00:38 -08001160
1161 return vh;
1162}
1163
stephen hemmingerd3428942012-10-01 12:32:35 +00001164/* Callback from net/ipv4/udp.c to receive packets */
1165static int vxlan_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
1166{
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001167 struct vxlan_sock *vs;
stephen hemmingerd3428942012-10-01 12:32:35 +00001168 struct vxlanhdr *vxh;
Tom Herbert3bf39472015-01-08 12:31:18 -08001169 u32 flags, vni;
Thomas Graf35114942015-01-15 03:53:55 +01001170 struct vxlan_metadata md = {0};
stephen hemmingerd3428942012-10-01 12:32:35 +00001171
stephen hemmingerd3428942012-10-01 12:32:35 +00001172 /* Need Vxlan and inner Ethernet header to be present */
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001173 if (!pskb_may_pull(skb, VXLAN_HLEN))
stephen hemmingerd3428942012-10-01 12:32:35 +00001174 goto error;
1175
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001176 vxh = (struct vxlanhdr *)(udp_hdr(skb) + 1);
Tom Herbert3bf39472015-01-08 12:31:18 -08001177 flags = ntohl(vxh->vx_flags);
1178 vni = ntohl(vxh->vx_vni);
1179
1180 if (flags & VXLAN_HF_VNI) {
1181 flags &= ~VXLAN_HF_VNI;
1182 } else {
1183 /* VNI flag always required to be set */
1184 goto bad_flags;
stephen hemmingerd3428942012-10-01 12:32:35 +00001185 }
1186
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001187 if (iptunnel_pull_header(skb, VXLAN_HLEN, htons(ETH_P_TEB)))
stephen hemmingerd3428942012-10-01 12:32:35 +00001188 goto drop;
Tom Herbertdfd86452015-01-12 17:00:38 -08001189 vxh = (struct vxlanhdr *)(udp_hdr(skb) + 1);
stephen hemmingerd3428942012-10-01 12:32:35 +00001190
Pravin B Shelar559835ea2013-09-24 10:25:40 -07001191 vs = rcu_dereference_sk_user_data(sk);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001192 if (!vs)
stephen hemmingerd3428942012-10-01 12:32:35 +00001193 goto drop;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001194
Tom Herbertdfd86452015-01-12 17:00:38 -08001195 if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) {
Tom Herbert0ace2ca2015-02-10 16:30:32 -08001196 vxh = vxlan_remcsum(skb, vxh, sizeof(struct vxlanhdr), vni,
1197 !!(vs->flags & VXLAN_F_REMCSUM_NOPARTIAL));
Tom Herbertdfd86452015-01-12 17:00:38 -08001198 if (!vxh)
1199 goto drop;
1200
1201 flags &= ~VXLAN_HF_RCO;
Alexey Kodanev40fb70f2015-03-13 19:13:53 +03001202 vni &= VXLAN_VNI_MASK;
Tom Herbertdfd86452015-01-12 17:00:38 -08001203 }
1204
Thomas Graf35114942015-01-15 03:53:55 +01001205 /* For backwards compatibility, only allow reserved fields to be
1206 * used by VXLAN extensions if explicitly requested.
1207 */
1208 if ((flags & VXLAN_HF_GBP) && (vs->flags & VXLAN_F_GBP)) {
1209 struct vxlanhdr_gbp *gbp;
1210
1211 gbp = (struct vxlanhdr_gbp *)vxh;
1212 md.gbp = ntohs(gbp->policy_id);
1213
1214 if (gbp->dont_learn)
1215 md.gbp |= VXLAN_GBP_DONT_LEARN;
1216
1217 if (gbp->policy_applied)
1218 md.gbp |= VXLAN_GBP_POLICY_APPLIED;
1219
1220 flags &= ~VXLAN_GBP_USED_BITS;
1221 }
1222
Alexey Kodanev40fb70f2015-03-13 19:13:53 +03001223 if (flags || vni & ~VXLAN_VNI_MASK) {
Tom Herbert3bf39472015-01-08 12:31:18 -08001224 /* If there are any unprocessed flags remaining treat
1225 * this as a malformed packet. This behavior diverges from
1226 * VXLAN RFC (RFC7348) which stipulates that bits in reserved
1227 * in reserved fields are to be ignored. The approach here
Simon Hormanc4b49512015-04-02 11:17:58 +09001228 * maintains compatibility with previous stack code, and also
Tom Herbert3bf39472015-01-08 12:31:18 -08001229 * is more robust and provides a little more security in
1230 * adding extensions to VXLAN.
1231 */
1232
1233 goto bad_flags;
1234 }
1235
Thomas Graf35114942015-01-15 03:53:55 +01001236 md.vni = vxh->vx_vni;
1237 vs->rcv(vs, skb, &md);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001238 return 0;
1239
1240drop:
1241 /* Consume bad packet */
1242 kfree_skb(skb);
1243 return 0;
1244
Tom Herbert3bf39472015-01-08 12:31:18 -08001245bad_flags:
1246 netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n",
1247 ntohl(vxh->vx_flags), ntohl(vxh->vx_vni));
1248
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001249error:
1250 /* Return non vxlan pkt */
1251 return 1;
1252}
1253
Thomas Graf35114942015-01-15 03:53:55 +01001254static void vxlan_rcv(struct vxlan_sock *vs, struct sk_buff *skb,
1255 struct vxlan_metadata *md)
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001256{
Cong Wange4c7ed42013-08-31 13:44:33 +08001257 struct iphdr *oip = NULL;
1258 struct ipv6hdr *oip6 = NULL;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001259 struct vxlan_dev *vxlan;
Li RongQing8f849852014-01-04 13:57:59 +08001260 struct pcpu_sw_netstats *stats;
Cong Wange4c7ed42013-08-31 13:44:33 +08001261 union vxlan_addr saddr;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001262 __u32 vni;
Cong Wange4c7ed42013-08-31 13:44:33 +08001263 int err = 0;
1264 union vxlan_addr *remote_ip;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001265
Thomas Graf35114942015-01-15 03:53:55 +01001266 vni = ntohl(md->vni) >> 8;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001267 /* Is this VNI defined? */
1268 vxlan = vxlan_vs_find_vni(vs, vni);
1269 if (!vxlan)
1270 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001271
Cong Wange4c7ed42013-08-31 13:44:33 +08001272 remote_ip = &vxlan->default_dst.remote_ip;
David Stevense4f67ad2012-11-20 02:50:14 +00001273 skb_reset_mac_header(skb);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02001274 skb_scrub_packet(skb, !net_eq(vxlan->net, dev_net(vxlan->dev)));
stephen hemmingerd3428942012-10-01 12:32:35 +00001275 skb->protocol = eth_type_trans(skb, vxlan->dev);
Tom Herbertf79b0642014-06-14 23:24:36 -07001276 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
stephen hemmingerd3428942012-10-01 12:32:35 +00001277
1278 /* Ignore packet loops (and multicast echo) */
Joe Perches7367d0b2013-09-01 11:51:23 -07001279 if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +00001280 goto drop;
1281
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001282 /* Re-examine inner Ethernet packet */
Cong Wange4c7ed42013-08-31 13:44:33 +08001283 if (remote_ip->sa.sa_family == AF_INET) {
1284 oip = ip_hdr(skb);
1285 saddr.sin.sin_addr.s_addr = oip->saddr;
1286 saddr.sa.sa_family = AF_INET;
1287#if IS_ENABLED(CONFIG_IPV6)
1288 } else {
1289 oip6 = ipv6_hdr(skb);
1290 saddr.sin6.sin6_addr = oip6->saddr;
1291 saddr.sa.sa_family = AF_INET6;
1292#endif
1293 }
1294
stephen hemminger26a41ae62013-06-17 12:09:58 -07001295 if ((vxlan->flags & VXLAN_F_LEARN) &&
Cong Wange4c7ed42013-08-31 13:44:33 +08001296 vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source))
stephen hemminger26a41ae62013-06-17 12:09:58 -07001297 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001298
stephen hemmingerd3428942012-10-01 12:32:35 +00001299 skb_reset_network_header(skb);
Thomas Graf35114942015-01-15 03:53:55 +01001300 skb->mark = md->gbp;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00001301
Cong Wange4c7ed42013-08-31 13:44:33 +08001302 if (oip6)
1303 err = IP6_ECN_decapsulate(oip6, skb);
1304 if (oip)
1305 err = IP_ECN_decapsulate(oip, skb);
1306
stephen hemmingerd3428942012-10-01 12:32:35 +00001307 if (unlikely(err)) {
Cong Wange4c7ed42013-08-31 13:44:33 +08001308 if (log_ecn_error) {
1309 if (oip6)
1310 net_info_ratelimited("non-ECT from %pI6\n",
1311 &oip6->saddr);
1312 if (oip)
1313 net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n",
1314 &oip->saddr, oip->tos);
1315 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001316 if (err > 1) {
1317 ++vxlan->dev->stats.rx_frame_errors;
1318 ++vxlan->dev->stats.rx_errors;
1319 goto drop;
1320 }
1321 }
1322
Pravin B Shelare8171042013-03-25 14:49:46 +00001323 stats = this_cpu_ptr(vxlan->dev->tstats);
stephen hemmingerd3428942012-10-01 12:32:35 +00001324 u64_stats_update_begin(&stats->syncp);
1325 stats->rx_packets++;
1326 stats->rx_bytes += skb->len;
1327 u64_stats_update_end(&stats->syncp);
1328
1329 netif_rx(skb);
1330
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001331 return;
stephen hemmingerd3428942012-10-01 12:32:35 +00001332drop:
1333 /* Consume bad packet */
1334 kfree_skb(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001335}
1336
David Stevense4f67ad2012-11-20 02:50:14 +00001337static int arp_reduce(struct net_device *dev, struct sk_buff *skb)
1338{
1339 struct vxlan_dev *vxlan = netdev_priv(dev);
1340 struct arphdr *parp;
1341 u8 *arpptr, *sha;
1342 __be32 sip, tip;
1343 struct neighbour *n;
1344
1345 if (dev->flags & IFF_NOARP)
1346 goto out;
1347
1348 if (!pskb_may_pull(skb, arp_hdr_len(dev))) {
1349 dev->stats.tx_dropped++;
1350 goto out;
1351 }
1352 parp = arp_hdr(skb);
1353
1354 if ((parp->ar_hrd != htons(ARPHRD_ETHER) &&
1355 parp->ar_hrd != htons(ARPHRD_IEEE802)) ||
1356 parp->ar_pro != htons(ETH_P_IP) ||
1357 parp->ar_op != htons(ARPOP_REQUEST) ||
1358 parp->ar_hln != dev->addr_len ||
1359 parp->ar_pln != 4)
1360 goto out;
1361 arpptr = (u8 *)parp + sizeof(struct arphdr);
1362 sha = arpptr;
1363 arpptr += dev->addr_len; /* sha */
1364 memcpy(&sip, arpptr, sizeof(sip));
1365 arpptr += sizeof(sip);
1366 arpptr += dev->addr_len; /* tha */
1367 memcpy(&tip, arpptr, sizeof(tip));
1368
1369 if (ipv4_is_loopback(tip) ||
1370 ipv4_is_multicast(tip))
1371 goto out;
1372
1373 n = neigh_lookup(&arp_tbl, &tip, dev);
1374
1375 if (n) {
David Stevense4f67ad2012-11-20 02:50:14 +00001376 struct vxlan_fdb *f;
1377 struct sk_buff *reply;
1378
1379 if (!(n->nud_state & NUD_CONNECTED)) {
1380 neigh_release(n);
1381 goto out;
1382 }
1383
1384 f = vxlan_find_mac(vxlan, n->ha);
Cong Wange4c7ed42013-08-31 13:44:33 +08001385 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
David Stevense4f67ad2012-11-20 02:50:14 +00001386 /* bridge-local neighbor */
1387 neigh_release(n);
1388 goto out;
1389 }
1390
1391 reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
1392 n->ha, sha);
1393
1394 neigh_release(n);
1395
David Stevens73461352014-03-18 12:32:29 -04001396 if (reply == NULL)
1397 goto out;
1398
David Stevense4f67ad2012-11-20 02:50:14 +00001399 skb_reset_mac_header(reply);
1400 __skb_pull(reply, skb_network_offset(reply));
1401 reply->ip_summed = CHECKSUM_UNNECESSARY;
1402 reply->pkt_type = PACKET_HOST;
1403
1404 if (netif_rx_ni(reply) == NET_RX_DROP)
1405 dev->stats.rx_dropped++;
Cong Wange4c7ed42013-08-31 13:44:33 +08001406 } else if (vxlan->flags & VXLAN_F_L3MISS) {
1407 union vxlan_addr ipa = {
1408 .sin.sin_addr.s_addr = tip,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001409 .sin.sin_family = AF_INET,
Cong Wange4c7ed42013-08-31 13:44:33 +08001410 };
1411
1412 vxlan_ip_miss(dev, &ipa);
1413 }
David Stevense4f67ad2012-11-20 02:50:14 +00001414out:
1415 consume_skb(skb);
1416 return NETDEV_TX_OK;
1417}
1418
Cong Wangf564f452013-08-31 13:44:36 +08001419#if IS_ENABLED(CONFIG_IPV6)
David Stevens4b29dba2014-03-24 10:39:58 -04001420static struct sk_buff *vxlan_na_create(struct sk_buff *request,
1421 struct neighbour *n, bool isrouter)
1422{
1423 struct net_device *dev = request->dev;
1424 struct sk_buff *reply;
1425 struct nd_msg *ns, *na;
1426 struct ipv6hdr *pip6;
1427 u8 *daddr;
1428 int na_olen = 8; /* opt hdr + ETH_ALEN for target */
1429 int ns_olen;
1430 int i, len;
1431
1432 if (dev == NULL)
1433 return NULL;
1434
1435 len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) +
1436 sizeof(*na) + na_olen + dev->needed_tailroom;
1437 reply = alloc_skb(len, GFP_ATOMIC);
1438 if (reply == NULL)
1439 return NULL;
1440
1441 reply->protocol = htons(ETH_P_IPV6);
1442 reply->dev = dev;
1443 skb_reserve(reply, LL_RESERVED_SPACE(request->dev));
1444 skb_push(reply, sizeof(struct ethhdr));
1445 skb_set_mac_header(reply, 0);
1446
1447 ns = (struct nd_msg *)skb_transport_header(request);
1448
1449 daddr = eth_hdr(request)->h_source;
1450 ns_olen = request->len - skb_transport_offset(request) - sizeof(*ns);
1451 for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) {
1452 if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) {
1453 daddr = ns->opt + i + sizeof(struct nd_opt_hdr);
1454 break;
1455 }
1456 }
1457
1458 /* Ethernet header */
1459 ether_addr_copy(eth_hdr(reply)->h_dest, daddr);
1460 ether_addr_copy(eth_hdr(reply)->h_source, n->ha);
1461 eth_hdr(reply)->h_proto = htons(ETH_P_IPV6);
1462 reply->protocol = htons(ETH_P_IPV6);
1463
1464 skb_pull(reply, sizeof(struct ethhdr));
1465 skb_set_network_header(reply, 0);
1466 skb_put(reply, sizeof(struct ipv6hdr));
1467
1468 /* IPv6 header */
1469
1470 pip6 = ipv6_hdr(reply);
1471 memset(pip6, 0, sizeof(struct ipv6hdr));
1472 pip6->version = 6;
1473 pip6->priority = ipv6_hdr(request)->priority;
1474 pip6->nexthdr = IPPROTO_ICMPV6;
1475 pip6->hop_limit = 255;
1476 pip6->daddr = ipv6_hdr(request)->saddr;
1477 pip6->saddr = *(struct in6_addr *)n->primary_key;
1478
1479 skb_pull(reply, sizeof(struct ipv6hdr));
1480 skb_set_transport_header(reply, 0);
1481
1482 na = (struct nd_msg *)skb_put(reply, sizeof(*na) + na_olen);
1483
1484 /* Neighbor Advertisement */
1485 memset(na, 0, sizeof(*na)+na_olen);
1486 na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT;
1487 na->icmph.icmp6_router = isrouter;
1488 na->icmph.icmp6_override = 1;
1489 na->icmph.icmp6_solicited = 1;
1490 na->target = ns->target;
1491 ether_addr_copy(&na->opt[2], n->ha);
1492 na->opt[0] = ND_OPT_TARGET_LL_ADDR;
1493 na->opt[1] = na_olen >> 3;
1494
1495 na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr,
1496 &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6,
1497 csum_partial(na, sizeof(*na)+na_olen, 0));
1498
1499 pip6->payload_len = htons(sizeof(*na)+na_olen);
1500
1501 skb_push(reply, sizeof(struct ipv6hdr));
1502
1503 reply->ip_summed = CHECKSUM_UNNECESSARY;
1504
1505 return reply;
1506}
1507
Cong Wangf564f452013-08-31 13:44:36 +08001508static int neigh_reduce(struct net_device *dev, struct sk_buff *skb)
1509{
1510 struct vxlan_dev *vxlan = netdev_priv(dev);
David Stevens4b29dba2014-03-24 10:39:58 -04001511 struct nd_msg *msg;
Cong Wangf564f452013-08-31 13:44:36 +08001512 const struct ipv6hdr *iphdr;
1513 const struct in6_addr *saddr, *daddr;
David Stevens4b29dba2014-03-24 10:39:58 -04001514 struct neighbour *n;
1515 struct inet6_dev *in6_dev;
Cong Wangf564f452013-08-31 13:44:36 +08001516
1517 in6_dev = __in6_dev_get(dev);
1518 if (!in6_dev)
1519 goto out;
1520
Cong Wangf564f452013-08-31 13:44:36 +08001521 iphdr = ipv6_hdr(skb);
1522 saddr = &iphdr->saddr;
1523 daddr = &iphdr->daddr;
1524
Cong Wangf564f452013-08-31 13:44:36 +08001525 msg = (struct nd_msg *)skb_transport_header(skb);
1526 if (msg->icmph.icmp6_code != 0 ||
1527 msg->icmph.icmp6_type != NDISC_NEIGHBOUR_SOLICITATION)
1528 goto out;
1529
David Stevens4b29dba2014-03-24 10:39:58 -04001530 if (ipv6_addr_loopback(daddr) ||
1531 ipv6_addr_is_multicast(&msg->target))
1532 goto out;
1533
1534 n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev);
Cong Wangf564f452013-08-31 13:44:36 +08001535
1536 if (n) {
1537 struct vxlan_fdb *f;
David Stevens4b29dba2014-03-24 10:39:58 -04001538 struct sk_buff *reply;
Cong Wangf564f452013-08-31 13:44:36 +08001539
1540 if (!(n->nud_state & NUD_CONNECTED)) {
1541 neigh_release(n);
1542 goto out;
1543 }
1544
1545 f = vxlan_find_mac(vxlan, n->ha);
1546 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
1547 /* bridge-local neighbor */
1548 neigh_release(n);
1549 goto out;
1550 }
1551
David Stevens4b29dba2014-03-24 10:39:58 -04001552 reply = vxlan_na_create(skb, n,
1553 !!(f ? f->flags & NTF_ROUTER : 0));
1554
Cong Wangf564f452013-08-31 13:44:36 +08001555 neigh_release(n);
David Stevens4b29dba2014-03-24 10:39:58 -04001556
1557 if (reply == NULL)
1558 goto out;
1559
1560 if (netif_rx_ni(reply) == NET_RX_DROP)
1561 dev->stats.rx_dropped++;
1562
Cong Wangf564f452013-08-31 13:44:36 +08001563 } else if (vxlan->flags & VXLAN_F_L3MISS) {
David Stevens4b29dba2014-03-24 10:39:58 -04001564 union vxlan_addr ipa = {
1565 .sin6.sin6_addr = msg->target,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001566 .sin6.sin6_family = AF_INET6,
David Stevens4b29dba2014-03-24 10:39:58 -04001567 };
1568
Cong Wangf564f452013-08-31 13:44:36 +08001569 vxlan_ip_miss(dev, &ipa);
1570 }
1571
1572out:
1573 consume_skb(skb);
1574 return NETDEV_TX_OK;
1575}
1576#endif
1577
David Stevense4f67ad2012-11-20 02:50:14 +00001578static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb)
1579{
1580 struct vxlan_dev *vxlan = netdev_priv(dev);
1581 struct neighbour *n;
David Stevense4f67ad2012-11-20 02:50:14 +00001582
1583 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
1584 return false;
1585
1586 n = NULL;
1587 switch (ntohs(eth_hdr(skb)->h_proto)) {
1588 case ETH_P_IP:
Cong Wange15a00a2013-08-31 13:44:34 +08001589 {
1590 struct iphdr *pip;
1591
David Stevense4f67ad2012-11-20 02:50:14 +00001592 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
1593 return false;
1594 pip = ip_hdr(skb);
1595 n = neigh_lookup(&arp_tbl, &pip->daddr, dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08001596 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1597 union vxlan_addr ipa = {
1598 .sin.sin_addr.s_addr = pip->daddr,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001599 .sin.sin_family = AF_INET,
Cong Wange4c7ed42013-08-31 13:44:33 +08001600 };
1601
1602 vxlan_ip_miss(dev, &ipa);
1603 return false;
1604 }
1605
David Stevense4f67ad2012-11-20 02:50:14 +00001606 break;
Cong Wange15a00a2013-08-31 13:44:34 +08001607 }
1608#if IS_ENABLED(CONFIG_IPV6)
1609 case ETH_P_IPV6:
1610 {
1611 struct ipv6hdr *pip6;
1612
1613 if (!pskb_may_pull(skb, sizeof(struct ipv6hdr)))
1614 return false;
1615 pip6 = ipv6_hdr(skb);
1616 n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev);
1617 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1618 union vxlan_addr ipa = {
1619 .sin6.sin6_addr = pip6->daddr,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001620 .sin6.sin6_family = AF_INET6,
Cong Wange15a00a2013-08-31 13:44:34 +08001621 };
1622
1623 vxlan_ip_miss(dev, &ipa);
1624 return false;
1625 }
1626
1627 break;
1628 }
1629#endif
David Stevense4f67ad2012-11-20 02:50:14 +00001630 default:
1631 return false;
1632 }
1633
1634 if (n) {
1635 bool diff;
1636
Joe Perches7367d0b2013-09-01 11:51:23 -07001637 diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha);
David Stevense4f67ad2012-11-20 02:50:14 +00001638 if (diff) {
1639 memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest,
1640 dev->addr_len);
1641 memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len);
1642 }
1643 neigh_release(n);
1644 return diff;
Cong Wange4c7ed42013-08-31 13:44:33 +08001645 }
1646
David Stevense4f67ad2012-11-20 02:50:14 +00001647 return false;
1648}
1649
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001650static void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, u32 vxflags,
Thomas Graf35114942015-01-15 03:53:55 +01001651 struct vxlan_metadata *md)
1652{
1653 struct vxlanhdr_gbp *gbp;
1654
Thomas Grafdb79a622015-02-04 17:00:04 +01001655 if (!md->gbp)
1656 return;
1657
Thomas Graf35114942015-01-15 03:53:55 +01001658 gbp = (struct vxlanhdr_gbp *)vxh;
1659 vxh->vx_flags |= htonl(VXLAN_HF_GBP);
1660
1661 if (md->gbp & VXLAN_GBP_DONT_LEARN)
1662 gbp->dont_learn = 1;
1663
1664 if (md->gbp & VXLAN_GBP_POLICY_APPLIED)
1665 gbp->policy_applied = 1;
1666
1667 gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK);
1668}
1669
Cong Wange4c7ed42013-08-31 13:44:33 +08001670#if IS_ENABLED(CONFIG_IPV6)
David Miller79b16aa2015-04-05 22:19:09 -04001671static int vxlan6_xmit_skb(struct dst_entry *dst, struct sock *sk,
1672 struct sk_buff *skb,
Cong Wange4c7ed42013-08-31 13:44:33 +08001673 struct net_device *dev, struct in6_addr *saddr,
1674 struct in6_addr *daddr, __u8 prio, __u8 ttl,
Thomas Graf35114942015-01-15 03:53:55 +01001675 __be16 src_port, __be16 dst_port,
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001676 struct vxlan_metadata *md, bool xnet, u32 vxflags)
Cong Wange4c7ed42013-08-31 13:44:33 +08001677{
Cong Wange4c7ed42013-08-31 13:44:33 +08001678 struct vxlanhdr *vxh;
Cong Wange4c7ed42013-08-31 13:44:33 +08001679 int min_headroom;
1680 int err;
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001681 bool udp_sum = !(vxflags & VXLAN_F_UDP_ZERO_CSUM6_TX);
Tom Herbertdfd86452015-01-12 17:00:38 -08001682 int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL;
1683 u16 hdrlen = sizeof(struct vxlanhdr);
Cong Wange4c7ed42013-08-31 13:44:33 +08001684
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001685 if ((vxflags & VXLAN_F_REMCSUM_TX) &&
Tom Herbertdfd86452015-01-12 17:00:38 -08001686 skb->ip_summed == CHECKSUM_PARTIAL) {
1687 int csum_start = skb_checksum_start_offset(skb);
1688
1689 if (csum_start <= VXLAN_MAX_REMCSUM_START &&
1690 !(csum_start & VXLAN_RCO_SHIFT_MASK) &&
1691 (skb->csum_offset == offsetof(struct udphdr, check) ||
1692 skb->csum_offset == offsetof(struct tcphdr, check))) {
1693 udp_sum = false;
1694 type |= SKB_GSO_TUNNEL_REMCSUM;
1695 }
1696 }
1697
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02001698 skb_scrub_packet(skb, xnet);
Nicolas Dichtel963a88b2013-09-02 15:34:57 +02001699
Cong Wange4c7ed42013-08-31 13:44:33 +08001700 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
1701 + VXLAN_HLEN + sizeof(struct ipv6hdr)
Jiri Pirkodf8a39d2015-01-13 17:13:44 +01001702 + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
Cong Wange4c7ed42013-08-31 13:44:33 +08001703
1704 /* Need space for new headers (invalidates iph ptr) */
1705 err = skb_cow_head(skb, min_headroom);
Pravin B Shelar74f47272014-12-23 16:20:36 -08001706 if (unlikely(err)) {
1707 kfree_skb(skb);
1708 goto err;
1709 }
Cong Wange4c7ed42013-08-31 13:44:33 +08001710
Jiri Pirko59682502014-11-19 14:04:59 +01001711 skb = vlan_hwaccel_push_inside(skb);
Pravin B Shelar74f47272014-12-23 16:20:36 -08001712 if (WARN_ON(!skb)) {
1713 err = -ENOMEM;
1714 goto err;
1715 }
Cong Wange4c7ed42013-08-31 13:44:33 +08001716
Jesse Grossb736a622015-04-09 11:19:14 -07001717 skb = iptunnel_handle_offloads(skb, udp_sum, type);
1718 if (IS_ERR(skb)) {
1719 err = -EINVAL;
1720 goto err;
1721 }
1722
Cong Wange4c7ed42013-08-31 13:44:33 +08001723 vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh));
Tom Herbert3bf39472015-01-08 12:31:18 -08001724 vxh->vx_flags = htonl(VXLAN_HF_VNI);
Thomas Graf35114942015-01-15 03:53:55 +01001725 vxh->vx_vni = md->vni;
Cong Wange4c7ed42013-08-31 13:44:33 +08001726
Tom Herbertdfd86452015-01-12 17:00:38 -08001727 if (type & SKB_GSO_TUNNEL_REMCSUM) {
1728 u32 data = (skb_checksum_start_offset(skb) - hdrlen) >>
1729 VXLAN_RCO_SHIFT;
1730
1731 if (skb->csum_offset == offsetof(struct udphdr, check))
1732 data |= VXLAN_RCO_UDP;
1733
1734 vxh->vx_vni |= htonl(data);
1735 vxh->vx_flags |= htonl(VXLAN_HF_RCO);
1736
1737 if (!skb_is_gso(skb)) {
1738 skb->ip_summed = CHECKSUM_NONE;
1739 skb->encapsulation = 0;
1740 }
1741 }
1742
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001743 if (vxflags & VXLAN_F_GBP)
1744 vxlan_build_gbp_hdr(vxh, vxflags, md);
Thomas Graf35114942015-01-15 03:53:55 +01001745
Tom Herbert996c9fd2014-09-29 20:22:33 -07001746 skb_set_inner_protocol(skb, htons(ETH_P_TEB));
1747
David Miller79b16aa2015-04-05 22:19:09 -04001748 udp_tunnel6_xmit_skb(dst, sk, skb, dev, saddr, daddr, prio,
Tom Herbertd998f8e2015-01-20 11:23:04 -08001749 ttl, src_port, dst_port,
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001750 !!(vxflags & VXLAN_F_UDP_ZERO_CSUM6_TX));
Cong Wange4c7ed42013-08-31 13:44:33 +08001751 return 0;
Pravin B Shelar74f47272014-12-23 16:20:36 -08001752err:
1753 dst_release(dst);
1754 return err;
Cong Wange4c7ed42013-08-31 13:44:33 +08001755}
1756#endif
1757
David Miller79b16aa2015-04-05 22:19:09 -04001758int vxlan_xmit_skb(struct rtable *rt, struct sock *sk, struct sk_buff *skb,
Pravin B Shelar49560532013-08-19 11:23:17 -07001759 __be32 src, __be32 dst, __u8 tos, __u8 ttl, __be16 df,
Thomas Graf35114942015-01-15 03:53:55 +01001760 __be16 src_port, __be16 dst_port,
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001761 struct vxlan_metadata *md, bool xnet, u32 vxflags)
Pravin B Shelar49560532013-08-19 11:23:17 -07001762{
1763 struct vxlanhdr *vxh;
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001764 int min_headroom;
Pravin B Shelar49560532013-08-19 11:23:17 -07001765 int err;
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001766 bool udp_sum = !!(vxflags & VXLAN_F_UDP_CSUM);
Tom Herbertdfd86452015-01-12 17:00:38 -08001767 int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL;
1768 u16 hdrlen = sizeof(struct vxlanhdr);
Pravin B Shelar49560532013-08-19 11:23:17 -07001769
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001770 if ((vxflags & VXLAN_F_REMCSUM_TX) &&
Tom Herbertdfd86452015-01-12 17:00:38 -08001771 skb->ip_summed == CHECKSUM_PARTIAL) {
1772 int csum_start = skb_checksum_start_offset(skb);
1773
1774 if (csum_start <= VXLAN_MAX_REMCSUM_START &&
1775 !(csum_start & VXLAN_RCO_SHIFT_MASK) &&
1776 (skb->csum_offset == offsetof(struct udphdr, check) ||
1777 skb->csum_offset == offsetof(struct tcphdr, check))) {
1778 udp_sum = false;
1779 type |= SKB_GSO_TUNNEL_REMCSUM;
1780 }
1781 }
1782
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001783 min_headroom = LL_RESERVED_SPACE(rt->dst.dev) + rt->dst.header_len
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07001784 + VXLAN_HLEN + sizeof(struct iphdr)
Jiri Pirkodf8a39d2015-01-13 17:13:44 +01001785 + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001786
1787 /* Need space for new headers (invalidates iph ptr) */
1788 err = skb_cow_head(skb, min_headroom);
Pravin B Shelar74f47272014-12-23 16:20:36 -08001789 if (unlikely(err)) {
1790 kfree_skb(skb);
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001791 return err;
Pravin B Shelar74f47272014-12-23 16:20:36 -08001792 }
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001793
Jiri Pirko59682502014-11-19 14:04:59 +01001794 skb = vlan_hwaccel_push_inside(skb);
1795 if (WARN_ON(!skb))
1796 return -ENOMEM;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07001797
Jesse Grossb736a622015-04-09 11:19:14 -07001798 skb = iptunnel_handle_offloads(skb, udp_sum, type);
1799 if (IS_ERR(skb))
1800 return PTR_ERR(skb);
1801
Pravin B Shelar49560532013-08-19 11:23:17 -07001802 vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh));
Tom Herbert3bf39472015-01-08 12:31:18 -08001803 vxh->vx_flags = htonl(VXLAN_HF_VNI);
Thomas Graf35114942015-01-15 03:53:55 +01001804 vxh->vx_vni = md->vni;
Pravin B Shelar49560532013-08-19 11:23:17 -07001805
Tom Herbertdfd86452015-01-12 17:00:38 -08001806 if (type & SKB_GSO_TUNNEL_REMCSUM) {
1807 u32 data = (skb_checksum_start_offset(skb) - hdrlen) >>
1808 VXLAN_RCO_SHIFT;
1809
1810 if (skb->csum_offset == offsetof(struct udphdr, check))
1811 data |= VXLAN_RCO_UDP;
1812
1813 vxh->vx_vni |= htonl(data);
1814 vxh->vx_flags |= htonl(VXLAN_HF_RCO);
1815
1816 if (!skb_is_gso(skb)) {
1817 skb->ip_summed = CHECKSUM_NONE;
1818 skb->encapsulation = 0;
1819 }
1820 }
1821
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001822 if (vxflags & VXLAN_F_GBP)
1823 vxlan_build_gbp_hdr(vxh, vxflags, md);
Thomas Graf35114942015-01-15 03:53:55 +01001824
Tom Herbert996c9fd2014-09-29 20:22:33 -07001825 skb_set_inner_protocol(skb, htons(ETH_P_TEB));
1826
David Miller79b16aa2015-04-05 22:19:09 -04001827 return udp_tunnel_xmit_skb(rt, sk, skb, src, dst, tos,
Tom Herbertd998f8e2015-01-20 11:23:04 -08001828 ttl, df, src_port, dst_port, xnet,
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001829 !(vxflags & VXLAN_F_UDP_CSUM));
Pravin B Shelar49560532013-08-19 11:23:17 -07001830}
1831EXPORT_SYMBOL_GPL(vxlan_xmit_skb);
1832
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001833/* Bypass encapsulation if the destination is local */
1834static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan,
1835 struct vxlan_dev *dst_vxlan)
1836{
Li RongQing8f849852014-01-04 13:57:59 +08001837 struct pcpu_sw_netstats *tx_stats, *rx_stats;
Cong Wange4c7ed42013-08-31 13:44:33 +08001838 union vxlan_addr loopback;
1839 union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip;
Li RongQingce6502a2014-10-16 08:49:41 +08001840 struct net_device *dev = skb->dev;
1841 int len = skb->len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001842
Li RongQing8f849852014-01-04 13:57:59 +08001843 tx_stats = this_cpu_ptr(src_vxlan->dev->tstats);
1844 rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001845 skb->pkt_type = PACKET_HOST;
1846 skb->encapsulation = 0;
1847 skb->dev = dst_vxlan->dev;
1848 __skb_pull(skb, skb_network_offset(skb));
1849
Cong Wange4c7ed42013-08-31 13:44:33 +08001850 if (remote_ip->sa.sa_family == AF_INET) {
1851 loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1852 loopback.sa.sa_family = AF_INET;
1853#if IS_ENABLED(CONFIG_IPV6)
1854 } else {
1855 loopback.sin6.sin6_addr = in6addr_loopback;
1856 loopback.sa.sa_family = AF_INET6;
1857#endif
1858 }
1859
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001860 if (dst_vxlan->flags & VXLAN_F_LEARN)
Cong Wange4c7ed42013-08-31 13:44:33 +08001861 vxlan_snoop(skb->dev, &loopback, eth_hdr(skb)->h_source);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001862
1863 u64_stats_update_begin(&tx_stats->syncp);
1864 tx_stats->tx_packets++;
Li RongQingce6502a2014-10-16 08:49:41 +08001865 tx_stats->tx_bytes += len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001866 u64_stats_update_end(&tx_stats->syncp);
1867
1868 if (netif_rx(skb) == NET_RX_SUCCESS) {
1869 u64_stats_update_begin(&rx_stats->syncp);
1870 rx_stats->rx_packets++;
Li RongQingce6502a2014-10-16 08:49:41 +08001871 rx_stats->rx_bytes += len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001872 u64_stats_update_end(&rx_stats->syncp);
1873 } else {
Li RongQingce6502a2014-10-16 08:49:41 +08001874 dev->stats.rx_dropped++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001875 }
1876}
1877
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001878static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev,
1879 struct vxlan_rdst *rdst, bool did_rsc)
stephen hemmingerd3428942012-10-01 12:32:35 +00001880{
1881 struct vxlan_dev *vxlan = netdev_priv(dev);
David Miller79b16aa2015-04-05 22:19:09 -04001882 struct sock *sk = vxlan->vn_sock->sock->sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001883 struct rtable *rt = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001884 const struct iphdr *old_iph;
stephen hemmingerd3428942012-10-01 12:32:35 +00001885 struct flowi4 fl4;
Cong Wange4c7ed42013-08-31 13:44:33 +08001886 union vxlan_addr *dst;
Thomas Graf35114942015-01-15 03:53:55 +01001887 struct vxlan_metadata md;
Cong Wange4c7ed42013-08-31 13:44:33 +08001888 __be16 src_port = 0, dst_port;
Stephen Hemminger234f5b72013-06-17 14:16:41 -07001889 u32 vni;
stephen hemmingerd3428942012-10-01 12:32:35 +00001890 __be16 df = 0;
1891 __u8 tos, ttl;
Pravin B Shelar0e6fbc52013-06-17 17:49:56 -07001892 int err;
stephen hemmingerd3428942012-10-01 12:32:35 +00001893
stephen hemminger823aa872013-04-27 11:31:57 +00001894 dst_port = rdst->remote_port ? rdst->remote_port : vxlan->dst_port;
David Stevens66817122013-03-15 04:35:51 +00001895 vni = rdst->remote_vni;
Cong Wange4c7ed42013-08-31 13:44:33 +08001896 dst = &rdst->remote_ip;
David Stevense4f67ad2012-11-20 02:50:14 +00001897
Cong Wange4c7ed42013-08-31 13:44:33 +08001898 if (vxlan_addr_any(dst)) {
David Stevense4f67ad2012-11-20 02:50:14 +00001899 if (did_rsc) {
David Stevense4f67ad2012-11-20 02:50:14 +00001900 /* short-circuited back to local bridge */
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001901 vxlan_encap_bypass(skb, vxlan, vxlan);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001902 return;
David Stevense4f67ad2012-11-20 02:50:14 +00001903 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001904 goto drop;
David Stevense4f67ad2012-11-20 02:50:14 +00001905 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001906
stephen hemmingerd3428942012-10-01 12:32:35 +00001907 old_iph = ip_hdr(skb);
1908
stephen hemmingerd3428942012-10-01 12:32:35 +00001909 ttl = vxlan->ttl;
Cong Wange4c7ed42013-08-31 13:44:33 +08001910 if (!ttl && vxlan_addr_multicast(dst))
stephen hemmingerd3428942012-10-01 12:32:35 +00001911 ttl = 1;
1912
1913 tos = vxlan->tos;
1914 if (tos == 1)
Pravin B Shelar206aaaf2013-03-25 14:49:53 +00001915 tos = ip_tunnel_get_dsfield(old_iph, skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001916
Tom Herbert535fb8d2014-07-01 21:32:49 -07001917 src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->port_min,
1918 vxlan->port_max, true);
stephen hemmingerd3428942012-10-01 12:32:35 +00001919
Cong Wange4c7ed42013-08-31 13:44:33 +08001920 if (dst->sa.sa_family == AF_INET) {
1921 memset(&fl4, 0, sizeof(fl4));
1922 fl4.flowi4_oif = rdst->remote_ifindex;
1923 fl4.flowi4_tos = RT_TOS(tos);
Thomas Graf239fb792015-05-05 15:09:21 +02001924 fl4.flowi4_mark = skb->mark;
1925 fl4.flowi4_proto = IPPROTO_UDP;
Cong Wange4c7ed42013-08-31 13:44:33 +08001926 fl4.daddr = dst->sin.sin_addr.s_addr;
1927 fl4.saddr = vxlan->saddr.sin.sin_addr.s_addr;
stephen hemmingerca78f182012-10-09 20:35:48 +00001928
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02001929 rt = ip_route_output_key(vxlan->net, &fl4);
Cong Wange4c7ed42013-08-31 13:44:33 +08001930 if (IS_ERR(rt)) {
1931 netdev_dbg(dev, "no route to %pI4\n",
1932 &dst->sin.sin_addr.s_addr);
1933 dev->stats.tx_carrier_errors++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001934 goto tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08001935 }
1936
1937 if (rt->dst.dev == dev) {
1938 netdev_dbg(dev, "circular route to %pI4\n",
1939 &dst->sin.sin_addr.s_addr);
1940 dev->stats.collisions++;
Fan Dufffc15a2013-12-09 10:33:53 +08001941 goto rt_tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08001942 }
1943
1944 /* Bypass encapsulation if the destination is local */
1945 if (rt->rt_flags & RTCF_LOCAL &&
1946 !(rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
1947 struct vxlan_dev *dst_vxlan;
1948
1949 ip_rt_put(rt);
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02001950 dst_vxlan = vxlan_find_vni(vxlan->net, vni,
Thomas Grafac5132d2015-01-15 03:53:56 +01001951 dst->sa.sa_family, dst_port,
1952 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08001953 if (!dst_vxlan)
1954 goto tx_error;
1955 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
1956 return;
1957 }
1958
1959 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
1960 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
Thomas Graf35114942015-01-15 03:53:55 +01001961 md.vni = htonl(vni << 8);
1962 md.gbp = skb->mark;
Cong Wange4c7ed42013-08-31 13:44:33 +08001963
David Miller79b16aa2015-04-05 22:19:09 -04001964 err = vxlan_xmit_skb(rt, sk, skb, fl4.saddr,
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001965 dst->sin.sin_addr.s_addr, tos, ttl, df,
1966 src_port, dst_port, &md,
1967 !net_eq(vxlan->net, dev_net(vxlan->dev)),
1968 vxlan->flags);
Pravin B Shelar74f47272014-12-23 16:20:36 -08001969 if (err < 0) {
1970 /* skb is already freed. */
1971 skb = NULL;
Cong Wange4c7ed42013-08-31 13:44:33 +08001972 goto rt_tx_error;
Pravin B Shelar74f47272014-12-23 16:20:36 -08001973 }
1974
Cong Wange4c7ed42013-08-31 13:44:33 +08001975 iptunnel_xmit_stats(err, &dev->stats, dev->tstats);
1976#if IS_ENABLED(CONFIG_IPV6)
1977 } else {
Cong Wange4c7ed42013-08-31 13:44:33 +08001978 struct dst_entry *ndst;
1979 struct flowi6 fl6;
1980 u32 flags;
1981
1982 memset(&fl6, 0, sizeof(fl6));
1983 fl6.flowi6_oif = rdst->remote_ifindex;
1984 fl6.daddr = dst->sin6.sin6_addr;
1985 fl6.saddr = vxlan->saddr.sin6.sin6_addr;
Thomas Graf239fb792015-05-05 15:09:21 +02001986 fl6.flowi6_mark = skb->mark;
Cong Wang8c1bb792013-09-02 10:06:51 +08001987 fl6.flowi6_proto = IPPROTO_UDP;
Cong Wange4c7ed42013-08-31 13:44:33 +08001988
1989 if (ipv6_stub->ipv6_dst_lookup(sk, &ndst, &fl6)) {
1990 netdev_dbg(dev, "no route to %pI6\n",
1991 &dst->sin6.sin6_addr);
1992 dev->stats.tx_carrier_errors++;
1993 goto tx_error;
1994 }
1995
1996 if (ndst->dev == dev) {
1997 netdev_dbg(dev, "circular route to %pI6\n",
1998 &dst->sin6.sin6_addr);
1999 dst_release(ndst);
2000 dev->stats.collisions++;
2001 goto tx_error;
2002 }
2003
2004 /* Bypass encapsulation if the destination is local */
2005 flags = ((struct rt6_info *)ndst)->rt6i_flags;
2006 if (flags & RTF_LOCAL &&
2007 !(flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
2008 struct vxlan_dev *dst_vxlan;
2009
2010 dst_release(ndst);
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02002011 dst_vxlan = vxlan_find_vni(vxlan->net, vni,
Thomas Grafac5132d2015-01-15 03:53:56 +01002012 dst->sa.sa_family, dst_port,
2013 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08002014 if (!dst_vxlan)
2015 goto tx_error;
2016 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
2017 return;
2018 }
2019
2020 ttl = ttl ? : ip6_dst_hoplimit(ndst);
Thomas Graf35114942015-01-15 03:53:55 +01002021 md.vni = htonl(vni << 8);
2022 md.gbp = skb->mark;
Cong Wange4c7ed42013-08-31 13:44:33 +08002023
David Miller79b16aa2015-04-05 22:19:09 -04002024 err = vxlan6_xmit_skb(ndst, sk, skb, dev, &fl6.saddr, &fl6.daddr,
Tom Herbertaf33c1a2015-01-20 11:23:05 -08002025 0, ttl, src_port, dst_port, &md,
2026 !net_eq(vxlan->net, dev_net(vxlan->dev)),
2027 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08002028#endif
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00002029 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002030
Stephen Hemminger4ad16932013-06-17 14:16:11 -07002031 return;
stephen hemmingerd3428942012-10-01 12:32:35 +00002032
2033drop:
2034 dev->stats.tx_dropped++;
2035 goto tx_free;
2036
Pravin B Shelar49560532013-08-19 11:23:17 -07002037rt_tx_error:
2038 ip_rt_put(rt);
stephen hemmingerd3428942012-10-01 12:32:35 +00002039tx_error:
2040 dev->stats.tx_errors++;
2041tx_free:
2042 dev_kfree_skb(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00002043}
2044
David Stevens66817122013-03-15 04:35:51 +00002045/* Transmit local packets over Vxlan
2046 *
2047 * Outer IP header inherits ECN and DF from inner header.
2048 * Outer UDP destination is the VXLAN assigned port.
2049 * source port is based on hash of flow
2050 */
2051static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev)
2052{
2053 struct vxlan_dev *vxlan = netdev_priv(dev);
2054 struct ethhdr *eth;
2055 bool did_rsc = false;
Eric Dumazet8f646c92014-01-06 09:54:31 -08002056 struct vxlan_rdst *rdst, *fdst = NULL;
David Stevens66817122013-03-15 04:35:51 +00002057 struct vxlan_fdb *f;
David Stevens66817122013-03-15 04:35:51 +00002058
2059 skb_reset_mac_header(skb);
2060 eth = eth_hdr(skb);
2061
Cong Wangf564f452013-08-31 13:44:36 +08002062 if ((vxlan->flags & VXLAN_F_PROXY)) {
2063 if (ntohs(eth->h_proto) == ETH_P_ARP)
2064 return arp_reduce(dev, skb);
2065#if IS_ENABLED(CONFIG_IPV6)
2066 else if (ntohs(eth->h_proto) == ETH_P_IPV6 &&
Li RongQing91269e32014-10-16 09:17:18 +08002067 pskb_may_pull(skb, sizeof(struct ipv6hdr)
2068 + sizeof(struct nd_msg)) &&
Cong Wangf564f452013-08-31 13:44:36 +08002069 ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) {
2070 struct nd_msg *msg;
2071
2072 msg = (struct nd_msg *)skb_transport_header(skb);
2073 if (msg->icmph.icmp6_code == 0 &&
2074 msg->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION)
2075 return neigh_reduce(dev, skb);
2076 }
Li RongQing7a9f5262014-10-17 14:06:16 +08002077 eth = eth_hdr(skb);
Cong Wangf564f452013-08-31 13:44:36 +08002078#endif
2079 }
David Stevens66817122013-03-15 04:35:51 +00002080
2081 f = vxlan_find_mac(vxlan, eth->h_dest);
David Stevensae884082013-04-19 00:36:26 +00002082 did_rsc = false;
2083
2084 if (f && (f->flags & NTF_ROUTER) && (vxlan->flags & VXLAN_F_RSC) &&
Cong Wange15a00a2013-08-31 13:44:34 +08002085 (ntohs(eth->h_proto) == ETH_P_IP ||
2086 ntohs(eth->h_proto) == ETH_P_IPV6)) {
David Stevensae884082013-04-19 00:36:26 +00002087 did_rsc = route_shortcircuit(dev, skb);
2088 if (did_rsc)
2089 f = vxlan_find_mac(vxlan, eth->h_dest);
2090 }
2091
David Stevens66817122013-03-15 04:35:51 +00002092 if (f == NULL) {
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002093 f = vxlan_find_mac(vxlan, all_zeros_mac);
2094 if (f == NULL) {
2095 if ((vxlan->flags & VXLAN_F_L2MISS) &&
2096 !is_multicast_ether_addr(eth->h_dest))
2097 vxlan_fdb_miss(vxlan, eth->h_dest);
David Stevens66817122013-03-15 04:35:51 +00002098
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002099 dev->stats.tx_dropped++;
Eric Dumazet8f646c92014-01-06 09:54:31 -08002100 kfree_skb(skb);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002101 return NETDEV_TX_OK;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -07002102 }
David Stevens66817122013-03-15 04:35:51 +00002103 }
2104
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002105 list_for_each_entry_rcu(rdst, &f->remotes, list) {
2106 struct sk_buff *skb1;
2107
Eric Dumazet8f646c92014-01-06 09:54:31 -08002108 if (!fdst) {
2109 fdst = rdst;
2110 continue;
2111 }
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002112 skb1 = skb_clone(skb, GFP_ATOMIC);
2113 if (skb1)
2114 vxlan_xmit_one(skb1, dev, rdst, did_rsc);
2115 }
2116
Eric Dumazet8f646c92014-01-06 09:54:31 -08002117 if (fdst)
2118 vxlan_xmit_one(skb, dev, fdst, did_rsc);
2119 else
2120 kfree_skb(skb);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07002121 return NETDEV_TX_OK;
David Stevens66817122013-03-15 04:35:51 +00002122}
2123
stephen hemmingerd3428942012-10-01 12:32:35 +00002124/* Walk the forwarding table and purge stale entries */
2125static void vxlan_cleanup(unsigned long arg)
2126{
2127 struct vxlan_dev *vxlan = (struct vxlan_dev *) arg;
2128 unsigned long next_timer = jiffies + FDB_AGE_INTERVAL;
2129 unsigned int h;
2130
2131 if (!netif_running(vxlan->dev))
2132 return;
2133
stephen hemmingerd3428942012-10-01 12:32:35 +00002134 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2135 struct hlist_node *p, *n;
Sorin Dumitru14e1d0f2015-05-26 10:42:04 +03002136
2137 spin_lock_bh(&vxlan->hash_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002138 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2139 struct vxlan_fdb *f
2140 = container_of(p, struct vxlan_fdb, hlist);
2141 unsigned long timeout;
2142
stephen hemminger3c172862012-10-26 06:24:34 +00002143 if (f->state & NUD_PERMANENT)
stephen hemmingerd3428942012-10-01 12:32:35 +00002144 continue;
2145
2146 timeout = f->used + vxlan->age_interval * HZ;
2147 if (time_before_eq(timeout, jiffies)) {
2148 netdev_dbg(vxlan->dev,
2149 "garbage collect %pM\n",
2150 f->eth_addr);
2151 f->state = NUD_STALE;
2152 vxlan_fdb_destroy(vxlan, f);
2153 } else if (time_before(timeout, next_timer))
2154 next_timer = timeout;
2155 }
Sorin Dumitru14e1d0f2015-05-26 10:42:04 +03002156 spin_unlock_bh(&vxlan->hash_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002157 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002158
2159 mod_timer(&vxlan->age_timer, next_timer);
2160}
2161
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002162static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan)
2163{
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002164 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002165 __u32 vni = vxlan->default_dst.remote_vni;
2166
2167 vxlan->vn_sock = vs;
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002168 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002169 hlist_add_head_rcu(&vxlan->hlist, vni_head(vs, vni));
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002170 spin_unlock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002171}
2172
stephen hemmingerd3428942012-10-01 12:32:35 +00002173/* Setup stats when device is created */
2174static int vxlan_init(struct net_device *dev)
2175{
WANG Cong1c213bd2014-02-13 11:46:28 -08002176 dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
Pravin B Shelare8171042013-03-25 14:49:46 +00002177 if (!dev->tstats)
stephen hemmingerd3428942012-10-01 12:32:35 +00002178 return -ENOMEM;
2179
2180 return 0;
2181}
2182
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002183static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan)
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002184{
2185 struct vxlan_fdb *f;
2186
2187 spin_lock_bh(&vxlan->hash_lock);
2188 f = __vxlan_find_mac(vxlan, all_zeros_mac);
2189 if (f)
2190 vxlan_fdb_destroy(vxlan, f);
2191 spin_unlock_bh(&vxlan->hash_lock);
2192}
2193
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002194static void vxlan_uninit(struct net_device *dev)
2195{
2196 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002197
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002198 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002199
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002200 free_percpu(dev->tstats);
2201}
2202
stephen hemmingerd3428942012-10-01 12:32:35 +00002203/* Start ageing timer and join group when device is brought up */
2204static int vxlan_open(struct net_device *dev)
2205{
2206 struct vxlan_dev *vxlan = netdev_priv(dev);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002207 struct vxlan_sock *vs;
2208 int ret = 0;
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002209
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002210 vs = vxlan_sock_add(vxlan->net, vxlan->dst_port, vxlan_rcv, NULL,
2211 false, vxlan->flags);
2212 if (IS_ERR(vs))
2213 return PTR_ERR(vs);
2214
2215 vxlan_vs_add_dev(vs, vxlan);
stephen hemmingerd3428942012-10-01 12:32:35 +00002216
Gao feng79d4a942013-12-10 16:37:32 +08002217 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002218 ret = vxlan_igmp_join(vxlan);
2219 if (ret) {
2220 vxlan_sock_release(vs);
2221 return ret;
2222 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002223 }
2224
2225 if (vxlan->age_interval)
2226 mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL);
2227
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002228 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002229}
2230
2231/* Purge the forwarding table */
2232static void vxlan_flush(struct vxlan_dev *vxlan)
2233{
Cong Wang31fec5a2013-05-27 22:35:52 +00002234 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002235
2236 spin_lock_bh(&vxlan->hash_lock);
2237 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2238 struct hlist_node *p, *n;
2239 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2240 struct vxlan_fdb *f
2241 = container_of(p, struct vxlan_fdb, hlist);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002242 /* the all_zeros_mac entry is deleted at vxlan_uninit */
2243 if (!is_zero_ether_addr(f->eth_addr))
2244 vxlan_fdb_destroy(vxlan, f);
stephen hemmingerd3428942012-10-01 12:32:35 +00002245 }
2246 }
2247 spin_unlock_bh(&vxlan->hash_lock);
2248}
2249
2250/* Cleanup timer and forwarding table on shutdown */
2251static int vxlan_stop(struct net_device *dev)
2252{
2253 struct vxlan_dev *vxlan = netdev_priv(dev);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002254 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002255 struct vxlan_sock *vs = vxlan->vn_sock;
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002256 int ret = 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002257
Marcelo Ricardo Leitner24c0e682015-03-23 16:23:12 -03002258 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip) &&
WANG Congf13b1682015-04-08 14:48:30 -07002259 !vxlan_group_used(vn, vxlan))
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002260 ret = vxlan_igmp_leave(vxlan);
stephen hemmingerd3428942012-10-01 12:32:35 +00002261
2262 del_timer_sync(&vxlan->age_timer);
2263
2264 vxlan_flush(vxlan);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002265 vxlan_sock_release(vs);
stephen hemmingerd3428942012-10-01 12:32:35 +00002266
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002267 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002268}
2269
stephen hemmingerd3428942012-10-01 12:32:35 +00002270/* Stub, nothing needs to be done. */
2271static void vxlan_set_multicast_list(struct net_device *dev)
2272{
2273}
2274
Daniel Borkmann345010b2013-12-18 00:21:08 +01002275static int vxlan_change_mtu(struct net_device *dev, int new_mtu)
2276{
2277 struct vxlan_dev *vxlan = netdev_priv(dev);
2278 struct vxlan_rdst *dst = &vxlan->default_dst;
2279 struct net_device *lowerdev;
2280 int max_mtu;
2281
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002282 lowerdev = __dev_get_by_index(vxlan->net, dst->remote_ifindex);
Daniel Borkmann345010b2013-12-18 00:21:08 +01002283 if (lowerdev == NULL)
2284 return eth_change_mtu(dev, new_mtu);
2285
2286 if (dst->remote_ip.sa.sa_family == AF_INET6)
2287 max_mtu = lowerdev->mtu - VXLAN6_HEADROOM;
2288 else
2289 max_mtu = lowerdev->mtu - VXLAN_HEADROOM;
2290
2291 if (new_mtu < 68 || new_mtu > max_mtu)
2292 return -EINVAL;
2293
2294 dev->mtu = new_mtu;
2295 return 0;
2296}
2297
stephen hemmingerd3428942012-10-01 12:32:35 +00002298static const struct net_device_ops vxlan_netdev_ops = {
2299 .ndo_init = vxlan_init,
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002300 .ndo_uninit = vxlan_uninit,
stephen hemmingerd3428942012-10-01 12:32:35 +00002301 .ndo_open = vxlan_open,
2302 .ndo_stop = vxlan_stop,
2303 .ndo_start_xmit = vxlan_xmit,
Pravin B Shelare8171042013-03-25 14:49:46 +00002304 .ndo_get_stats64 = ip_tunnel_get_stats64,
stephen hemmingerd3428942012-10-01 12:32:35 +00002305 .ndo_set_rx_mode = vxlan_set_multicast_list,
Daniel Borkmann345010b2013-12-18 00:21:08 +01002306 .ndo_change_mtu = vxlan_change_mtu,
stephen hemmingerd3428942012-10-01 12:32:35 +00002307 .ndo_validate_addr = eth_validate_addr,
2308 .ndo_set_mac_address = eth_mac_addr,
2309 .ndo_fdb_add = vxlan_fdb_add,
2310 .ndo_fdb_del = vxlan_fdb_delete,
2311 .ndo_fdb_dump = vxlan_fdb_dump,
2312};
2313
2314/* Info for udev, that this is a virtual tunnel endpoint */
2315static struct device_type vxlan_type = {
2316 .name = "vxlan",
2317};
2318
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002319/* Calls the ndo_add_vxlan_port of the caller in order to
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002320 * supply the listening VXLAN udp ports. Callers are expected
2321 * to implement the ndo_add_vxlan_port.
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002322 */
2323void vxlan_get_rx_port(struct net_device *dev)
2324{
2325 struct vxlan_sock *vs;
2326 struct net *net = dev_net(dev);
2327 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2328 sa_family_t sa_family;
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002329 __be16 port;
2330 unsigned int i;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002331
2332 spin_lock(&vn->sock_lock);
2333 for (i = 0; i < PORT_HASH_SIZE; ++i) {
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002334 hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) {
2335 port = inet_sk(vs->sock->sk)->inet_sport;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002336 sa_family = vs->sock->sk->sk_family;
2337 dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
2338 port);
2339 }
2340 }
2341 spin_unlock(&vn->sock_lock);
2342}
2343EXPORT_SYMBOL_GPL(vxlan_get_rx_port);
2344
stephen hemmingerd3428942012-10-01 12:32:35 +00002345/* Initialize the device structure. */
2346static void vxlan_setup(struct net_device *dev)
2347{
2348 struct vxlan_dev *vxlan = netdev_priv(dev);
Cong Wang31fec5a2013-05-27 22:35:52 +00002349 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002350
2351 eth_hw_addr_random(dev);
2352 ether_setup(dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08002353 if (vxlan->default_dst.remote_ip.sa.sa_family == AF_INET6)
Cong Wang2853af62014-06-12 11:53:10 -07002354 dev->needed_headroom = ETH_HLEN + VXLAN6_HEADROOM;
Cong Wange4c7ed42013-08-31 13:44:33 +08002355 else
Cong Wang2853af62014-06-12 11:53:10 -07002356 dev->needed_headroom = ETH_HLEN + VXLAN_HEADROOM;
stephen hemmingerd3428942012-10-01 12:32:35 +00002357
2358 dev->netdev_ops = &vxlan_netdev_ops;
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002359 dev->destructor = free_netdev;
stephen hemmingerd3428942012-10-01 12:32:35 +00002360 SET_NETDEV_DEVTYPE(dev, &vxlan_type);
2361
2362 dev->tx_queue_len = 0;
2363 dev->features |= NETIF_F_LLTX;
Joseph Gasparakisd6727fe2012-12-07 14:14:16 +00002364 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002365 dev->features |= NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002366 dev->features |= NETIF_F_GSO_SOFTWARE;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002367
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002368 dev->vlan_features = dev->features;
2369 dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002370 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002371 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002372 dev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Eric Dumazet02875872014-10-05 18:38:35 -07002373 netif_keep_dst(dev);
stephen hemminger6602d002012-12-31 12:00:21 +00002374 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
stephen hemmingerd3428942012-10-01 12:32:35 +00002375
stephen hemminger553675f2013-05-16 11:35:20 +00002376 INIT_LIST_HEAD(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00002377 spin_lock_init(&vxlan->hash_lock);
2378
2379 init_timer_deferrable(&vxlan->age_timer);
2380 vxlan->age_timer.function = vxlan_cleanup;
2381 vxlan->age_timer.data = (unsigned long) vxlan;
2382
stephen hemminger823aa872013-04-27 11:31:57 +00002383 vxlan->dst_port = htons(vxlan_port);
stephen hemminger05f47d62012-10-09 20:35:50 +00002384
stephen hemmingerd3428942012-10-01 12:32:35 +00002385 vxlan->dev = dev;
2386
2387 for (h = 0; h < FDB_HASH_SIZE; ++h)
2388 INIT_HLIST_HEAD(&vxlan->fdb_head[h]);
2389}
2390
2391static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = {
2392 [IFLA_VXLAN_ID] = { .type = NLA_U32 },
stephen hemminger5d174dd2013-04-27 11:31:55 +00002393 [IFLA_VXLAN_GROUP] = { .len = FIELD_SIZEOF(struct iphdr, daddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002394 [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002395 [IFLA_VXLAN_LINK] = { .type = NLA_U32 },
2396 [IFLA_VXLAN_LOCAL] = { .len = FIELD_SIZEOF(struct iphdr, saddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002397 [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002398 [IFLA_VXLAN_TOS] = { .type = NLA_U8 },
2399 [IFLA_VXLAN_TTL] = { .type = NLA_U8 },
2400 [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 },
2401 [IFLA_VXLAN_AGEING] = { .type = NLA_U32 },
2402 [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 },
stephen hemminger05f47d62012-10-09 20:35:50 +00002403 [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) },
David Stevense4f67ad2012-11-20 02:50:14 +00002404 [IFLA_VXLAN_PROXY] = { .type = NLA_U8 },
2405 [IFLA_VXLAN_RSC] = { .type = NLA_U8 },
2406 [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 },
2407 [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 },
stephen hemminger823aa872013-04-27 11:31:57 +00002408 [IFLA_VXLAN_PORT] = { .type = NLA_U16 },
Tom Herbert5c91ae02014-11-06 18:06:01 -08002409 [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 },
2410 [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 },
2411 [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 },
Tom Herbertdfd86452015-01-12 17:00:38 -08002412 [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 },
2413 [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 },
Thomas Graf35114942015-01-15 03:53:55 +01002414 [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, },
Tom Herbert0ace2ca2015-02-10 16:30:32 -08002415 [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG },
stephen hemmingerd3428942012-10-01 12:32:35 +00002416};
2417
2418static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[])
2419{
2420 if (tb[IFLA_ADDRESS]) {
2421 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) {
2422 pr_debug("invalid link address (not ethernet)\n");
2423 return -EINVAL;
2424 }
2425
2426 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) {
2427 pr_debug("invalid all zero ethernet address\n");
2428 return -EADDRNOTAVAIL;
2429 }
2430 }
2431
2432 if (!data)
2433 return -EINVAL;
2434
2435 if (data[IFLA_VXLAN_ID]) {
2436 __u32 id = nla_get_u32(data[IFLA_VXLAN_ID]);
2437 if (id >= VXLAN_VID_MASK)
2438 return -ERANGE;
2439 }
2440
stephen hemminger05f47d62012-10-09 20:35:50 +00002441 if (data[IFLA_VXLAN_PORT_RANGE]) {
2442 const struct ifla_vxlan_port_range *p
2443 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2444
2445 if (ntohs(p->high) < ntohs(p->low)) {
2446 pr_debug("port range %u .. %u not valid\n",
2447 ntohs(p->low), ntohs(p->high));
2448 return -EINVAL;
2449 }
2450 }
2451
stephen hemmingerd3428942012-10-01 12:32:35 +00002452 return 0;
2453}
2454
Yan Burman1b13c972013-01-29 23:43:07 +00002455static void vxlan_get_drvinfo(struct net_device *netdev,
2456 struct ethtool_drvinfo *drvinfo)
2457{
2458 strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version));
2459 strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver));
2460}
2461
2462static const struct ethtool_ops vxlan_ethtool_ops = {
2463 .get_drvinfo = vxlan_get_drvinfo,
2464 .get_link = ethtool_op_get_link,
2465};
2466
stephen hemminger553675f2013-05-16 11:35:20 +00002467static void vxlan_del_work(struct work_struct *work)
2468{
2469 struct vxlan_sock *vs = container_of(work, struct vxlan_sock, del_work);
Andy Zhouacbf74a2014-09-16 17:31:18 -07002470 udp_tunnel_sock_release(vs->sock);
stephen hemminger553675f2013-05-16 11:35:20 +00002471 kfree_rcu(vs, rcu);
2472}
2473
Tom Herbert3ee64f32014-07-13 19:49:42 -07002474static struct socket *vxlan_create_sock(struct net *net, bool ipv6,
2475 __be16 port, u32 flags)
stephen hemminger553675f2013-05-16 11:35:20 +00002476{
Cong Wange4c7ed42013-08-31 13:44:33 +08002477 struct socket *sock;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002478 struct udp_port_cfg udp_conf;
2479 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002480
Tom Herbert3ee64f32014-07-13 19:49:42 -07002481 memset(&udp_conf, 0, sizeof(udp_conf));
2482
2483 if (ipv6) {
2484 udp_conf.family = AF_INET6;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002485 udp_conf.use_udp6_rx_checksums =
Alexander Duyck3dc2b6a2014-11-24 20:08:38 -08002486 !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX);
Tom Herbert3ee64f32014-07-13 19:49:42 -07002487 } else {
2488 udp_conf.family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +08002489 }
2490
Tom Herbert3ee64f32014-07-13 19:49:42 -07002491 udp_conf.local_udp_port = port;
Cong Wange4c7ed42013-08-31 13:44:33 +08002492
Tom Herbert3ee64f32014-07-13 19:49:42 -07002493 /* Open UDP socket */
2494 err = udp_sock_create(net, &udp_conf, &sock);
2495 if (err < 0)
2496 return ERR_PTR(err);
Cong Wange4c7ed42013-08-31 13:44:33 +08002497
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002498 return sock;
Cong Wange4c7ed42013-08-31 13:44:33 +08002499}
2500
2501/* Create new listen socket if needed */
2502static struct vxlan_sock *vxlan_socket_create(struct net *net, __be16 port,
Tom Herbert359a0ea2014-06-04 17:20:29 -07002503 vxlan_rcv_t *rcv, void *data,
2504 u32 flags)
Cong Wange4c7ed42013-08-31 13:44:33 +08002505{
2506 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2507 struct vxlan_sock *vs;
2508 struct socket *sock;
Cong Wang31fec5a2013-05-27 22:35:52 +00002509 unsigned int h;
Tom Herbert359a0ea2014-06-04 17:20:29 -07002510 bool ipv6 = !!(flags & VXLAN_F_IPV6);
Andy Zhouacbf74a2014-09-16 17:31:18 -07002511 struct udp_tunnel_sock_cfg tunnel_cfg;
stephen hemminger553675f2013-05-16 11:35:20 +00002512
Or Gerlitzdc01e7d2014-01-20 13:59:21 +02002513 vs = kzalloc(sizeof(*vs), GFP_KERNEL);
Cong Wange4c7ed42013-08-31 13:44:33 +08002514 if (!vs)
stephen hemminger553675f2013-05-16 11:35:20 +00002515 return ERR_PTR(-ENOMEM);
2516
2517 for (h = 0; h < VNI_HASH_SIZE; ++h)
2518 INIT_HLIST_HEAD(&vs->vni_list[h]);
2519
2520 INIT_WORK(&vs->del_work, vxlan_del_work);
2521
Tom Herbert3ee64f32014-07-13 19:49:42 -07002522 sock = vxlan_create_sock(net, ipv6, port, flags);
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002523 if (IS_ERR(sock)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002524 pr_info("Cannot bind port %d, err=%ld\n", ntohs(port),
2525 PTR_ERR(sock));
stephen hemminger553675f2013-05-16 11:35:20 +00002526 kfree(vs);
Duan Jionge50fddc2013-11-01 13:09:43 +08002527 return ERR_CAST(sock);
stephen hemminger553675f2013-05-16 11:35:20 +00002528 }
2529
Cong Wange4c7ed42013-08-31 13:44:33 +08002530 vs->sock = sock;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002531 atomic_set(&vs->refcnt, 1);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07002532 vs->rcv = rcv;
Pravin B Shelar012a5722013-08-19 11:23:07 -07002533 vs->data = data;
Tom Herbertaf33c1a2015-01-20 11:23:05 -08002534 vs->flags = (flags & VXLAN_F_RCV_FLAGS);
stephen hemminger553675f2013-05-16 11:35:20 +00002535
Or Gerlitzdc01e7d2014-01-20 13:59:21 +02002536 /* Initialize the vxlan udp offloads structure */
2537 vs->udp_offloads.port = port;
2538 vs->udp_offloads.callbacks.gro_receive = vxlan_gro_receive;
2539 vs->udp_offloads.callbacks.gro_complete = vxlan_gro_complete;
2540
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002541 spin_lock(&vn->sock_lock);
2542 hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
Or Gerlitzdc01e7d2014-01-20 13:59:21 +02002543 vxlan_notify_add_rx_port(vs);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002544 spin_unlock(&vn->sock_lock);
stephen hemminger553675f2013-05-16 11:35:20 +00002545
2546 /* Mark socket as an encapsulation socket. */
Andy Zhouacbf74a2014-09-16 17:31:18 -07002547 tunnel_cfg.sk_user_data = vs;
2548 tunnel_cfg.encap_type = 1;
2549 tunnel_cfg.encap_rcv = vxlan_udp_encap_recv;
2550 tunnel_cfg.encap_destroy = NULL;
2551
2552 setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
Cong Wange4c7ed42013-08-31 13:44:33 +08002553
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002554 return vs;
2555}
stephen hemminger553675f2013-05-16 11:35:20 +00002556
Pravin B Shelar012a5722013-08-19 11:23:07 -07002557struct vxlan_sock *vxlan_sock_add(struct net *net, __be16 port,
2558 vxlan_rcv_t *rcv, void *data,
Tom Herbert359a0ea2014-06-04 17:20:29 -07002559 bool no_share, u32 flags)
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002560{
2561 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2562 struct vxlan_sock *vs;
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02002563 bool ipv6 = flags & VXLAN_F_IPV6;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002564
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002565 if (!no_share) {
2566 spin_lock(&vn->sock_lock);
2567 vs = vxlan_find_sock(net, ipv6 ? AF_INET6 : AF_INET, port,
2568 flags);
2569 if (vs && vs->rcv == rcv) {
2570 if (!atomic_add_unless(&vs->refcnt, 1, 0))
2571 vs = ERR_PTR(-EBUSY);
2572 spin_unlock(&vn->sock_lock);
2573 return vs;
2574 }
2575 spin_unlock(&vn->sock_lock);
2576 }
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002577
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002578 return vxlan_socket_create(net, port, rcv, data, flags);
stephen hemminger553675f2013-05-16 11:35:20 +00002579}
Pravin B Shelar012a5722013-08-19 11:23:07 -07002580EXPORT_SYMBOL_GPL(vxlan_sock_add);
stephen hemminger553675f2013-05-16 11:35:20 +00002581
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002582static int vxlan_newlink(struct net *src_net, struct net_device *dev,
stephen hemmingerd3428942012-10-01 12:32:35 +00002583 struct nlattr *tb[], struct nlattr *data[])
2584{
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002585 struct vxlan_net *vn = net_generic(src_net, vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00002586 struct vxlan_dev *vxlan = netdev_priv(dev);
Atzm Watanabec7995c42013-04-16 02:50:52 +00002587 struct vxlan_rdst *dst = &vxlan->default_dst;
stephen hemmingerd3428942012-10-01 12:32:35 +00002588 __u32 vni;
2589 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002590 bool use_ipv6 = false;
stephen hemmingerd3428942012-10-01 12:32:35 +00002591
2592 if (!data[IFLA_VXLAN_ID])
2593 return -EINVAL;
2594
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002595 vxlan->net = src_net;
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002596
stephen hemmingerd3428942012-10-01 12:32:35 +00002597 vni = nla_get_u32(data[IFLA_VXLAN_ID]);
Atzm Watanabec7995c42013-04-16 02:50:52 +00002598 dst->remote_vni = vni;
stephen hemmingerd3428942012-10-01 12:32:35 +00002599
Mike Rapoport5933a7b2014-04-01 09:23:01 +03002600 /* Unless IPv6 is explicitly requested, assume IPv4 */
2601 dst->remote_ip.sa.sa_family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +08002602 if (data[IFLA_VXLAN_GROUP]) {
Jiri Benc67b61f62015-03-29 16:59:26 +02002603 dst->remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]);
Cong Wange4c7ed42013-08-31 13:44:33 +08002604 } else if (data[IFLA_VXLAN_GROUP6]) {
2605 if (!IS_ENABLED(CONFIG_IPV6))
2606 return -EPFNOSUPPORT;
stephen hemmingerd3428942012-10-01 12:32:35 +00002607
Jiri Benc67b61f62015-03-29 16:59:26 +02002608 dst->remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]);
Cong Wange4c7ed42013-08-31 13:44:33 +08002609 dst->remote_ip.sa.sa_family = AF_INET6;
2610 use_ipv6 = true;
2611 }
2612
2613 if (data[IFLA_VXLAN_LOCAL]) {
Jiri Benc67b61f62015-03-29 16:59:26 +02002614 vxlan->saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]);
Cong Wange4c7ed42013-08-31 13:44:33 +08002615 vxlan->saddr.sa.sa_family = AF_INET;
2616 } else if (data[IFLA_VXLAN_LOCAL6]) {
2617 if (!IS_ENABLED(CONFIG_IPV6))
2618 return -EPFNOSUPPORT;
2619
2620 /* TODO: respect scope id */
Jiri Benc67b61f62015-03-29 16:59:26 +02002621 vxlan->saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]);
Cong Wange4c7ed42013-08-31 13:44:33 +08002622 vxlan->saddr.sa.sa_family = AF_INET6;
2623 use_ipv6 = true;
2624 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002625
stephen hemminger34e02aa2012-10-09 20:35:53 +00002626 if (data[IFLA_VXLAN_LINK] &&
Atzm Watanabec7995c42013-04-16 02:50:52 +00002627 (dst->remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]))) {
stephen hemminger34e02aa2012-10-09 20:35:53 +00002628 struct net_device *lowerdev
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002629 = __dev_get_by_index(src_net, dst->remote_ifindex);
stephen hemmingerd3428942012-10-01 12:32:35 +00002630
stephen hemminger34e02aa2012-10-09 20:35:53 +00002631 if (!lowerdev) {
Atzm Watanabec7995c42013-04-16 02:50:52 +00002632 pr_info("ifindex %d does not exist\n", dst->remote_ifindex);
stephen hemminger34e02aa2012-10-09 20:35:53 +00002633 return -ENODEV;
stephen hemmingerd3428942012-10-01 12:32:35 +00002634 }
stephen hemminger34e02aa2012-10-09 20:35:53 +00002635
Cong Wange4c7ed42013-08-31 13:44:33 +08002636#if IS_ENABLED(CONFIG_IPV6)
2637 if (use_ipv6) {
2638 struct inet6_dev *idev = __in6_dev_get(lowerdev);
2639 if (idev && idev->cnf.disable_ipv6) {
2640 pr_info("IPv6 is disabled via sysctl\n");
2641 return -EPERM;
2642 }
2643 vxlan->flags |= VXLAN_F_IPV6;
2644 }
2645#endif
2646
stephen hemminger34e02aa2012-10-09 20:35:53 +00002647 if (!tb[IFLA_MTU])
Cong Wange4c7ed42013-08-31 13:44:33 +08002648 dev->mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
Alexander Duyck1ba56fb2012-11-13 13:10:59 +00002649
Cong Wang2853af62014-06-12 11:53:10 -07002650 dev->needed_headroom = lowerdev->hard_header_len +
Cong Wange4c7ed42013-08-31 13:44:33 +08002651 (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
fan.du7bda7012014-01-03 10:18:58 +08002652 } else if (use_ipv6)
2653 vxlan->flags |= VXLAN_F_IPV6;
stephen hemmingerd3428942012-10-01 12:32:35 +00002654
2655 if (data[IFLA_VXLAN_TOS])
2656 vxlan->tos = nla_get_u8(data[IFLA_VXLAN_TOS]);
2657
Vincent Bernatafb97182012-10-30 10:27:16 +00002658 if (data[IFLA_VXLAN_TTL])
2659 vxlan->ttl = nla_get_u8(data[IFLA_VXLAN_TTL]);
2660
stephen hemmingerd3428942012-10-01 12:32:35 +00002661 if (!data[IFLA_VXLAN_LEARNING] || nla_get_u8(data[IFLA_VXLAN_LEARNING]))
David Stevense4f67ad2012-11-20 02:50:14 +00002662 vxlan->flags |= VXLAN_F_LEARN;
stephen hemmingerd3428942012-10-01 12:32:35 +00002663
2664 if (data[IFLA_VXLAN_AGEING])
2665 vxlan->age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]);
2666 else
2667 vxlan->age_interval = FDB_AGE_DEFAULT;
2668
David Stevense4f67ad2012-11-20 02:50:14 +00002669 if (data[IFLA_VXLAN_PROXY] && nla_get_u8(data[IFLA_VXLAN_PROXY]))
2670 vxlan->flags |= VXLAN_F_PROXY;
2671
2672 if (data[IFLA_VXLAN_RSC] && nla_get_u8(data[IFLA_VXLAN_RSC]))
2673 vxlan->flags |= VXLAN_F_RSC;
2674
2675 if (data[IFLA_VXLAN_L2MISS] && nla_get_u8(data[IFLA_VXLAN_L2MISS]))
2676 vxlan->flags |= VXLAN_F_L2MISS;
2677
2678 if (data[IFLA_VXLAN_L3MISS] && nla_get_u8(data[IFLA_VXLAN_L3MISS]))
2679 vxlan->flags |= VXLAN_F_L3MISS;
2680
stephen hemmingerd3428942012-10-01 12:32:35 +00002681 if (data[IFLA_VXLAN_LIMIT])
2682 vxlan->addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]);
2683
stephen hemminger05f47d62012-10-09 20:35:50 +00002684 if (data[IFLA_VXLAN_PORT_RANGE]) {
2685 const struct ifla_vxlan_port_range *p
2686 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2687 vxlan->port_min = ntohs(p->low);
2688 vxlan->port_max = ntohs(p->high);
2689 }
2690
stephen hemminger823aa872013-04-27 11:31:57 +00002691 if (data[IFLA_VXLAN_PORT])
2692 vxlan->dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]);
2693
Tom Herbert359a0ea2014-06-04 17:20:29 -07002694 if (data[IFLA_VXLAN_UDP_CSUM] && nla_get_u8(data[IFLA_VXLAN_UDP_CSUM]))
2695 vxlan->flags |= VXLAN_F_UDP_CSUM;
2696
2697 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX] &&
2698 nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX]))
2699 vxlan->flags |= VXLAN_F_UDP_ZERO_CSUM6_TX;
2700
2701 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX] &&
2702 nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX]))
2703 vxlan->flags |= VXLAN_F_UDP_ZERO_CSUM6_RX;
2704
Tom Herbertdfd86452015-01-12 17:00:38 -08002705 if (data[IFLA_VXLAN_REMCSUM_TX] &&
2706 nla_get_u8(data[IFLA_VXLAN_REMCSUM_TX]))
2707 vxlan->flags |= VXLAN_F_REMCSUM_TX;
2708
2709 if (data[IFLA_VXLAN_REMCSUM_RX] &&
2710 nla_get_u8(data[IFLA_VXLAN_REMCSUM_RX]))
2711 vxlan->flags |= VXLAN_F_REMCSUM_RX;
2712
Thomas Graf35114942015-01-15 03:53:55 +01002713 if (data[IFLA_VXLAN_GBP])
2714 vxlan->flags |= VXLAN_F_GBP;
2715
Tom Herbert0ace2ca2015-02-10 16:30:32 -08002716 if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL])
2717 vxlan->flags |= VXLAN_F_REMCSUM_NOPARTIAL;
2718
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002719 if (vxlan_find_vni(src_net, vni, use_ipv6 ? AF_INET6 : AF_INET,
Thomas Grafac5132d2015-01-15 03:53:56 +01002720 vxlan->dst_port, vxlan->flags)) {
stephen hemminger553675f2013-05-16 11:35:20 +00002721 pr_info("duplicate VNI %u\n", vni);
2722 return -EEXIST;
2723 }
2724
Wilfried Klaebe7ad24ea2014-05-11 00:12:32 +00002725 dev->ethtool_ops = &vxlan_ethtool_ops;
Yan Burman1b13c972013-01-29 23:43:07 +00002726
Sridhar Samudrala2936b6a2013-09-17 12:12:40 -07002727 /* create an fdb entry for a valid default destination */
2728 if (!vxlan_addr_any(&vxlan->default_dst.remote_ip)) {
2729 err = vxlan_fdb_create(vxlan, all_zeros_mac,
2730 &vxlan->default_dst.remote_ip,
2731 NUD_REACHABLE|NUD_PERMANENT,
2732 NLM_F_EXCL|NLM_F_CREATE,
2733 vxlan->dst_port,
2734 vxlan->default_dst.remote_vni,
2735 vxlan->default_dst.remote_ifindex,
2736 NTF_SELF);
2737 if (err)
2738 return err;
2739 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002740
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002741 err = register_netdevice(dev);
2742 if (err) {
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002743 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002744 return err;
2745 }
2746
stephen hemminger553675f2013-05-16 11:35:20 +00002747 list_add(&vxlan->next, &vn->vxlan_list);
stephen hemminger553675f2013-05-16 11:35:20 +00002748
2749 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002750}
2751
2752static void vxlan_dellink(struct net_device *dev, struct list_head *head)
2753{
2754 struct vxlan_dev *vxlan = netdev_priv(dev);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002755 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00002756
stephen hemmingerfe5c3562013-07-13 10:18:18 -07002757 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002758 if (!hlist_unhashed(&vxlan->hlist))
2759 hlist_del_rcu(&vxlan->hlist);
stephen hemmingerfe5c3562013-07-13 10:18:18 -07002760 spin_unlock(&vn->sock_lock);
2761
stephen hemminger553675f2013-05-16 11:35:20 +00002762 list_del(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00002763 unregister_netdevice_queue(dev, head);
2764}
2765
2766static size_t vxlan_get_size(const struct net_device *dev)
2767{
2768
2769 return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */
Cong Wange4c7ed42013-08-31 13:44:33 +08002770 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00002771 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */
Cong Wange4c7ed42013-08-31 13:44:33 +08002772 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00002773 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */
2774 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */
2775 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */
David Stevense4f67ad2012-11-20 02:50:14 +00002776 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */
2777 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */
2778 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */
2779 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */
stephen hemmingerd3428942012-10-01 12:32:35 +00002780 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */
2781 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */
stephen hemminger05f47d62012-10-09 20:35:50 +00002782 nla_total_size(sizeof(struct ifla_vxlan_port_range)) +
Tom Herbert359a0ea2014-06-04 17:20:29 -07002783 nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */
2784 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */
2785 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */
2786 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */
Tom Herbertdfd86452015-01-12 17:00:38 -08002787 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */
2788 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */
stephen hemmingerd3428942012-10-01 12:32:35 +00002789 0;
2790}
2791
2792static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev)
2793{
2794 const struct vxlan_dev *vxlan = netdev_priv(dev);
Atzm Watanabec7995c42013-04-16 02:50:52 +00002795 const struct vxlan_rdst *dst = &vxlan->default_dst;
stephen hemminger05f47d62012-10-09 20:35:50 +00002796 struct ifla_vxlan_port_range ports = {
2797 .low = htons(vxlan->port_min),
2798 .high = htons(vxlan->port_max),
2799 };
stephen hemmingerd3428942012-10-01 12:32:35 +00002800
Atzm Watanabec7995c42013-04-16 02:50:52 +00002801 if (nla_put_u32(skb, IFLA_VXLAN_ID, dst->remote_vni))
stephen hemmingerd3428942012-10-01 12:32:35 +00002802 goto nla_put_failure;
2803
Cong Wange4c7ed42013-08-31 13:44:33 +08002804 if (!vxlan_addr_any(&dst->remote_ip)) {
2805 if (dst->remote_ip.sa.sa_family == AF_INET) {
Jiri Benc930345e2015-03-29 16:59:25 +02002806 if (nla_put_in_addr(skb, IFLA_VXLAN_GROUP,
2807 dst->remote_ip.sin.sin_addr.s_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08002808 goto nla_put_failure;
2809#if IS_ENABLED(CONFIG_IPV6)
2810 } else {
Jiri Benc930345e2015-03-29 16:59:25 +02002811 if (nla_put_in6_addr(skb, IFLA_VXLAN_GROUP6,
2812 &dst->remote_ip.sin6.sin6_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08002813 goto nla_put_failure;
2814#endif
2815 }
2816 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002817
Atzm Watanabec7995c42013-04-16 02:50:52 +00002818 if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +00002819 goto nla_put_failure;
2820
Cong Wange4c7ed42013-08-31 13:44:33 +08002821 if (!vxlan_addr_any(&vxlan->saddr)) {
2822 if (vxlan->saddr.sa.sa_family == AF_INET) {
Jiri Benc930345e2015-03-29 16:59:25 +02002823 if (nla_put_in_addr(skb, IFLA_VXLAN_LOCAL,
2824 vxlan->saddr.sin.sin_addr.s_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08002825 goto nla_put_failure;
2826#if IS_ENABLED(CONFIG_IPV6)
2827 } else {
Jiri Benc930345e2015-03-29 16:59:25 +02002828 if (nla_put_in6_addr(skb, IFLA_VXLAN_LOCAL6,
2829 &vxlan->saddr.sin6.sin6_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08002830 goto nla_put_failure;
2831#endif
2832 }
2833 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002834
2835 if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->ttl) ||
2836 nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->tos) ||
David Stevense4f67ad2012-11-20 02:50:14 +00002837 nla_put_u8(skb, IFLA_VXLAN_LEARNING,
2838 !!(vxlan->flags & VXLAN_F_LEARN)) ||
2839 nla_put_u8(skb, IFLA_VXLAN_PROXY,
2840 !!(vxlan->flags & VXLAN_F_PROXY)) ||
2841 nla_put_u8(skb, IFLA_VXLAN_RSC, !!(vxlan->flags & VXLAN_F_RSC)) ||
2842 nla_put_u8(skb, IFLA_VXLAN_L2MISS,
2843 !!(vxlan->flags & VXLAN_F_L2MISS)) ||
2844 nla_put_u8(skb, IFLA_VXLAN_L3MISS,
2845 !!(vxlan->flags & VXLAN_F_L3MISS)) ||
stephen hemmingerd3428942012-10-01 12:32:35 +00002846 nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->age_interval) ||
stephen hemminger823aa872013-04-27 11:31:57 +00002847 nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->addrmax) ||
Tom Herbert359a0ea2014-06-04 17:20:29 -07002848 nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->dst_port) ||
2849 nla_put_u8(skb, IFLA_VXLAN_UDP_CSUM,
2850 !!(vxlan->flags & VXLAN_F_UDP_CSUM)) ||
2851 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_TX,
2852 !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_TX)) ||
2853 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_RX,
Tom Herbertdfd86452015-01-12 17:00:38 -08002854 !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_RX)) ||
2855 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_TX,
2856 !!(vxlan->flags & VXLAN_F_REMCSUM_TX)) ||
2857 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_RX,
2858 !!(vxlan->flags & VXLAN_F_REMCSUM_RX)))
stephen hemmingerd3428942012-10-01 12:32:35 +00002859 goto nla_put_failure;
2860
stephen hemminger05f47d62012-10-09 20:35:50 +00002861 if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports))
2862 goto nla_put_failure;
2863
Thomas Graf35114942015-01-15 03:53:55 +01002864 if (vxlan->flags & VXLAN_F_GBP &&
2865 nla_put_flag(skb, IFLA_VXLAN_GBP))
2866 goto nla_put_failure;
2867
Tom Herbert0ace2ca2015-02-10 16:30:32 -08002868 if (vxlan->flags & VXLAN_F_REMCSUM_NOPARTIAL &&
2869 nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL))
2870 goto nla_put_failure;
2871
stephen hemmingerd3428942012-10-01 12:32:35 +00002872 return 0;
2873
2874nla_put_failure:
2875 return -EMSGSIZE;
2876}
2877
Nicolas Dichtel1728d4f2015-01-15 15:11:17 +01002878static struct net *vxlan_get_link_net(const struct net_device *dev)
2879{
2880 struct vxlan_dev *vxlan = netdev_priv(dev);
2881
2882 return vxlan->net;
2883}
2884
stephen hemmingerd3428942012-10-01 12:32:35 +00002885static struct rtnl_link_ops vxlan_link_ops __read_mostly = {
2886 .kind = "vxlan",
2887 .maxtype = IFLA_VXLAN_MAX,
2888 .policy = vxlan_policy,
2889 .priv_size = sizeof(struct vxlan_dev),
2890 .setup = vxlan_setup,
2891 .validate = vxlan_validate,
2892 .newlink = vxlan_newlink,
2893 .dellink = vxlan_dellink,
2894 .get_size = vxlan_get_size,
2895 .fill_info = vxlan_fill_info,
Nicolas Dichtel1728d4f2015-01-15 15:11:17 +01002896 .get_link_net = vxlan_get_link_net,
stephen hemmingerd3428942012-10-01 12:32:35 +00002897};
2898
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01002899static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn,
2900 struct net_device *dev)
2901{
2902 struct vxlan_dev *vxlan, *next;
2903 LIST_HEAD(list_kill);
2904
2905 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
2906 struct vxlan_rdst *dst = &vxlan->default_dst;
2907
2908 /* In case we created vxlan device with carrier
2909 * and we loose the carrier due to module unload
2910 * we also need to remove vxlan device. In other
2911 * cases, it's not necessary and remote_ifindex
2912 * is 0 here, so no matches.
2913 */
2914 if (dst->remote_ifindex == dev->ifindex)
2915 vxlan_dellink(vxlan->dev, &list_kill);
2916 }
2917
2918 unregister_netdevice_many(&list_kill);
2919}
2920
2921static int vxlan_lowerdev_event(struct notifier_block *unused,
2922 unsigned long event, void *ptr)
2923{
2924 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
Daniel Borkmann783c1462014-01-22 21:07:53 +01002925 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01002926
Daniel Borkmann783c1462014-01-22 21:07:53 +01002927 if (event == NETDEV_UNREGISTER)
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01002928 vxlan_handle_lowerdev_unregister(vn, dev);
2929
2930 return NOTIFY_DONE;
2931}
2932
2933static struct notifier_block vxlan_notifier_block __read_mostly = {
2934 .notifier_call = vxlan_lowerdev_event,
2935};
2936
stephen hemmingerd3428942012-10-01 12:32:35 +00002937static __net_init int vxlan_init_net(struct net *net)
2938{
2939 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Cong Wang31fec5a2013-05-27 22:35:52 +00002940 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002941
stephen hemminger553675f2013-05-16 11:35:20 +00002942 INIT_LIST_HEAD(&vn->vxlan_list);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002943 spin_lock_init(&vn->sock_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002944
stephen hemminger553675f2013-05-16 11:35:20 +00002945 for (h = 0; h < PORT_HASH_SIZE; ++h)
2946 INIT_HLIST_HEAD(&vn->sock_list[h]);
stephen hemmingerd3428942012-10-01 12:32:35 +00002947
2948 return 0;
2949}
2950
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002951static void __net_exit vxlan_exit_net(struct net *net)
2952{
2953 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2954 struct vxlan_dev *vxlan, *next;
2955 struct net_device *dev, *aux;
2956 LIST_HEAD(list);
2957
2958 rtnl_lock();
2959 for_each_netdev_safe(net, dev, aux)
2960 if (dev->rtnl_link_ops == &vxlan_link_ops)
2961 unregister_netdevice_queue(dev, &list);
2962
2963 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
2964 /* If vxlan->dev is in the same netns, it has already been added
2965 * to the list by the previous loop.
2966 */
2967 if (!net_eq(dev_net(vxlan->dev), net))
John W. Linville13c3ed62015-05-18 13:51:24 -04002968 unregister_netdevice_queue(vxlan->dev, &list);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002969 }
2970
2971 unregister_netdevice_many(&list);
2972 rtnl_unlock();
2973}
2974
stephen hemmingerd3428942012-10-01 12:32:35 +00002975static struct pernet_operations vxlan_net_ops = {
2976 .init = vxlan_init_net,
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002977 .exit = vxlan_exit_net,
stephen hemmingerd3428942012-10-01 12:32:35 +00002978 .id = &vxlan_net_id,
2979 .size = sizeof(struct vxlan_net),
2980};
2981
2982static int __init vxlan_init_module(void)
2983{
2984 int rc;
2985
Stephen Hemminger758c57d2013-06-17 14:16:09 -07002986 vxlan_wq = alloc_workqueue("vxlan", 0, 0);
2987 if (!vxlan_wq)
2988 return -ENOMEM;
2989
stephen hemmingerd3428942012-10-01 12:32:35 +00002990 get_random_bytes(&vxlan_salt, sizeof(vxlan_salt));
2991
Daniel Borkmann783c1462014-01-22 21:07:53 +01002992 rc = register_pernet_subsys(&vxlan_net_ops);
stephen hemmingerd3428942012-10-01 12:32:35 +00002993 if (rc)
2994 goto out1;
2995
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01002996 rc = register_netdevice_notifier(&vxlan_notifier_block);
stephen hemmingerd3428942012-10-01 12:32:35 +00002997 if (rc)
2998 goto out2;
2999
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003000 rc = rtnl_link_register(&vxlan_link_ops);
3001 if (rc)
3002 goto out3;
stephen hemmingerd3428942012-10-01 12:32:35 +00003003
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003004 return 0;
3005out3:
3006 unregister_netdevice_notifier(&vxlan_notifier_block);
stephen hemmingerd3428942012-10-01 12:32:35 +00003007out2:
Daniel Borkmann783c1462014-01-22 21:07:53 +01003008 unregister_pernet_subsys(&vxlan_net_ops);
stephen hemmingerd3428942012-10-01 12:32:35 +00003009out1:
Stephen Hemminger758c57d2013-06-17 14:16:09 -07003010 destroy_workqueue(vxlan_wq);
stephen hemmingerd3428942012-10-01 12:32:35 +00003011 return rc;
3012}
Cong Wang7332a132013-05-27 22:35:53 +00003013late_initcall(vxlan_init_module);
stephen hemmingerd3428942012-10-01 12:32:35 +00003014
3015static void __exit vxlan_cleanup_module(void)
3016{
Stephen Hemmingerb7153982013-06-17 14:16:09 -07003017 rtnl_link_unregister(&vxlan_link_ops);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003018 unregister_netdevice_notifier(&vxlan_notifier_block);
Stephen Hemminger758c57d2013-06-17 14:16:09 -07003019 destroy_workqueue(vxlan_wq);
Daniel Borkmann783c1462014-01-22 21:07:53 +01003020 unregister_pernet_subsys(&vxlan_net_ops);
3021 /* rcu_barrier() is called by netns */
stephen hemmingerd3428942012-10-01 12:32:35 +00003022}
3023module_exit(vxlan_cleanup_module);
3024
3025MODULE_LICENSE("GPL");
3026MODULE_VERSION(VXLAN_VERSION);
stephen hemminger3b8df3c2013-04-27 11:31:52 +00003027MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>");
Jesse Brandeburgead51392014-01-17 11:00:33 -08003028MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic");
stephen hemmingerd3428942012-10-01 12:32:35 +00003029MODULE_ALIAS_RTNL_LINK("vxlan");