blob: 58f6a0c02b173f7ad2d753630e1fc69f778d2789 [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>
36#include <net/rtnetlink.h>
37#include <net/route.h>
38#include <net/dsfield.h>
39#include <net/inet_ecn.h>
40#include <net/net_namespace.h>
41#include <net/netns/generic.h>
Pravin B Shelar012a5722013-08-19 11:23:07 -070042#include <net/vxlan.h>
Cong Wange4c7ed42013-08-31 13:44:33 +080043#if IS_ENABLED(CONFIG_IPV6)
44#include <net/ipv6.h>
45#include <net/addrconf.h>
46#include <net/ip6_tunnel.h>
Cong Wang660d98c2013-09-02 10:06:52 +080047#include <net/ip6_checksum.h>
Cong Wange4c7ed42013-08-31 13:44:33 +080048#endif
stephen hemmingerd3428942012-10-01 12:32:35 +000049
50#define VXLAN_VERSION "0.1"
51
stephen hemminger553675f2013-05-16 11:35:20 +000052#define PORT_HASH_BITS 8
53#define PORT_HASH_SIZE (1<<PORT_HASH_BITS)
stephen hemmingerd3428942012-10-01 12:32:35 +000054#define VNI_HASH_BITS 10
55#define VNI_HASH_SIZE (1<<VNI_HASH_BITS)
56#define FDB_HASH_BITS 8
57#define FDB_HASH_SIZE (1<<FDB_HASH_BITS)
58#define FDB_AGE_DEFAULT 300 /* 5 min */
59#define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */
60
61#define VXLAN_N_VID (1u << 24)
62#define VXLAN_VID_MASK (VXLAN_N_VID - 1)
Pravin B Shelar7ce04752013-08-19 11:22:54 -070063#define VXLAN_HLEN (sizeof(struct udphdr) + sizeof(struct vxlanhdr))
stephen hemmingerd3428942012-10-01 12:32:35 +000064
65#define VXLAN_FLAGS 0x08000000 /* struct vxlanhdr.vx_flags required value. */
66
67/* VXLAN protocol header */
68struct vxlanhdr {
69 __be32 vx_flags;
70 __be32 vx_vni;
71};
72
stephen hemminger23c578b2013-04-27 11:31:53 +000073/* UDP port for VXLAN traffic.
74 * The IANA assigned port is 4789, but the Linux default is 8472
Stephen Hemminger234f5b72013-06-17 14:16:41 -070075 * for compatibility with early adopters.
stephen hemminger23c578b2013-04-27 11:31:53 +000076 */
Stephen Hemminger9daaa392013-06-17 14:16:12 -070077static unsigned short vxlan_port __read_mostly = 8472;
78module_param_named(udp_port, vxlan_port, ushort, 0444);
stephen hemmingerd3428942012-10-01 12:32:35 +000079MODULE_PARM_DESC(udp_port, "Destination UDP port");
80
81static bool log_ecn_error = true;
82module_param(log_ecn_error, bool, 0644);
83MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
84
Pravin B Shelar60d9d4c2013-06-20 00:26:31 -070085static int vxlan_net_id;
stephen hemminger553675f2013-05-16 11:35:20 +000086
Mike Rapoportafbd8ba2013-06-25 16:01:51 +030087static const u8 all_zeros_mac[ETH_ALEN];
88
stephen hemminger553675f2013-05-16 11:35:20 +000089/* per-network namespace private data for this module */
90struct vxlan_net {
91 struct list_head vxlan_list;
92 struct hlist_head sock_list[PORT_HASH_SIZE];
Stephen Hemminger1c51a912013-06-17 14:16:11 -070093 spinlock_t sock_lock;
stephen hemminger553675f2013-05-16 11:35:20 +000094};
95
Cong Wange4c7ed42013-08-31 13:44:33 +080096union vxlan_addr {
97 struct sockaddr_in sin;
98 struct sockaddr_in6 sin6;
99 struct sockaddr sa;
100};
101
David Stevens66817122013-03-15 04:35:51 +0000102struct vxlan_rdst {
Cong Wange4c7ed42013-08-31 13:44:33 +0800103 union vxlan_addr remote_ip;
David Stevens66817122013-03-15 04:35:51 +0000104 __be16 remote_port;
105 u32 remote_vni;
106 u32 remote_ifindex;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700107 struct list_head list;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300108 struct rcu_head rcu;
David Stevens66817122013-03-15 04:35:51 +0000109};
110
stephen hemmingerd3428942012-10-01 12:32:35 +0000111/* Forwarding table entry */
112struct vxlan_fdb {
113 struct hlist_node hlist; /* linked list of entries */
114 struct rcu_head rcu;
115 unsigned long updated; /* jiffies */
116 unsigned long used;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700117 struct list_head remotes;
stephen hemmingerd3428942012-10-01 12:32:35 +0000118 u16 state; /* see ndm_state */
David Stevensae884082013-04-19 00:36:26 +0000119 u8 flags; /* see ndm_flags */
stephen hemmingerd3428942012-10-01 12:32:35 +0000120 u8 eth_addr[ETH_ALEN];
121};
122
stephen hemmingerd3428942012-10-01 12:32:35 +0000123/* Pseudo network device */
124struct vxlan_dev {
stephen hemminger553675f2013-05-16 11:35:20 +0000125 struct hlist_node hlist; /* vni hash table */
126 struct list_head next; /* vxlan's per namespace list */
127 struct vxlan_sock *vn_sock; /* listening socket */
stephen hemmingerd3428942012-10-01 12:32:35 +0000128 struct net_device *dev;
Atzm Watanabec7995c42013-04-16 02:50:52 +0000129 struct vxlan_rdst default_dst; /* default destination */
Cong Wange4c7ed42013-08-31 13:44:33 +0800130 union vxlan_addr saddr; /* source address */
stephen hemminger823aa872013-04-27 11:31:57 +0000131 __be16 dst_port;
stephen hemminger05f47d62012-10-09 20:35:50 +0000132 __u16 port_min; /* source port range */
133 __u16 port_max;
stephen hemmingerd3428942012-10-01 12:32:35 +0000134 __u8 tos; /* TOS override */
135 __u8 ttl;
David Stevense4f67ad2012-11-20 02:50:14 +0000136 u32 flags; /* VXLAN_F_* below */
stephen hemmingerd3428942012-10-01 12:32:35 +0000137
Stephen Hemminger1c51a912013-06-17 14:16:11 -0700138 struct work_struct sock_work;
stephen hemminger3fc2de22013-07-18 08:40:15 -0700139 struct work_struct igmp_join;
140 struct work_struct igmp_leave;
Stephen Hemminger1c51a912013-06-17 14:16:11 -0700141
stephen hemmingerd3428942012-10-01 12:32:35 +0000142 unsigned long age_interval;
143 struct timer_list age_timer;
144 spinlock_t hash_lock;
145 unsigned int addrcnt;
146 unsigned int addrmax;
stephen hemmingerd3428942012-10-01 12:32:35 +0000147
148 struct hlist_head fdb_head[FDB_HASH_SIZE];
149};
150
David Stevense4f67ad2012-11-20 02:50:14 +0000151#define VXLAN_F_LEARN 0x01
152#define VXLAN_F_PROXY 0x02
153#define VXLAN_F_RSC 0x04
154#define VXLAN_F_L2MISS 0x08
155#define VXLAN_F_L3MISS 0x10
Cong Wange4c7ed42013-08-31 13:44:33 +0800156#define VXLAN_F_IPV6 0x20 /* internal flag */
David Stevense4f67ad2012-11-20 02:50:14 +0000157
stephen hemmingerd3428942012-10-01 12:32:35 +0000158/* salt for hash table */
159static u32 vxlan_salt __read_mostly;
Stephen Hemminger758c57d2013-06-17 14:16:09 -0700160static struct workqueue_struct *vxlan_wq;
stephen hemmingerd3428942012-10-01 12:32:35 +0000161
Stephen Hemminger1c51a912013-06-17 14:16:11 -0700162static void vxlan_sock_work(struct work_struct *work);
163
Cong Wange4c7ed42013-08-31 13:44:33 +0800164#if IS_ENABLED(CONFIG_IPV6)
165static inline
166bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
167{
168 if (a->sa.sa_family != b->sa.sa_family)
169 return false;
170 if (a->sa.sa_family == AF_INET6)
171 return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr);
172 else
173 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
174}
175
176static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
177{
178 if (ipa->sa.sa_family == AF_INET6)
179 return ipv6_addr_any(&ipa->sin6.sin6_addr);
180 else
181 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
182}
183
184static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
185{
186 if (ipa->sa.sa_family == AF_INET6)
187 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
188 else
189 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
190}
191
192static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
193{
194 if (nla_len(nla) >= sizeof(struct in6_addr)) {
195 nla_memcpy(&ip->sin6.sin6_addr, nla, sizeof(struct in6_addr));
196 ip->sa.sa_family = AF_INET6;
197 return 0;
198 } else if (nla_len(nla) >= sizeof(__be32)) {
199 ip->sin.sin_addr.s_addr = nla_get_be32(nla);
200 ip->sa.sa_family = AF_INET;
201 return 0;
202 } else {
203 return -EAFNOSUPPORT;
204 }
205}
206
207static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
208 const union vxlan_addr *ip)
209{
210 if (ip->sa.sa_family == AF_INET6)
211 return nla_put(skb, attr, sizeof(struct in6_addr), &ip->sin6.sin6_addr);
212 else
213 return nla_put_be32(skb, attr, ip->sin.sin_addr.s_addr);
214}
215
216#else /* !CONFIG_IPV6 */
217
218static inline
219bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
220{
221 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
222}
223
224static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
225{
226 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
227}
228
229static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
230{
231 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
232}
233
234static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
235{
236 if (nla_len(nla) >= sizeof(struct in6_addr)) {
237 return -EAFNOSUPPORT;
238 } else if (nla_len(nla) >= sizeof(__be32)) {
239 ip->sin.sin_addr.s_addr = nla_get_be32(nla);
240 ip->sa.sa_family = AF_INET;
241 return 0;
242 } else {
243 return -EAFNOSUPPORT;
244 }
245}
246
247static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
248 const union vxlan_addr *ip)
249{
250 return nla_put_be32(skb, attr, ip->sin.sin_addr.s_addr);
251}
252#endif
253
stephen hemminger553675f2013-05-16 11:35:20 +0000254/* Virtual Network hash table head */
255static inline struct hlist_head *vni_head(struct vxlan_sock *vs, u32 id)
256{
257 return &vs->vni_list[hash_32(id, VNI_HASH_BITS)];
258}
259
260/* Socket hash table head */
261static inline struct hlist_head *vs_head(struct net *net, __be16 port)
stephen hemmingerd3428942012-10-01 12:32:35 +0000262{
263 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
264
stephen hemminger553675f2013-05-16 11:35:20 +0000265 return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)];
266}
267
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700268/* First remote destination for a forwarding entry.
269 * Guaranteed to be non-NULL because remotes are never deleted.
270 */
stephen hemminger5ca54612013-08-04 17:17:39 -0700271static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700272{
stephen hemminger5ca54612013-08-04 17:17:39 -0700273 return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list);
274}
275
276static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb)
277{
278 return list_first_entry(&fdb->remotes, struct vxlan_rdst, list);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700279}
280
stephen hemminger553675f2013-05-16 11:35:20 +0000281/* Find VXLAN socket based on network namespace and UDP port */
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -0700282static struct vxlan_sock *vxlan_find_sock(struct net *net, __be16 port)
stephen hemminger553675f2013-05-16 11:35:20 +0000283{
284 struct vxlan_sock *vs;
285
286 hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) {
287 if (inet_sk(vs->sock->sk)->inet_sport == port)
288 return vs;
289 }
290 return NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000291}
292
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700293static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, u32 id)
stephen hemmingerd3428942012-10-01 12:32:35 +0000294{
295 struct vxlan_dev *vxlan;
stephen hemmingerd3428942012-10-01 12:32:35 +0000296
stephen hemminger553675f2013-05-16 11:35:20 +0000297 hlist_for_each_entry_rcu(vxlan, vni_head(vs, id), hlist) {
Atzm Watanabec7995c42013-04-16 02:50:52 +0000298 if (vxlan->default_dst.remote_vni == id)
stephen hemmingerd3428942012-10-01 12:32:35 +0000299 return vxlan;
300 }
301
302 return NULL;
303}
304
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700305/* Look up VNI in a per net namespace table */
306static struct vxlan_dev *vxlan_find_vni(struct net *net, u32 id, __be16 port)
307{
308 struct vxlan_sock *vs;
309
310 vs = vxlan_find_sock(net, port);
311 if (!vs)
312 return NULL;
313
314 return vxlan_vs_find_vni(vs, id);
315}
316
stephen hemmingerd3428942012-10-01 12:32:35 +0000317/* Fill in neighbour message in skbuff. */
318static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan,
Stephen Hemminger234f5b72013-06-17 14:16:41 -0700319 const struct vxlan_fdb *fdb,
320 u32 portid, u32 seq, int type, unsigned int flags,
321 const struct vxlan_rdst *rdst)
stephen hemmingerd3428942012-10-01 12:32:35 +0000322{
323 unsigned long now = jiffies;
324 struct nda_cacheinfo ci;
325 struct nlmsghdr *nlh;
326 struct ndmsg *ndm;
David Stevense4f67ad2012-11-20 02:50:14 +0000327 bool send_ip, send_eth;
stephen hemmingerd3428942012-10-01 12:32:35 +0000328
329 nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags);
330 if (nlh == NULL)
331 return -EMSGSIZE;
332
333 ndm = nlmsg_data(nlh);
334 memset(ndm, 0, sizeof(*ndm));
David Stevense4f67ad2012-11-20 02:50:14 +0000335
336 send_eth = send_ip = true;
337
338 if (type == RTM_GETNEIGH) {
339 ndm->ndm_family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +0800340 send_ip = !vxlan_addr_any(&rdst->remote_ip);
David Stevense4f67ad2012-11-20 02:50:14 +0000341 send_eth = !is_zero_ether_addr(fdb->eth_addr);
342 } else
343 ndm->ndm_family = AF_BRIDGE;
stephen hemmingerd3428942012-10-01 12:32:35 +0000344 ndm->ndm_state = fdb->state;
345 ndm->ndm_ifindex = vxlan->dev->ifindex;
David Stevensae884082013-04-19 00:36:26 +0000346 ndm->ndm_flags = fdb->flags;
stephen hemmingerd3428942012-10-01 12:32:35 +0000347 ndm->ndm_type = NDA_DST;
348
David Stevense4f67ad2012-11-20 02:50:14 +0000349 if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000350 goto nla_put_failure;
351
Cong Wange4c7ed42013-08-31 13:44:33 +0800352 if (send_ip && vxlan_nla_put_addr(skb, NDA_DST, &rdst->remote_ip))
David Stevens66817122013-03-15 04:35:51 +0000353 goto nla_put_failure;
354
stephen hemminger823aa872013-04-27 11:31:57 +0000355 if (rdst->remote_port && rdst->remote_port != vxlan->dst_port &&
David Stevens66817122013-03-15 04:35:51 +0000356 nla_put_be16(skb, NDA_PORT, rdst->remote_port))
357 goto nla_put_failure;
Atzm Watanabec7995c42013-04-16 02:50:52 +0000358 if (rdst->remote_vni != vxlan->default_dst.remote_vni &&
Pravin B Shelar60d9d4c2013-06-20 00:26:31 -0700359 nla_put_u32(skb, NDA_VNI, rdst->remote_vni))
David Stevens66817122013-03-15 04:35:51 +0000360 goto nla_put_failure;
361 if (rdst->remote_ifindex &&
362 nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +0000363 goto nla_put_failure;
364
365 ci.ndm_used = jiffies_to_clock_t(now - fdb->used);
366 ci.ndm_confirmed = 0;
367 ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated);
368 ci.ndm_refcnt = 0;
369
370 if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci))
371 goto nla_put_failure;
372
373 return nlmsg_end(skb, nlh);
374
375nla_put_failure:
376 nlmsg_cancel(skb, nlh);
377 return -EMSGSIZE;
378}
379
380static inline size_t vxlan_nlmsg_size(void)
381{
382 return NLMSG_ALIGN(sizeof(struct ndmsg))
383 + nla_total_size(ETH_ALEN) /* NDA_LLADDR */
Cong Wange4c7ed42013-08-31 13:44:33 +0800384 + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */
stephen hemminger73cf3312013-04-27 11:31:54 +0000385 + nla_total_size(sizeof(__be16)) /* NDA_PORT */
David Stevens66817122013-03-15 04:35:51 +0000386 + nla_total_size(sizeof(__be32)) /* NDA_VNI */
387 + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */
stephen hemmingerd3428942012-10-01 12:32:35 +0000388 + nla_total_size(sizeof(struct nda_cacheinfo));
389}
390
391static void vxlan_fdb_notify(struct vxlan_dev *vxlan,
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700392 struct vxlan_fdb *fdb, int type)
stephen hemmingerd3428942012-10-01 12:32:35 +0000393{
394 struct net *net = dev_net(vxlan->dev);
395 struct sk_buff *skb;
396 int err = -ENOBUFS;
397
398 skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC);
399 if (skb == NULL)
400 goto errout;
401
stephen hemminger5ca54612013-08-04 17:17:39 -0700402 err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0,
403 first_remote_rtnl(fdb));
stephen hemmingerd3428942012-10-01 12:32:35 +0000404 if (err < 0) {
405 /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */
406 WARN_ON(err == -EMSGSIZE);
407 kfree_skb(skb);
408 goto errout;
409 }
410
411 rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC);
412 return;
413errout:
414 if (err < 0)
415 rtnl_set_sk_err(net, RTNLGRP_NEIGH, err);
416}
417
Cong Wange4c7ed42013-08-31 13:44:33 +0800418static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa)
David Stevense4f67ad2012-11-20 02:50:14 +0000419{
420 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700421 struct vxlan_fdb f = {
422 .state = NUD_STALE,
423 };
424 struct vxlan_rdst remote = {
Cong Wange4c7ed42013-08-31 13:44:33 +0800425 .remote_ip = *ipa, /* goes to NDA_DST */
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700426 .remote_vni = VXLAN_N_VID,
427 };
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700428
429 INIT_LIST_HEAD(&f.remotes);
430 list_add_rcu(&remote.list, &f.remotes);
David Stevense4f67ad2012-11-20 02:50:14 +0000431
432 vxlan_fdb_notify(vxlan, &f, RTM_GETNEIGH);
433}
434
435static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN])
436{
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700437 struct vxlan_fdb f = {
438 .state = NUD_STALE,
439 };
David Stevense4f67ad2012-11-20 02:50:14 +0000440
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700441 INIT_LIST_HEAD(&f.remotes);
David Stevense4f67ad2012-11-20 02:50:14 +0000442 memcpy(f.eth_addr, eth_addr, ETH_ALEN);
443
444 vxlan_fdb_notify(vxlan, &f, RTM_GETNEIGH);
445}
446
stephen hemmingerd3428942012-10-01 12:32:35 +0000447/* Hash Ethernet address */
448static u32 eth_hash(const unsigned char *addr)
449{
450 u64 value = get_unaligned((u64 *)addr);
451
452 /* only want 6 bytes */
453#ifdef __BIG_ENDIAN
stephen hemmingerd3428942012-10-01 12:32:35 +0000454 value >>= 16;
stephen hemminger321fb992012-10-09 20:35:47 +0000455#else
456 value <<= 16;
stephen hemmingerd3428942012-10-01 12:32:35 +0000457#endif
458 return hash_64(value, FDB_HASH_BITS);
459}
460
461/* Hash chain to use given mac address */
462static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan,
463 const u8 *mac)
464{
465 return &vxlan->fdb_head[eth_hash(mac)];
466}
467
468/* Look up Ethernet address in forwarding table */
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000469static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan,
stephen hemmingerd3428942012-10-01 12:32:35 +0000470 const u8 *mac)
471
472{
473 struct hlist_head *head = vxlan_fdb_head(vxlan, mac);
474 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000475
Sasha Levinb67bfe02013-02-27 17:06:00 -0800476 hlist_for_each_entry_rcu(f, head, hlist) {
Joe Perches7367d0b2013-09-01 11:51:23 -0700477 if (ether_addr_equal(mac, f->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000478 return f;
479 }
480
481 return NULL;
482}
483
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000484static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan,
485 const u8 *mac)
486{
487 struct vxlan_fdb *f;
488
489 f = __vxlan_find_mac(vxlan, mac);
490 if (f)
491 f->used = jiffies;
492
493 return f;
494}
495
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300496/* caller should hold vxlan->hash_lock */
497static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800498 union vxlan_addr *ip, __be16 port,
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300499 __u32 vni, __u32 ifindex)
500{
501 struct vxlan_rdst *rd;
502
503 list_for_each_entry(rd, &f->remotes, list) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800504 if (vxlan_addr_equal(&rd->remote_ip, ip) &&
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300505 rd->remote_port == port &&
506 rd->remote_vni == vni &&
507 rd->remote_ifindex == ifindex)
508 return rd;
509 }
510
511 return NULL;
512}
513
Thomas Richter906dc182013-07-19 17:20:07 +0200514/* Replace destination of unicast mac */
515static int vxlan_fdb_replace(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800516 union vxlan_addr *ip, __be16 port, __u32 vni, __u32 ifindex)
Thomas Richter906dc182013-07-19 17:20:07 +0200517{
518 struct vxlan_rdst *rd;
519
520 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
521 if (rd)
522 return 0;
523
524 rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list);
525 if (!rd)
526 return 0;
Cong Wange4c7ed42013-08-31 13:44:33 +0800527 rd->remote_ip = *ip;
Thomas Richter906dc182013-07-19 17:20:07 +0200528 rd->remote_port = port;
529 rd->remote_vni = vni;
530 rd->remote_ifindex = ifindex;
531 return 1;
532}
533
David Stevens66817122013-03-15 04:35:51 +0000534/* Add/update destinations for multicast */
535static int vxlan_fdb_append(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800536 union vxlan_addr *ip, __be16 port, __u32 vni, __u32 ifindex)
David Stevens66817122013-03-15 04:35:51 +0000537{
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700538 struct vxlan_rdst *rd;
David Stevens66817122013-03-15 04:35:51 +0000539
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300540 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
541 if (rd)
542 return 0;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700543
David Stevens66817122013-03-15 04:35:51 +0000544 rd = kmalloc(sizeof(*rd), GFP_ATOMIC);
545 if (rd == NULL)
546 return -ENOBUFS;
Cong Wange4c7ed42013-08-31 13:44:33 +0800547 rd->remote_ip = *ip;
David Stevens66817122013-03-15 04:35:51 +0000548 rd->remote_port = port;
549 rd->remote_vni = vni;
550 rd->remote_ifindex = ifindex;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700551
552 list_add_tail_rcu(&rd->list, &f->remotes);
553
David Stevens66817122013-03-15 04:35:51 +0000554 return 1;
555}
556
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700557/* Notify netdevs that UDP port started listening */
558static void vxlan_notify_add_rx_port(struct sock *sk)
559{
560 struct net_device *dev;
561 struct net *net = sock_net(sk);
562 sa_family_t sa_family = sk->sk_family;
Joseph Gasparakis35e42372013-09-13 07:34:13 -0700563 __be16 port = inet_sk(sk)->inet_sport;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700564
565 rcu_read_lock();
566 for_each_netdev_rcu(net, dev) {
567 if (dev->netdev_ops->ndo_add_vxlan_port)
568 dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
569 port);
570 }
571 rcu_read_unlock();
572}
573
574/* Notify netdevs that UDP port is no more listening */
575static void vxlan_notify_del_rx_port(struct sock *sk)
576{
577 struct net_device *dev;
578 struct net *net = sock_net(sk);
579 sa_family_t sa_family = sk->sk_family;
Joseph Gasparakis35e42372013-09-13 07:34:13 -0700580 __be16 port = inet_sk(sk)->inet_sport;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700581
582 rcu_read_lock();
583 for_each_netdev_rcu(net, dev) {
584 if (dev->netdev_ops->ndo_del_vxlan_port)
585 dev->netdev_ops->ndo_del_vxlan_port(dev, sa_family,
586 port);
587 }
588 rcu_read_unlock();
589}
590
stephen hemmingerd3428942012-10-01 12:32:35 +0000591/* Add new entry to forwarding table -- assumes lock held */
592static int vxlan_fdb_create(struct vxlan_dev *vxlan,
Cong Wange4c7ed42013-08-31 13:44:33 +0800593 const u8 *mac, union vxlan_addr *ip,
David Stevens66817122013-03-15 04:35:51 +0000594 __u16 state, __u16 flags,
stephen hemminger73cf3312013-04-27 11:31:54 +0000595 __be16 port, __u32 vni, __u32 ifindex,
David Stevensae884082013-04-19 00:36:26 +0000596 __u8 ndm_flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000597{
598 struct vxlan_fdb *f;
599 int notify = 0;
600
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000601 f = __vxlan_find_mac(vxlan, mac);
stephen hemmingerd3428942012-10-01 12:32:35 +0000602 if (f) {
603 if (flags & NLM_F_EXCL) {
604 netdev_dbg(vxlan->dev,
605 "lost race to create %pM\n", mac);
606 return -EEXIST;
607 }
608 if (f->state != state) {
609 f->state = state;
610 f->updated = jiffies;
611 notify = 1;
612 }
David Stevensae884082013-04-19 00:36:26 +0000613 if (f->flags != ndm_flags) {
614 f->flags = ndm_flags;
615 f->updated = jiffies;
616 notify = 1;
617 }
Thomas Richter906dc182013-07-19 17:20:07 +0200618 if ((flags & NLM_F_REPLACE)) {
619 /* Only change unicasts */
620 if (!(is_multicast_ether_addr(f->eth_addr) ||
621 is_zero_ether_addr(f->eth_addr))) {
622 int rc = vxlan_fdb_replace(f, ip, port, vni,
623 ifindex);
624
625 if (rc < 0)
626 return rc;
627 notify |= rc;
628 } else
629 return -EOPNOTSUPP;
630 }
David Stevens66817122013-03-15 04:35:51 +0000631 if ((flags & NLM_F_APPEND) &&
Mike Rapoport58e4c762013-06-25 16:01:56 +0300632 (is_multicast_ether_addr(f->eth_addr) ||
633 is_zero_ether_addr(f->eth_addr))) {
David Stevens66817122013-03-15 04:35:51 +0000634 int rc = vxlan_fdb_append(f, ip, port, vni, ifindex);
635
636 if (rc < 0)
637 return rc;
638 notify |= rc;
639 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000640 } else {
641 if (!(flags & NLM_F_CREATE))
642 return -ENOENT;
643
644 if (vxlan->addrmax && vxlan->addrcnt >= vxlan->addrmax)
645 return -ENOSPC;
646
Thomas Richter906dc182013-07-19 17:20:07 +0200647 /* Disallow replace to add a multicast entry */
648 if ((flags & NLM_F_REPLACE) &&
649 (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac)))
650 return -EOPNOTSUPP;
651
Cong Wange4c7ed42013-08-31 13:44:33 +0800652 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
stephen hemmingerd3428942012-10-01 12:32:35 +0000653 f = kmalloc(sizeof(*f), GFP_ATOMIC);
654 if (!f)
655 return -ENOMEM;
656
657 notify = 1;
stephen hemmingerd3428942012-10-01 12:32:35 +0000658 f->state = state;
David Stevensae884082013-04-19 00:36:26 +0000659 f->flags = ndm_flags;
stephen hemmingerd3428942012-10-01 12:32:35 +0000660 f->updated = f->used = jiffies;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700661 INIT_LIST_HEAD(&f->remotes);
stephen hemmingerd3428942012-10-01 12:32:35 +0000662 memcpy(f->eth_addr, mac, ETH_ALEN);
663
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700664 vxlan_fdb_append(f, ip, port, vni, ifindex);
665
stephen hemmingerd3428942012-10-01 12:32:35 +0000666 ++vxlan->addrcnt;
667 hlist_add_head_rcu(&f->hlist,
668 vxlan_fdb_head(vxlan, mac));
669 }
670
671 if (notify)
672 vxlan_fdb_notify(vxlan, f, RTM_NEWNEIGH);
673
674 return 0;
675}
676
Wei Yongjun6706c822013-04-11 19:00:35 +0000677static void vxlan_fdb_free(struct rcu_head *head)
David Stevens66817122013-03-15 04:35:51 +0000678{
679 struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700680 struct vxlan_rdst *rd, *nd;
David Stevens66817122013-03-15 04:35:51 +0000681
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700682 list_for_each_entry_safe(rd, nd, &f->remotes, list)
David Stevens66817122013-03-15 04:35:51 +0000683 kfree(rd);
David Stevens66817122013-03-15 04:35:51 +0000684 kfree(f);
685}
686
stephen hemmingerd3428942012-10-01 12:32:35 +0000687static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f)
688{
689 netdev_dbg(vxlan->dev,
690 "delete %pM\n", f->eth_addr);
691
692 --vxlan->addrcnt;
693 vxlan_fdb_notify(vxlan, f, RTM_DELNEIGH);
694
695 hlist_del_rcu(&f->hlist);
David Stevens66817122013-03-15 04:35:51 +0000696 call_rcu(&f->rcu, vxlan_fdb_free);
stephen hemmingerd3428942012-10-01 12:32:35 +0000697}
698
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300699static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan,
Cong Wange4c7ed42013-08-31 13:44:33 +0800700 union vxlan_addr *ip, __be16 *port, u32 *vni, u32 *ifindex)
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300701{
702 struct net *net = dev_net(vxlan->dev);
Cong Wange4c7ed42013-08-31 13:44:33 +0800703 int err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300704
705 if (tb[NDA_DST]) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800706 err = vxlan_nla_get_addr(ip, tb[NDA_DST]);
707 if (err)
708 return err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300709 } else {
Cong Wange4c7ed42013-08-31 13:44:33 +0800710 union vxlan_addr *remote = &vxlan->default_dst.remote_ip;
711 if (remote->sa.sa_family == AF_INET) {
712 ip->sin.sin_addr.s_addr = htonl(INADDR_ANY);
713 ip->sa.sa_family = AF_INET;
714#if IS_ENABLED(CONFIG_IPV6)
715 } else {
716 ip->sin6.sin6_addr = in6addr_any;
717 ip->sa.sa_family = AF_INET6;
718#endif
719 }
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300720 }
721
722 if (tb[NDA_PORT]) {
723 if (nla_len(tb[NDA_PORT]) != sizeof(__be16))
724 return -EINVAL;
725 *port = nla_get_be16(tb[NDA_PORT]);
726 } else {
727 *port = vxlan->dst_port;
728 }
729
730 if (tb[NDA_VNI]) {
731 if (nla_len(tb[NDA_VNI]) != sizeof(u32))
732 return -EINVAL;
733 *vni = nla_get_u32(tb[NDA_VNI]);
734 } else {
735 *vni = vxlan->default_dst.remote_vni;
736 }
737
738 if (tb[NDA_IFINDEX]) {
739 struct net_device *tdev;
740
741 if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32))
742 return -EINVAL;
743 *ifindex = nla_get_u32(tb[NDA_IFINDEX]);
744 tdev = dev_get_by_index(net, *ifindex);
745 if (!tdev)
746 return -EADDRNOTAVAIL;
747 dev_put(tdev);
748 } else {
749 *ifindex = 0;
750 }
751
752 return 0;
753}
754
stephen hemmingerd3428942012-10-01 12:32:35 +0000755/* Add static entry (via netlink) */
756static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[],
757 struct net_device *dev,
758 const unsigned char *addr, u16 flags)
759{
760 struct vxlan_dev *vxlan = netdev_priv(dev);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300761 /* struct net *net = dev_net(vxlan->dev); */
Cong Wange4c7ed42013-08-31 13:44:33 +0800762 union vxlan_addr ip;
stephen hemminger73cf3312013-04-27 11:31:54 +0000763 __be16 port;
764 u32 vni, ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +0000765 int err;
766
767 if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) {
768 pr_info("RTM_NEWNEIGH with invalid state %#x\n",
769 ndm->ndm_state);
770 return -EINVAL;
771 }
772
773 if (tb[NDA_DST] == NULL)
774 return -EINVAL;
775
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300776 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
777 if (err)
778 return err;
David Stevens66817122013-03-15 04:35:51 +0000779
stephen hemmingerd3428942012-10-01 12:32:35 +0000780 spin_lock_bh(&vxlan->hash_lock);
Cong Wange4c7ed42013-08-31 13:44:33 +0800781 err = vxlan_fdb_create(vxlan, addr, &ip, ndm->ndm_state, flags,
stephen hemminger73cf3312013-04-27 11:31:54 +0000782 port, vni, ifindex, ndm->ndm_flags);
stephen hemmingerd3428942012-10-01 12:32:35 +0000783 spin_unlock_bh(&vxlan->hash_lock);
784
785 return err;
786}
787
788/* Delete entry (via netlink) */
Vlad Yasevich1690be62013-02-13 12:00:18 +0000789static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[],
790 struct net_device *dev,
stephen hemmingerd3428942012-10-01 12:32:35 +0000791 const unsigned char *addr)
792{
793 struct vxlan_dev *vxlan = netdev_priv(dev);
794 struct vxlan_fdb *f;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300795 struct vxlan_rdst *rd = NULL;
Cong Wange4c7ed42013-08-31 13:44:33 +0800796 union vxlan_addr ip;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300797 __be16 port;
798 u32 vni, ifindex;
799 int err;
800
801 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
802 if (err)
803 return err;
804
805 err = -ENOENT;
stephen hemmingerd3428942012-10-01 12:32:35 +0000806
807 spin_lock_bh(&vxlan->hash_lock);
808 f = vxlan_find_mac(vxlan, addr);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300809 if (!f)
810 goto out;
811
Cong Wange4c7ed42013-08-31 13:44:33 +0800812 if (!vxlan_addr_any(&ip)) {
813 rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300814 if (!rd)
815 goto out;
stephen hemmingerd3428942012-10-01 12:32:35 +0000816 }
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300817
818 err = 0;
819
820 /* remove a destination if it's not the only one on the list,
821 * otherwise destroy the fdb entry
822 */
823 if (rd && !list_is_singular(&f->remotes)) {
824 list_del_rcu(&rd->list);
Wei Yongjundcdc7a52013-08-17 07:32:09 +0800825 kfree_rcu(rd, rcu);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300826 goto out;
827 }
828
829 vxlan_fdb_destroy(vxlan, f);
830
831out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000832 spin_unlock_bh(&vxlan->hash_lock);
833
834 return err;
835}
836
837/* Dump forwarding table */
838static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb,
839 struct net_device *dev, int idx)
840{
841 struct vxlan_dev *vxlan = netdev_priv(dev);
842 unsigned int h;
843
844 for (h = 0; h < FDB_HASH_SIZE; ++h) {
845 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000846 int err;
847
Sasha Levinb67bfe02013-02-27 17:06:00 -0800848 hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) {
David Stevens66817122013-03-15 04:35:51 +0000849 struct vxlan_rdst *rd;
stephen hemmingerd3428942012-10-01 12:32:35 +0000850
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700851 if (idx < cb->args[0])
852 goto skip;
853
854 list_for_each_entry_rcu(rd, &f->remotes, list) {
David Stevens66817122013-03-15 04:35:51 +0000855 err = vxlan_fdb_info(skb, vxlan, f,
856 NETLINK_CB(cb->skb).portid,
857 cb->nlh->nlmsg_seq,
858 RTM_NEWNEIGH,
859 NLM_F_MULTI, rd);
860 if (err < 0)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700861 goto out;
David Stevens66817122013-03-15 04:35:51 +0000862 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700863skip:
864 ++idx;
stephen hemmingerd3428942012-10-01 12:32:35 +0000865 }
866 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700867out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000868 return idx;
869}
870
871/* Watch incoming packets to learn mapping between Ethernet address
872 * and Tunnel endpoint.
stephen hemminger26a41ae62013-06-17 12:09:58 -0700873 * Return true if packet is bogus and should be droppped.
stephen hemmingerd3428942012-10-01 12:32:35 +0000874 */
stephen hemminger26a41ae62013-06-17 12:09:58 -0700875static bool vxlan_snoop(struct net_device *dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800876 union vxlan_addr *src_ip, const u8 *src_mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000877{
878 struct vxlan_dev *vxlan = netdev_priv(dev);
879 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000880
881 f = vxlan_find_mac(vxlan, src_mac);
882 if (likely(f)) {
stephen hemminger5ca54612013-08-04 17:17:39 -0700883 struct vxlan_rdst *rdst = first_remote_rcu(f);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700884
Cong Wange4c7ed42013-08-31 13:44:33 +0800885 if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip)))
stephen hemminger26a41ae62013-06-17 12:09:58 -0700886 return false;
887
888 /* Don't migrate static entries, drop packets */
stephen hemmingereb064c32013-06-18 14:27:01 -0700889 if (f->state & NUD_NOARP)
stephen hemminger26a41ae62013-06-17 12:09:58 -0700890 return true;
stephen hemmingerd3428942012-10-01 12:32:35 +0000891
892 if (net_ratelimit())
893 netdev_info(dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800894 "%pM migrated from %pIS to %pIS\n",
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700895 src_mac, &rdst->remote_ip, &src_ip);
stephen hemmingerd3428942012-10-01 12:32:35 +0000896
Cong Wange4c7ed42013-08-31 13:44:33 +0800897 rdst->remote_ip = *src_ip;
stephen hemmingerd3428942012-10-01 12:32:35 +0000898 f->updated = jiffies;
Stephen Hemminger8385f502013-06-17 14:16:10 -0700899 vxlan_fdb_notify(vxlan, f, RTM_NEWNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +0000900 } else {
901 /* learned new entry */
902 spin_lock(&vxlan->hash_lock);
stephen hemminger3bf74b12013-06-17 12:09:57 -0700903
904 /* close off race between vxlan_flush and incoming packets */
905 if (netif_running(dev))
906 vxlan_fdb_create(vxlan, src_mac, src_ip,
907 NUD_REACHABLE,
908 NLM_F_EXCL|NLM_F_CREATE,
909 vxlan->dst_port,
910 vxlan->default_dst.remote_vni,
911 0, NTF_SELF);
stephen hemmingerd3428942012-10-01 12:32:35 +0000912 spin_unlock(&vxlan->hash_lock);
913 }
stephen hemminger26a41ae62013-06-17 12:09:58 -0700914
915 return false;
stephen hemmingerd3428942012-10-01 12:32:35 +0000916}
917
stephen hemmingerd3428942012-10-01 12:32:35 +0000918/* See if multicast group is already in use by other ID */
Gao feng95ab0992013-12-10 16:37:33 +0800919static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev)
stephen hemmingerd3428942012-10-01 12:32:35 +0000920{
stephen hemminger553675f2013-05-16 11:35:20 +0000921 struct vxlan_dev *vxlan;
stephen hemmingerd3428942012-10-01 12:32:35 +0000922
Gao feng95ab0992013-12-10 16:37:33 +0800923 /* The vxlan_sock is only used by dev, leaving group has
924 * no effect on other vxlan devices.
925 */
926 if (atomic_read(&dev->vn_sock->refcnt) == 1)
927 return false;
928
stephen hemminger553675f2013-05-16 11:35:20 +0000929 list_for_each_entry(vxlan, &vn->vxlan_list, next) {
Gao feng95ab0992013-12-10 16:37:33 +0800930 if (!netif_running(vxlan->dev) || vxlan == dev)
stephen hemminger553675f2013-05-16 11:35:20 +0000931 continue;
stephen hemmingerd3428942012-10-01 12:32:35 +0000932
Gao feng95ab0992013-12-10 16:37:33 +0800933 if (vxlan->vn_sock != dev->vn_sock)
934 continue;
935
936 if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip,
937 &dev->default_dst.remote_ip))
938 continue;
939
940 if (vxlan->default_dst.remote_ifindex !=
941 dev->default_dst.remote_ifindex)
942 continue;
943
944 return true;
stephen hemminger553675f2013-05-16 11:35:20 +0000945 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000946
947 return false;
948}
949
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700950static void vxlan_sock_hold(struct vxlan_sock *vs)
stephen hemmingerd3428942012-10-01 12:32:35 +0000951{
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700952 atomic_inc(&vs->refcnt);
stephen hemmingerd3428942012-10-01 12:32:35 +0000953}
954
Pravin B Shelar012a5722013-08-19 11:23:07 -0700955void vxlan_sock_release(struct vxlan_sock *vs)
stephen hemmingerd3428942012-10-01 12:32:35 +0000956{
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700957 struct sock *sk = vs->sock->sk;
958 struct net *net = sock_net(sk);
959 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Pravin B Shelar012a5722013-08-19 11:23:07 -0700960
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700961 if (!atomic_dec_and_test(&vs->refcnt))
962 return;
963
Stephen Hemminger1c51a912013-06-17 14:16:11 -0700964 spin_lock(&vn->sock_lock);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700965 hlist_del_rcu(&vs->hlist);
Pravin B Shelar559835ea2013-09-24 10:25:40 -0700966 rcu_assign_sk_user_data(vs->sock->sk, NULL);
Joseph Gasparakis53cf52752013-09-04 02:13:38 -0700967 vxlan_notify_del_rx_port(sk);
Stephen Hemminger1c51a912013-06-17 14:16:11 -0700968 spin_unlock(&vn->sock_lock);
969
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700970 queue_work(vxlan_wq, &vs->del_work);
971}
Pravin B Shelar012a5722013-08-19 11:23:07 -0700972EXPORT_SYMBOL_GPL(vxlan_sock_release);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700973
stephen hemminger3fc2de22013-07-18 08:40:15 -0700974/* Callback to update multicast group membership when first VNI on
975 * multicast asddress is brought up
976 * Done as workqueue because ip_mc_join_group acquires RTNL.
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700977 */
stephen hemminger3fc2de22013-07-18 08:40:15 -0700978static void vxlan_igmp_join(struct work_struct *work)
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700979{
stephen hemminger3fc2de22013-07-18 08:40:15 -0700980 struct vxlan_dev *vxlan = container_of(work, struct vxlan_dev, igmp_join);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -0700981 struct vxlan_sock *vs = vxlan->vn_sock;
982 struct sock *sk = vs->sock->sk;
Cong Wange4c7ed42013-08-31 13:44:33 +0800983 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
984 int ifindex = vxlan->default_dst.remote_ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +0000985
stephen hemmingerd3428942012-10-01 12:32:35 +0000986 lock_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +0800987 if (ip->sa.sa_family == AF_INET) {
988 struct ip_mreqn mreq = {
989 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
990 .imr_ifindex = ifindex,
991 };
992
993 ip_mc_join_group(sk, &mreq);
994#if IS_ENABLED(CONFIG_IPV6)
995 } else {
996 ipv6_stub->ipv6_sock_mc_join(sk, ifindex,
997 &ip->sin6.sin6_addr);
998#endif
999 }
stephen hemminger3fc2de22013-07-18 08:40:15 -07001000 release_sock(sk);
1001
Pravin B Shelar012a5722013-08-19 11:23:07 -07001002 vxlan_sock_release(vs);
stephen hemminger3fc2de22013-07-18 08:40:15 -07001003 dev_put(vxlan->dev);
1004}
1005
1006/* Inverse of vxlan_igmp_join when last VNI is brought down */
1007static void vxlan_igmp_leave(struct work_struct *work)
1008{
1009 struct vxlan_dev *vxlan = container_of(work, struct vxlan_dev, igmp_leave);
stephen hemminger3fc2de22013-07-18 08:40:15 -07001010 struct vxlan_sock *vs = vxlan->vn_sock;
1011 struct sock *sk = vs->sock->sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001012 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1013 int ifindex = vxlan->default_dst.remote_ifindex;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001014
1015 lock_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001016 if (ip->sa.sa_family == AF_INET) {
1017 struct ip_mreqn mreq = {
1018 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1019 .imr_ifindex = ifindex,
1020 };
1021
1022 ip_mc_leave_group(sk, &mreq);
1023#if IS_ENABLED(CONFIG_IPV6)
1024 } else {
1025 ipv6_stub->ipv6_sock_mc_drop(sk, ifindex,
1026 &ip->sin6.sin6_addr);
1027#endif
1028 }
1029
stephen hemmingerd3428942012-10-01 12:32:35 +00001030 release_sock(sk);
stephen hemmingerd3428942012-10-01 12:32:35 +00001031
Pravin B Shelar012a5722013-08-19 11:23:07 -07001032 vxlan_sock_release(vs);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001033 dev_put(vxlan->dev);
stephen hemmingerd3428942012-10-01 12:32:35 +00001034}
1035
1036/* Callback from net/ipv4/udp.c to receive packets */
1037static int vxlan_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
1038{
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001039 struct vxlan_sock *vs;
stephen hemmingerd3428942012-10-01 12:32:35 +00001040 struct vxlanhdr *vxh;
stephen hemminger553675f2013-05-16 11:35:20 +00001041 __be16 port;
stephen hemmingerd3428942012-10-01 12:32:35 +00001042
stephen hemmingerd3428942012-10-01 12:32:35 +00001043 /* Need Vxlan and inner Ethernet header to be present */
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001044 if (!pskb_may_pull(skb, VXLAN_HLEN))
stephen hemmingerd3428942012-10-01 12:32:35 +00001045 goto error;
1046
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001047 /* Return packets with reserved bits set */
1048 vxh = (struct vxlanhdr *)(udp_hdr(skb) + 1);
stephen hemmingerd3428942012-10-01 12:32:35 +00001049 if (vxh->vx_flags != htonl(VXLAN_FLAGS) ||
1050 (vxh->vx_vni & htonl(0xff))) {
1051 netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n",
1052 ntohl(vxh->vx_flags), ntohl(vxh->vx_vni));
1053 goto error;
1054 }
1055
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001056 if (iptunnel_pull_header(skb, VXLAN_HLEN, htons(ETH_P_TEB)))
stephen hemmingerd3428942012-10-01 12:32:35 +00001057 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001058
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001059 port = inet_sk(sk)->inet_sport;
1060
Pravin B Shelar559835ea2013-09-24 10:25:40 -07001061 vs = rcu_dereference_sk_user_data(sk);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001062 if (!vs)
stephen hemmingerd3428942012-10-01 12:32:35 +00001063 goto drop;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001064
1065 vs->rcv(vs, skb, vxh->vx_vni);
1066 return 0;
1067
1068drop:
1069 /* Consume bad packet */
1070 kfree_skb(skb);
1071 return 0;
1072
1073error:
1074 /* Return non vxlan pkt */
1075 return 1;
1076}
1077
1078static void vxlan_rcv(struct vxlan_sock *vs,
1079 struct sk_buff *skb, __be32 vx_vni)
1080{
Cong Wange4c7ed42013-08-31 13:44:33 +08001081 struct iphdr *oip = NULL;
1082 struct ipv6hdr *oip6 = NULL;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001083 struct vxlan_dev *vxlan;
1084 struct pcpu_tstats *stats;
Cong Wange4c7ed42013-08-31 13:44:33 +08001085 union vxlan_addr saddr;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001086 __u32 vni;
Cong Wange4c7ed42013-08-31 13:44:33 +08001087 int err = 0;
1088 union vxlan_addr *remote_ip;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001089
1090 vni = ntohl(vx_vni) >> 8;
1091 /* Is this VNI defined? */
1092 vxlan = vxlan_vs_find_vni(vs, vni);
1093 if (!vxlan)
1094 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001095
Cong Wange4c7ed42013-08-31 13:44:33 +08001096 remote_ip = &vxlan->default_dst.remote_ip;
David Stevense4f67ad2012-11-20 02:50:14 +00001097 skb_reset_mac_header(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001098 skb->protocol = eth_type_trans(skb, vxlan->dev);
stephen hemmingerd3428942012-10-01 12:32:35 +00001099
1100 /* Ignore packet loops (and multicast echo) */
Joe Perches7367d0b2013-09-01 11:51:23 -07001101 if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +00001102 goto drop;
1103
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001104 /* Re-examine inner Ethernet packet */
Cong Wange4c7ed42013-08-31 13:44:33 +08001105 if (remote_ip->sa.sa_family == AF_INET) {
1106 oip = ip_hdr(skb);
1107 saddr.sin.sin_addr.s_addr = oip->saddr;
1108 saddr.sa.sa_family = AF_INET;
1109#if IS_ENABLED(CONFIG_IPV6)
1110 } else {
1111 oip6 = ipv6_hdr(skb);
1112 saddr.sin6.sin6_addr = oip6->saddr;
1113 saddr.sa.sa_family = AF_INET6;
1114#endif
1115 }
1116
stephen hemminger26a41ae62013-06-17 12:09:58 -07001117 if ((vxlan->flags & VXLAN_F_LEARN) &&
Cong Wange4c7ed42013-08-31 13:44:33 +08001118 vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source))
stephen hemminger26a41ae62013-06-17 12:09:58 -07001119 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001120
stephen hemmingerd3428942012-10-01 12:32:35 +00001121 skb_reset_network_header(skb);
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00001122
1123 /* If the NIC driver gave us an encapsulated packet with
1124 * CHECKSUM_UNNECESSARY and Rx checksum feature is enabled,
1125 * leave the CHECKSUM_UNNECESSARY, the device checksummed it
1126 * for us. Otherwise force the upper layers to verify it.
1127 */
1128 if (skb->ip_summed != CHECKSUM_UNNECESSARY || !skb->encapsulation ||
1129 !(vxlan->dev->features & NETIF_F_RXCSUM))
1130 skb->ip_summed = CHECKSUM_NONE;
1131
1132 skb->encapsulation = 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00001133
Cong Wange4c7ed42013-08-31 13:44:33 +08001134 if (oip6)
1135 err = IP6_ECN_decapsulate(oip6, skb);
1136 if (oip)
1137 err = IP_ECN_decapsulate(oip, skb);
1138
stephen hemmingerd3428942012-10-01 12:32:35 +00001139 if (unlikely(err)) {
Cong Wange4c7ed42013-08-31 13:44:33 +08001140 if (log_ecn_error) {
1141 if (oip6)
1142 net_info_ratelimited("non-ECT from %pI6\n",
1143 &oip6->saddr);
1144 if (oip)
1145 net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n",
1146 &oip->saddr, oip->tos);
1147 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001148 if (err > 1) {
1149 ++vxlan->dev->stats.rx_frame_errors;
1150 ++vxlan->dev->stats.rx_errors;
1151 goto drop;
1152 }
1153 }
1154
Pravin B Shelare8171042013-03-25 14:49:46 +00001155 stats = this_cpu_ptr(vxlan->dev->tstats);
stephen hemmingerd3428942012-10-01 12:32:35 +00001156 u64_stats_update_begin(&stats->syncp);
1157 stats->rx_packets++;
1158 stats->rx_bytes += skb->len;
1159 u64_stats_update_end(&stats->syncp);
1160
1161 netif_rx(skb);
1162
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001163 return;
stephen hemmingerd3428942012-10-01 12:32:35 +00001164drop:
1165 /* Consume bad packet */
1166 kfree_skb(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001167}
1168
David Stevense4f67ad2012-11-20 02:50:14 +00001169static int arp_reduce(struct net_device *dev, struct sk_buff *skb)
1170{
1171 struct vxlan_dev *vxlan = netdev_priv(dev);
1172 struct arphdr *parp;
1173 u8 *arpptr, *sha;
1174 __be32 sip, tip;
1175 struct neighbour *n;
1176
1177 if (dev->flags & IFF_NOARP)
1178 goto out;
1179
1180 if (!pskb_may_pull(skb, arp_hdr_len(dev))) {
1181 dev->stats.tx_dropped++;
1182 goto out;
1183 }
1184 parp = arp_hdr(skb);
1185
1186 if ((parp->ar_hrd != htons(ARPHRD_ETHER) &&
1187 parp->ar_hrd != htons(ARPHRD_IEEE802)) ||
1188 parp->ar_pro != htons(ETH_P_IP) ||
1189 parp->ar_op != htons(ARPOP_REQUEST) ||
1190 parp->ar_hln != dev->addr_len ||
1191 parp->ar_pln != 4)
1192 goto out;
1193 arpptr = (u8 *)parp + sizeof(struct arphdr);
1194 sha = arpptr;
1195 arpptr += dev->addr_len; /* sha */
1196 memcpy(&sip, arpptr, sizeof(sip));
1197 arpptr += sizeof(sip);
1198 arpptr += dev->addr_len; /* tha */
1199 memcpy(&tip, arpptr, sizeof(tip));
1200
1201 if (ipv4_is_loopback(tip) ||
1202 ipv4_is_multicast(tip))
1203 goto out;
1204
1205 n = neigh_lookup(&arp_tbl, &tip, dev);
1206
1207 if (n) {
David Stevense4f67ad2012-11-20 02:50:14 +00001208 struct vxlan_fdb *f;
1209 struct sk_buff *reply;
1210
1211 if (!(n->nud_state & NUD_CONNECTED)) {
1212 neigh_release(n);
1213 goto out;
1214 }
1215
1216 f = vxlan_find_mac(vxlan, n->ha);
Cong Wange4c7ed42013-08-31 13:44:33 +08001217 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
David Stevense4f67ad2012-11-20 02:50:14 +00001218 /* bridge-local neighbor */
1219 neigh_release(n);
1220 goto out;
1221 }
1222
1223 reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
1224 n->ha, sha);
1225
1226 neigh_release(n);
1227
1228 skb_reset_mac_header(reply);
1229 __skb_pull(reply, skb_network_offset(reply));
1230 reply->ip_summed = CHECKSUM_UNNECESSARY;
1231 reply->pkt_type = PACKET_HOST;
1232
1233 if (netif_rx_ni(reply) == NET_RX_DROP)
1234 dev->stats.rx_dropped++;
Cong Wange4c7ed42013-08-31 13:44:33 +08001235 } else if (vxlan->flags & VXLAN_F_L3MISS) {
1236 union vxlan_addr ipa = {
1237 .sin.sin_addr.s_addr = tip,
1238 .sa.sa_family = AF_INET,
1239 };
1240
1241 vxlan_ip_miss(dev, &ipa);
1242 }
David Stevense4f67ad2012-11-20 02:50:14 +00001243out:
1244 consume_skb(skb);
1245 return NETDEV_TX_OK;
1246}
1247
Cong Wangf564f452013-08-31 13:44:36 +08001248#if IS_ENABLED(CONFIG_IPV6)
1249static int neigh_reduce(struct net_device *dev, struct sk_buff *skb)
1250{
1251 struct vxlan_dev *vxlan = netdev_priv(dev);
1252 struct neighbour *n;
1253 union vxlan_addr ipa;
1254 const struct ipv6hdr *iphdr;
1255 const struct in6_addr *saddr, *daddr;
1256 struct nd_msg *msg;
1257 struct inet6_dev *in6_dev = NULL;
1258
1259 in6_dev = __in6_dev_get(dev);
1260 if (!in6_dev)
1261 goto out;
1262
1263 if (!pskb_may_pull(skb, skb->len))
1264 goto out;
1265
1266 iphdr = ipv6_hdr(skb);
1267 saddr = &iphdr->saddr;
1268 daddr = &iphdr->daddr;
1269
1270 if (ipv6_addr_loopback(daddr) ||
1271 ipv6_addr_is_multicast(daddr))
1272 goto out;
1273
1274 msg = (struct nd_msg *)skb_transport_header(skb);
1275 if (msg->icmph.icmp6_code != 0 ||
1276 msg->icmph.icmp6_type != NDISC_NEIGHBOUR_SOLICITATION)
1277 goto out;
1278
1279 n = neigh_lookup(ipv6_stub->nd_tbl, daddr, dev);
1280
1281 if (n) {
1282 struct vxlan_fdb *f;
1283
1284 if (!(n->nud_state & NUD_CONNECTED)) {
1285 neigh_release(n);
1286 goto out;
1287 }
1288
1289 f = vxlan_find_mac(vxlan, n->ha);
1290 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
1291 /* bridge-local neighbor */
1292 neigh_release(n);
1293 goto out;
1294 }
1295
1296 ipv6_stub->ndisc_send_na(dev, n, saddr, &msg->target,
1297 !!in6_dev->cnf.forwarding,
1298 true, false, false);
1299 neigh_release(n);
1300 } else if (vxlan->flags & VXLAN_F_L3MISS) {
1301 ipa.sin6.sin6_addr = *daddr;
1302 ipa.sa.sa_family = AF_INET6;
1303 vxlan_ip_miss(dev, &ipa);
1304 }
1305
1306out:
1307 consume_skb(skb);
1308 return NETDEV_TX_OK;
1309}
1310#endif
1311
David Stevense4f67ad2012-11-20 02:50:14 +00001312static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb)
1313{
1314 struct vxlan_dev *vxlan = netdev_priv(dev);
1315 struct neighbour *n;
David Stevense4f67ad2012-11-20 02:50:14 +00001316
1317 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
1318 return false;
1319
1320 n = NULL;
1321 switch (ntohs(eth_hdr(skb)->h_proto)) {
1322 case ETH_P_IP:
Cong Wange15a00a2013-08-31 13:44:34 +08001323 {
1324 struct iphdr *pip;
1325
David Stevense4f67ad2012-11-20 02:50:14 +00001326 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
1327 return false;
1328 pip = ip_hdr(skb);
1329 n = neigh_lookup(&arp_tbl, &pip->daddr, dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08001330 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1331 union vxlan_addr ipa = {
1332 .sin.sin_addr.s_addr = pip->daddr,
1333 .sa.sa_family = AF_INET,
1334 };
1335
1336 vxlan_ip_miss(dev, &ipa);
1337 return false;
1338 }
1339
David Stevense4f67ad2012-11-20 02:50:14 +00001340 break;
Cong Wange15a00a2013-08-31 13:44:34 +08001341 }
1342#if IS_ENABLED(CONFIG_IPV6)
1343 case ETH_P_IPV6:
1344 {
1345 struct ipv6hdr *pip6;
1346
1347 if (!pskb_may_pull(skb, sizeof(struct ipv6hdr)))
1348 return false;
1349 pip6 = ipv6_hdr(skb);
1350 n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev);
1351 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1352 union vxlan_addr ipa = {
1353 .sin6.sin6_addr = pip6->daddr,
1354 .sa.sa_family = AF_INET6,
1355 };
1356
1357 vxlan_ip_miss(dev, &ipa);
1358 return false;
1359 }
1360
1361 break;
1362 }
1363#endif
David Stevense4f67ad2012-11-20 02:50:14 +00001364 default:
1365 return false;
1366 }
1367
1368 if (n) {
1369 bool diff;
1370
Joe Perches7367d0b2013-09-01 11:51:23 -07001371 diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha);
David Stevense4f67ad2012-11-20 02:50:14 +00001372 if (diff) {
1373 memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest,
1374 dev->addr_len);
1375 memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len);
1376 }
1377 neigh_release(n);
1378 return diff;
Cong Wange4c7ed42013-08-31 13:44:33 +08001379 }
1380
David Stevense4f67ad2012-11-20 02:50:14 +00001381 return false;
1382}
1383
stephen hemminger553675f2013-05-16 11:35:20 +00001384static void vxlan_sock_put(struct sk_buff *skb)
stephen hemminger1cad8712012-10-09 20:35:49 +00001385{
1386 sock_put(skb->sk);
1387}
1388
1389/* On transmit, associate with the tunnel socket */
Pravin B Shelar49560532013-08-19 11:23:17 -07001390static void vxlan_set_owner(struct sock *sk, struct sk_buff *skb)
stephen hemminger1cad8712012-10-09 20:35:49 +00001391{
stephen hemminger1cad8712012-10-09 20:35:49 +00001392 skb_orphan(skb);
1393 sock_hold(sk);
1394 skb->sk = sk;
stephen hemminger553675f2013-05-16 11:35:20 +00001395 skb->destructor = vxlan_sock_put;
stephen hemminger1cad8712012-10-09 20:35:49 +00001396}
1397
stephen hemminger05f47d62012-10-09 20:35:50 +00001398/* Compute source port for outgoing packet
1399 * first choice to use L4 flow hash since it will spread
1400 * better and maybe available from hardware
1401 * secondary choice is to use jhash on the Ethernet header
1402 */
Pravin B Shelar49560532013-08-19 11:23:17 -07001403__be16 vxlan_src_port(__u16 port_min, __u16 port_max, struct sk_buff *skb)
stephen hemminger05f47d62012-10-09 20:35:50 +00001404{
Pravin B Shelar49560532013-08-19 11:23:17 -07001405 unsigned int range = (port_max - port_min) + 1;
stephen hemminger05f47d62012-10-09 20:35:50 +00001406 u32 hash;
1407
1408 hash = skb_get_rxhash(skb);
1409 if (!hash)
1410 hash = jhash(skb->data, 2 * ETH_ALEN,
1411 (__force u32) skb->protocol);
1412
Pravin B Shelar49560532013-08-19 11:23:17 -07001413 return htons((((u64) hash * range) >> 32) + port_min);
stephen hemminger05f47d62012-10-09 20:35:50 +00001414}
Pravin B Shelar49560532013-08-19 11:23:17 -07001415EXPORT_SYMBOL_GPL(vxlan_src_port);
stephen hemminger05f47d62012-10-09 20:35:50 +00001416
Pravin B Shelar05c0db02013-03-07 13:22:36 +00001417static int handle_offloads(struct sk_buff *skb)
1418{
1419 if (skb_is_gso(skb)) {
1420 int err = skb_unclone(skb, GFP_ATOMIC);
1421 if (unlikely(err))
1422 return err;
1423
Dmitry Kravkovf6ace502013-04-28 08:16:01 +00001424 skb_shinfo(skb)->gso_type |= SKB_GSO_UDP_TUNNEL;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00001425 } else if (skb->ip_summed != CHECKSUM_PARTIAL)
1426 skb->ip_summed = CHECKSUM_NONE;
1427
1428 return 0;
1429}
1430
Cong Wange4c7ed42013-08-31 13:44:33 +08001431#if IS_ENABLED(CONFIG_IPV6)
Nicolas Dichtel11796182013-09-02 15:34:55 +02001432static int vxlan6_xmit_skb(struct vxlan_sock *vs,
Cong Wange4c7ed42013-08-31 13:44:33 +08001433 struct dst_entry *dst, struct sk_buff *skb,
1434 struct net_device *dev, struct in6_addr *saddr,
1435 struct in6_addr *daddr, __u8 prio, __u8 ttl,
1436 __be16 src_port, __be16 dst_port, __be32 vni)
1437{
1438 struct ipv6hdr *ip6h;
1439 struct vxlanhdr *vxh;
1440 struct udphdr *uh;
1441 int min_headroom;
1442 int err;
1443
1444 if (!skb->encapsulation) {
1445 skb_reset_inner_headers(skb);
1446 skb->encapsulation = 1;
1447 }
1448
Nicolas Dichtel963a88b2013-09-02 15:34:57 +02001449 skb_scrub_packet(skb, false);
1450
Cong Wange4c7ed42013-08-31 13:44:33 +08001451 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
1452 + VXLAN_HLEN + sizeof(struct ipv6hdr)
1453 + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
1454
1455 /* Need space for new headers (invalidates iph ptr) */
1456 err = skb_cow_head(skb, min_headroom);
1457 if (unlikely(err))
1458 return err;
1459
1460 if (vlan_tx_tag_present(skb)) {
1461 if (WARN_ON(!__vlan_put_tag(skb,
1462 skb->vlan_proto,
1463 vlan_tx_tag_get(skb))))
1464 return -ENOMEM;
1465
1466 skb->vlan_tci = 0;
1467 }
1468
1469 vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh));
1470 vxh->vx_flags = htonl(VXLAN_FLAGS);
1471 vxh->vx_vni = vni;
1472
1473 __skb_push(skb, sizeof(*uh));
1474 skb_reset_transport_header(skb);
1475 uh = udp_hdr(skb);
1476
1477 uh->dest = dst_port;
1478 uh->source = src_port;
1479
1480 uh->len = htons(skb->len);
1481 uh->check = 0;
1482
1483 memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
1484 IPCB(skb)->flags &= ~(IPSKB_XFRM_TUNNEL_SIZE | IPSKB_XFRM_TRANSFORMED |
1485 IPSKB_REROUTED);
Cong Wange4c7ed42013-08-31 13:44:33 +08001486 skb_dst_set(skb, dst);
1487
1488 if (!skb_is_gso(skb) && !(dst->dev->features & NETIF_F_IPV6_CSUM)) {
1489 __wsum csum = skb_checksum(skb, 0, skb->len, 0);
1490 skb->ip_summed = CHECKSUM_UNNECESSARY;
1491 uh->check = csum_ipv6_magic(saddr, daddr, skb->len,
1492 IPPROTO_UDP, csum);
1493 if (uh->check == 0)
1494 uh->check = CSUM_MANGLED_0;
1495 } else {
1496 skb->ip_summed = CHECKSUM_PARTIAL;
1497 skb->csum_start = skb_transport_header(skb) - skb->head;
1498 skb->csum_offset = offsetof(struct udphdr, check);
1499 uh->check = ~csum_ipv6_magic(saddr, daddr,
1500 skb->len, IPPROTO_UDP, 0);
1501 }
1502
1503 __skb_push(skb, sizeof(*ip6h));
1504 skb_reset_network_header(skb);
1505 ip6h = ipv6_hdr(skb);
1506 ip6h->version = 6;
1507 ip6h->priority = prio;
1508 ip6h->flow_lbl[0] = 0;
1509 ip6h->flow_lbl[1] = 0;
1510 ip6h->flow_lbl[2] = 0;
1511 ip6h->payload_len = htons(skb->len);
1512 ip6h->nexthdr = IPPROTO_UDP;
1513 ip6h->hop_limit = ttl;
1514 ip6h->daddr = *daddr;
1515 ip6h->saddr = *saddr;
1516
1517 vxlan_set_owner(vs->sock->sk, skb);
1518
1519 err = handle_offloads(skb);
1520 if (err)
1521 return err;
1522
1523 ip6tunnel_xmit(skb, dev);
1524 return 0;
1525}
1526#endif
1527
Nicolas Dichtel11796182013-09-02 15:34:55 +02001528int vxlan_xmit_skb(struct vxlan_sock *vs,
Pravin B Shelar49560532013-08-19 11:23:17 -07001529 struct rtable *rt, struct sk_buff *skb,
1530 __be32 src, __be32 dst, __u8 tos, __u8 ttl, __be16 df,
1531 __be16 src_port, __be16 dst_port, __be32 vni)
1532{
1533 struct vxlanhdr *vxh;
1534 struct udphdr *uh;
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001535 int min_headroom;
Pravin B Shelar49560532013-08-19 11:23:17 -07001536 int err;
1537
1538 if (!skb->encapsulation) {
1539 skb_reset_inner_headers(skb);
1540 skb->encapsulation = 1;
1541 }
1542
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001543 min_headroom = LL_RESERVED_SPACE(rt->dst.dev) + rt->dst.header_len
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07001544 + VXLAN_HLEN + sizeof(struct iphdr)
1545 + (vlan_tx_tag_present(skb) ? VLAN_HLEN : 0);
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001546
1547 /* Need space for new headers (invalidates iph ptr) */
1548 err = skb_cow_head(skb, min_headroom);
1549 if (unlikely(err))
1550 return err;
1551
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07001552 if (vlan_tx_tag_present(skb)) {
1553 if (WARN_ON(!__vlan_put_tag(skb,
1554 skb->vlan_proto,
1555 vlan_tx_tag_get(skb))))
1556 return -ENOMEM;
1557
1558 skb->vlan_tci = 0;
1559 }
1560
Pravin B Shelar49560532013-08-19 11:23:17 -07001561 vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh));
1562 vxh->vx_flags = htonl(VXLAN_FLAGS);
1563 vxh->vx_vni = vni;
1564
1565 __skb_push(skb, sizeof(*uh));
1566 skb_reset_transport_header(skb);
1567 uh = udp_hdr(skb);
1568
1569 uh->dest = dst_port;
1570 uh->source = src_port;
1571
1572 uh->len = htons(skb->len);
1573 uh->check = 0;
1574
1575 vxlan_set_owner(vs->sock->sk, skb);
1576
1577 err = handle_offloads(skb);
1578 if (err)
1579 return err;
1580
Nicolas Dichtel963a88b2013-09-02 15:34:57 +02001581 return iptunnel_xmit(rt, skb, src, dst, IPPROTO_UDP, tos, ttl, df,
1582 false);
Pravin B Shelar49560532013-08-19 11:23:17 -07001583}
1584EXPORT_SYMBOL_GPL(vxlan_xmit_skb);
1585
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001586/* Bypass encapsulation if the destination is local */
1587static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan,
1588 struct vxlan_dev *dst_vxlan)
1589{
1590 struct pcpu_tstats *tx_stats = this_cpu_ptr(src_vxlan->dev->tstats);
1591 struct pcpu_tstats *rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats);
Cong Wange4c7ed42013-08-31 13:44:33 +08001592 union vxlan_addr loopback;
1593 union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001594
1595 skb->pkt_type = PACKET_HOST;
1596 skb->encapsulation = 0;
1597 skb->dev = dst_vxlan->dev;
1598 __skb_pull(skb, skb_network_offset(skb));
1599
Cong Wange4c7ed42013-08-31 13:44:33 +08001600 if (remote_ip->sa.sa_family == AF_INET) {
1601 loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1602 loopback.sa.sa_family = AF_INET;
1603#if IS_ENABLED(CONFIG_IPV6)
1604 } else {
1605 loopback.sin6.sin6_addr = in6addr_loopback;
1606 loopback.sa.sa_family = AF_INET6;
1607#endif
1608 }
1609
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001610 if (dst_vxlan->flags & VXLAN_F_LEARN)
Cong Wange4c7ed42013-08-31 13:44:33 +08001611 vxlan_snoop(skb->dev, &loopback, eth_hdr(skb)->h_source);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001612
1613 u64_stats_update_begin(&tx_stats->syncp);
1614 tx_stats->tx_packets++;
1615 tx_stats->tx_bytes += skb->len;
1616 u64_stats_update_end(&tx_stats->syncp);
1617
1618 if (netif_rx(skb) == NET_RX_SUCCESS) {
1619 u64_stats_update_begin(&rx_stats->syncp);
1620 rx_stats->rx_packets++;
1621 rx_stats->rx_bytes += skb->len;
1622 u64_stats_update_end(&rx_stats->syncp);
1623 } else {
1624 skb->dev->stats.rx_dropped++;
1625 }
1626}
1627
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001628static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev,
1629 struct vxlan_rdst *rdst, bool did_rsc)
stephen hemmingerd3428942012-10-01 12:32:35 +00001630{
1631 struct vxlan_dev *vxlan = netdev_priv(dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08001632 struct rtable *rt = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001633 const struct iphdr *old_iph;
stephen hemmingerd3428942012-10-01 12:32:35 +00001634 struct flowi4 fl4;
Cong Wange4c7ed42013-08-31 13:44:33 +08001635 union vxlan_addr *dst;
1636 __be16 src_port = 0, dst_port;
Stephen Hemminger234f5b72013-06-17 14:16:41 -07001637 u32 vni;
stephen hemmingerd3428942012-10-01 12:32:35 +00001638 __be16 df = 0;
1639 __u8 tos, ttl;
Pravin B Shelar0e6fbc52013-06-17 17:49:56 -07001640 int err;
stephen hemmingerd3428942012-10-01 12:32:35 +00001641
stephen hemminger823aa872013-04-27 11:31:57 +00001642 dst_port = rdst->remote_port ? rdst->remote_port : vxlan->dst_port;
David Stevens66817122013-03-15 04:35:51 +00001643 vni = rdst->remote_vni;
Cong Wange4c7ed42013-08-31 13:44:33 +08001644 dst = &rdst->remote_ip;
David Stevense4f67ad2012-11-20 02:50:14 +00001645
Cong Wange4c7ed42013-08-31 13:44:33 +08001646 if (vxlan_addr_any(dst)) {
David Stevense4f67ad2012-11-20 02:50:14 +00001647 if (did_rsc) {
David Stevense4f67ad2012-11-20 02:50:14 +00001648 /* short-circuited back to local bridge */
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001649 vxlan_encap_bypass(skb, vxlan, vxlan);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001650 return;
David Stevense4f67ad2012-11-20 02:50:14 +00001651 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001652 goto drop;
David Stevense4f67ad2012-11-20 02:50:14 +00001653 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001654
stephen hemmingerd3428942012-10-01 12:32:35 +00001655 old_iph = ip_hdr(skb);
1656
stephen hemmingerd3428942012-10-01 12:32:35 +00001657 ttl = vxlan->ttl;
Cong Wange4c7ed42013-08-31 13:44:33 +08001658 if (!ttl && vxlan_addr_multicast(dst))
stephen hemmingerd3428942012-10-01 12:32:35 +00001659 ttl = 1;
1660
1661 tos = vxlan->tos;
1662 if (tos == 1)
Pravin B Shelar206aaaf2013-03-25 14:49:53 +00001663 tos = ip_tunnel_get_dsfield(old_iph, skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001664
Pravin B Shelar49560532013-08-19 11:23:17 -07001665 src_port = vxlan_src_port(vxlan->port_min, vxlan->port_max, skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001666
Cong Wange4c7ed42013-08-31 13:44:33 +08001667 if (dst->sa.sa_family == AF_INET) {
1668 memset(&fl4, 0, sizeof(fl4));
1669 fl4.flowi4_oif = rdst->remote_ifindex;
1670 fl4.flowi4_tos = RT_TOS(tos);
1671 fl4.daddr = dst->sin.sin_addr.s_addr;
1672 fl4.saddr = vxlan->saddr.sin.sin_addr.s_addr;
stephen hemmingerca78f182012-10-09 20:35:48 +00001673
Cong Wange4c7ed42013-08-31 13:44:33 +08001674 rt = ip_route_output_key(dev_net(dev), &fl4);
1675 if (IS_ERR(rt)) {
1676 netdev_dbg(dev, "no route to %pI4\n",
1677 &dst->sin.sin_addr.s_addr);
1678 dev->stats.tx_carrier_errors++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001679 goto tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08001680 }
1681
1682 if (rt->dst.dev == dev) {
1683 netdev_dbg(dev, "circular route to %pI4\n",
1684 &dst->sin.sin_addr.s_addr);
1685 dev->stats.collisions++;
1686 goto tx_error;
1687 }
1688
1689 /* Bypass encapsulation if the destination is local */
1690 if (rt->rt_flags & RTCF_LOCAL &&
1691 !(rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
1692 struct vxlan_dev *dst_vxlan;
1693
1694 ip_rt_put(rt);
1695 dst_vxlan = vxlan_find_vni(dev_net(dev), vni, dst_port);
1696 if (!dst_vxlan)
1697 goto tx_error;
1698 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
1699 return;
1700 }
1701
1702 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
1703 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
1704
Nicolas Dichtel11796182013-09-02 15:34:55 +02001705 err = vxlan_xmit_skb(vxlan->vn_sock, rt, skb,
Cong Wange4c7ed42013-08-31 13:44:33 +08001706 fl4.saddr, dst->sin.sin_addr.s_addr,
1707 tos, ttl, df, src_port, dst_port,
1708 htonl(vni << 8));
1709
1710 if (err < 0)
1711 goto rt_tx_error;
1712 iptunnel_xmit_stats(err, &dev->stats, dev->tstats);
1713#if IS_ENABLED(CONFIG_IPV6)
1714 } else {
1715 struct sock *sk = vxlan->vn_sock->sock->sk;
1716 struct dst_entry *ndst;
1717 struct flowi6 fl6;
1718 u32 flags;
1719
1720 memset(&fl6, 0, sizeof(fl6));
1721 fl6.flowi6_oif = rdst->remote_ifindex;
1722 fl6.daddr = dst->sin6.sin6_addr;
1723 fl6.saddr = vxlan->saddr.sin6.sin6_addr;
Cong Wang8c1bb792013-09-02 10:06:51 +08001724 fl6.flowi6_proto = IPPROTO_UDP;
Cong Wange4c7ed42013-08-31 13:44:33 +08001725
1726 if (ipv6_stub->ipv6_dst_lookup(sk, &ndst, &fl6)) {
1727 netdev_dbg(dev, "no route to %pI6\n",
1728 &dst->sin6.sin6_addr);
1729 dev->stats.tx_carrier_errors++;
1730 goto tx_error;
1731 }
1732
1733 if (ndst->dev == dev) {
1734 netdev_dbg(dev, "circular route to %pI6\n",
1735 &dst->sin6.sin6_addr);
1736 dst_release(ndst);
1737 dev->stats.collisions++;
1738 goto tx_error;
1739 }
1740
1741 /* Bypass encapsulation if the destination is local */
1742 flags = ((struct rt6_info *)ndst)->rt6i_flags;
1743 if (flags & RTF_LOCAL &&
1744 !(flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
1745 struct vxlan_dev *dst_vxlan;
1746
1747 dst_release(ndst);
1748 dst_vxlan = vxlan_find_vni(dev_net(dev), vni, dst_port);
1749 if (!dst_vxlan)
1750 goto tx_error;
1751 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
1752 return;
1753 }
1754
1755 ttl = ttl ? : ip6_dst_hoplimit(ndst);
1756
Nicolas Dichtel11796182013-09-02 15:34:55 +02001757 err = vxlan6_xmit_skb(vxlan->vn_sock, ndst, skb,
Cong Wange4c7ed42013-08-31 13:44:33 +08001758 dev, &fl6.saddr, &fl6.daddr, 0, ttl,
1759 src_port, dst_port, htonl(vni << 8));
1760#endif
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001761 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001762
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001763 return;
stephen hemmingerd3428942012-10-01 12:32:35 +00001764
1765drop:
1766 dev->stats.tx_dropped++;
1767 goto tx_free;
1768
Pravin B Shelar49560532013-08-19 11:23:17 -07001769rt_tx_error:
1770 ip_rt_put(rt);
stephen hemmingerd3428942012-10-01 12:32:35 +00001771tx_error:
1772 dev->stats.tx_errors++;
1773tx_free:
1774 dev_kfree_skb(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00001775}
1776
David Stevens66817122013-03-15 04:35:51 +00001777/* Transmit local packets over Vxlan
1778 *
1779 * Outer IP header inherits ECN and DF from inner header.
1780 * Outer UDP destination is the VXLAN assigned port.
1781 * source port is based on hash of flow
1782 */
1783static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev)
1784{
1785 struct vxlan_dev *vxlan = netdev_priv(dev);
1786 struct ethhdr *eth;
1787 bool did_rsc = false;
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001788 struct vxlan_rdst *rdst;
David Stevens66817122013-03-15 04:35:51 +00001789 struct vxlan_fdb *f;
David Stevens66817122013-03-15 04:35:51 +00001790
1791 skb_reset_mac_header(skb);
1792 eth = eth_hdr(skb);
1793
Cong Wangf564f452013-08-31 13:44:36 +08001794 if ((vxlan->flags & VXLAN_F_PROXY)) {
1795 if (ntohs(eth->h_proto) == ETH_P_ARP)
1796 return arp_reduce(dev, skb);
1797#if IS_ENABLED(CONFIG_IPV6)
1798 else if (ntohs(eth->h_proto) == ETH_P_IPV6 &&
1799 skb->len >= sizeof(struct ipv6hdr) + sizeof(struct nd_msg) &&
1800 ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) {
1801 struct nd_msg *msg;
1802
1803 msg = (struct nd_msg *)skb_transport_header(skb);
1804 if (msg->icmph.icmp6_code == 0 &&
1805 msg->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION)
1806 return neigh_reduce(dev, skb);
1807 }
1808#endif
1809 }
David Stevens66817122013-03-15 04:35:51 +00001810
1811 f = vxlan_find_mac(vxlan, eth->h_dest);
David Stevensae884082013-04-19 00:36:26 +00001812 did_rsc = false;
1813
1814 if (f && (f->flags & NTF_ROUTER) && (vxlan->flags & VXLAN_F_RSC) &&
Cong Wange15a00a2013-08-31 13:44:34 +08001815 (ntohs(eth->h_proto) == ETH_P_IP ||
1816 ntohs(eth->h_proto) == ETH_P_IPV6)) {
David Stevensae884082013-04-19 00:36:26 +00001817 did_rsc = route_shortcircuit(dev, skb);
1818 if (did_rsc)
1819 f = vxlan_find_mac(vxlan, eth->h_dest);
1820 }
1821
David Stevens66817122013-03-15 04:35:51 +00001822 if (f == NULL) {
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001823 f = vxlan_find_mac(vxlan, all_zeros_mac);
1824 if (f == NULL) {
1825 if ((vxlan->flags & VXLAN_F_L2MISS) &&
1826 !is_multicast_ether_addr(eth->h_dest))
1827 vxlan_fdb_miss(vxlan, eth->h_dest);
David Stevens66817122013-03-15 04:35:51 +00001828
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001829 dev->stats.tx_dropped++;
1830 dev_kfree_skb(skb);
1831 return NETDEV_TX_OK;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -07001832 }
David Stevens66817122013-03-15 04:35:51 +00001833 }
1834
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001835 list_for_each_entry_rcu(rdst, &f->remotes, list) {
1836 struct sk_buff *skb1;
1837
1838 skb1 = skb_clone(skb, GFP_ATOMIC);
1839 if (skb1)
1840 vxlan_xmit_one(skb1, dev, rdst, did_rsc);
1841 }
1842
1843 dev_kfree_skb(skb);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001844 return NETDEV_TX_OK;
David Stevens66817122013-03-15 04:35:51 +00001845}
1846
stephen hemmingerd3428942012-10-01 12:32:35 +00001847/* Walk the forwarding table and purge stale entries */
1848static void vxlan_cleanup(unsigned long arg)
1849{
1850 struct vxlan_dev *vxlan = (struct vxlan_dev *) arg;
1851 unsigned long next_timer = jiffies + FDB_AGE_INTERVAL;
1852 unsigned int h;
1853
1854 if (!netif_running(vxlan->dev))
1855 return;
1856
1857 spin_lock_bh(&vxlan->hash_lock);
1858 for (h = 0; h < FDB_HASH_SIZE; ++h) {
1859 struct hlist_node *p, *n;
1860 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
1861 struct vxlan_fdb *f
1862 = container_of(p, struct vxlan_fdb, hlist);
1863 unsigned long timeout;
1864
stephen hemminger3c172862012-10-26 06:24:34 +00001865 if (f->state & NUD_PERMANENT)
stephen hemmingerd3428942012-10-01 12:32:35 +00001866 continue;
1867
1868 timeout = f->used + vxlan->age_interval * HZ;
1869 if (time_before_eq(timeout, jiffies)) {
1870 netdev_dbg(vxlan->dev,
1871 "garbage collect %pM\n",
1872 f->eth_addr);
1873 f->state = NUD_STALE;
1874 vxlan_fdb_destroy(vxlan, f);
1875 } else if (time_before(timeout, next_timer))
1876 next_timer = timeout;
1877 }
1878 }
1879 spin_unlock_bh(&vxlan->hash_lock);
1880
1881 mod_timer(&vxlan->age_timer, next_timer);
1882}
1883
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07001884static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan)
1885{
1886 __u32 vni = vxlan->default_dst.remote_vni;
1887
1888 vxlan->vn_sock = vs;
1889 hlist_add_head_rcu(&vxlan->hlist, vni_head(vs, vni));
1890}
1891
stephen hemmingerd3428942012-10-01 12:32:35 +00001892/* Setup stats when device is created */
1893static int vxlan_init(struct net_device *dev)
1894{
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001895 struct vxlan_dev *vxlan = netdev_priv(dev);
1896 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
1897 struct vxlan_sock *vs;
John Stultz827da442013-10-07 15:51:58 -07001898 int i;
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001899
Pravin B Shelare8171042013-03-25 14:49:46 +00001900 dev->tstats = alloc_percpu(struct pcpu_tstats);
1901 if (!dev->tstats)
stephen hemmingerd3428942012-10-01 12:32:35 +00001902 return -ENOMEM;
1903
John Stultz827da442013-10-07 15:51:58 -07001904 for_each_possible_cpu(i) {
1905 struct pcpu_tstats *vxlan_stats;
1906 vxlan_stats = per_cpu_ptr(dev->tstats, i);
1907 u64_stats_init(&vxlan_stats->syncp);
1908 }
1909
1910
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001911 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07001912 vs = vxlan_find_sock(dev_net(dev), vxlan->dst_port);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001913 if (vs) {
1914 /* If we have a socket with same port already, reuse it */
1915 atomic_inc(&vs->refcnt);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07001916 vxlan_vs_add_dev(vs, vxlan);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001917 } else {
1918 /* otherwise make new socket outside of RTNL */
1919 dev_hold(dev);
1920 queue_work(vxlan_wq, &vxlan->sock_work);
1921 }
1922 spin_unlock(&vn->sock_lock);
1923
stephen hemmingerd3428942012-10-01 12:32:35 +00001924 return 0;
1925}
1926
Stephen Hemmingerba609e92013-06-25 17:06:01 -07001927static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan)
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001928{
1929 struct vxlan_fdb *f;
1930
1931 spin_lock_bh(&vxlan->hash_lock);
1932 f = __vxlan_find_mac(vxlan, all_zeros_mac);
1933 if (f)
1934 vxlan_fdb_destroy(vxlan, f);
1935 spin_unlock_bh(&vxlan->hash_lock);
1936}
1937
Stephen Hemmingerebf40632013-06-17 14:16:11 -07001938static void vxlan_uninit(struct net_device *dev)
1939{
1940 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerebf40632013-06-17 14:16:11 -07001941 struct vxlan_sock *vs = vxlan->vn_sock;
1942
Stephen Hemmingerba609e92013-06-25 17:06:01 -07001943 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001944
Stephen Hemmingerebf40632013-06-17 14:16:11 -07001945 if (vs)
Pravin B Shelar012a5722013-08-19 11:23:07 -07001946 vxlan_sock_release(vs);
Stephen Hemmingerebf40632013-06-17 14:16:11 -07001947 free_percpu(dev->tstats);
1948}
1949
stephen hemmingerd3428942012-10-01 12:32:35 +00001950/* Start ageing timer and join group when device is brought up */
1951static int vxlan_open(struct net_device *dev)
1952{
1953 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001954 struct vxlan_sock *vs = vxlan->vn_sock;
1955
1956 /* socket hasn't been created */
1957 if (!vs)
1958 return -ENOTCONN;
stephen hemmingerd3428942012-10-01 12:32:35 +00001959
Gao feng79d4a942013-12-10 16:37:32 +08001960 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) {
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001961 vxlan_sock_hold(vs);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001962 dev_hold(dev);
stephen hemminger3fc2de22013-07-18 08:40:15 -07001963 queue_work(vxlan_wq, &vxlan->igmp_join);
stephen hemmingerd3428942012-10-01 12:32:35 +00001964 }
1965
1966 if (vxlan->age_interval)
1967 mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL);
1968
1969 return 0;
1970}
1971
1972/* Purge the forwarding table */
1973static void vxlan_flush(struct vxlan_dev *vxlan)
1974{
Cong Wang31fec5a2013-05-27 22:35:52 +00001975 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00001976
1977 spin_lock_bh(&vxlan->hash_lock);
1978 for (h = 0; h < FDB_HASH_SIZE; ++h) {
1979 struct hlist_node *p, *n;
1980 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
1981 struct vxlan_fdb *f
1982 = container_of(p, struct vxlan_fdb, hlist);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03001983 /* the all_zeros_mac entry is deleted at vxlan_uninit */
1984 if (!is_zero_ether_addr(f->eth_addr))
1985 vxlan_fdb_destroy(vxlan, f);
stephen hemmingerd3428942012-10-01 12:32:35 +00001986 }
1987 }
1988 spin_unlock_bh(&vxlan->hash_lock);
1989}
1990
1991/* Cleanup timer and forwarding table on shutdown */
1992static int vxlan_stop(struct net_device *dev)
1993{
stephen hemminger3fc2de22013-07-18 08:40:15 -07001994 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00001995 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001996 struct vxlan_sock *vs = vxlan->vn_sock;
stephen hemmingerd3428942012-10-01 12:32:35 +00001997
Cong Wange4c7ed42013-08-31 13:44:33 +08001998 if (vs && vxlan_addr_multicast(&vxlan->default_dst.remote_ip) &&
Gao feng95ab0992013-12-10 16:37:33 +08001999 !vxlan_group_used(vn, vxlan)) {
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002000 vxlan_sock_hold(vs);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07002001 dev_hold(dev);
stephen hemminger3fc2de22013-07-18 08:40:15 -07002002 queue_work(vxlan_wq, &vxlan->igmp_leave);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07002003 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002004
2005 del_timer_sync(&vxlan->age_timer);
2006
2007 vxlan_flush(vxlan);
2008
2009 return 0;
2010}
2011
stephen hemmingerd3428942012-10-01 12:32:35 +00002012/* Stub, nothing needs to be done. */
2013static void vxlan_set_multicast_list(struct net_device *dev)
2014{
2015}
2016
2017static const struct net_device_ops vxlan_netdev_ops = {
2018 .ndo_init = vxlan_init,
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002019 .ndo_uninit = vxlan_uninit,
stephen hemmingerd3428942012-10-01 12:32:35 +00002020 .ndo_open = vxlan_open,
2021 .ndo_stop = vxlan_stop,
2022 .ndo_start_xmit = vxlan_xmit,
Pravin B Shelare8171042013-03-25 14:49:46 +00002023 .ndo_get_stats64 = ip_tunnel_get_stats64,
stephen hemmingerd3428942012-10-01 12:32:35 +00002024 .ndo_set_rx_mode = vxlan_set_multicast_list,
2025 .ndo_change_mtu = eth_change_mtu,
2026 .ndo_validate_addr = eth_validate_addr,
2027 .ndo_set_mac_address = eth_mac_addr,
2028 .ndo_fdb_add = vxlan_fdb_add,
2029 .ndo_fdb_del = vxlan_fdb_delete,
2030 .ndo_fdb_dump = vxlan_fdb_dump,
2031};
2032
2033/* Info for udev, that this is a virtual tunnel endpoint */
2034static struct device_type vxlan_type = {
2035 .name = "vxlan",
2036};
2037
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002038/* Calls the ndo_add_vxlan_port of the caller in order to
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002039 * supply the listening VXLAN udp ports. Callers are expected
2040 * to implement the ndo_add_vxlan_port.
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002041 */
2042void vxlan_get_rx_port(struct net_device *dev)
2043{
2044 struct vxlan_sock *vs;
2045 struct net *net = dev_net(dev);
2046 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2047 sa_family_t sa_family;
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002048 __be16 port;
2049 unsigned int i;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002050
2051 spin_lock(&vn->sock_lock);
2052 for (i = 0; i < PORT_HASH_SIZE; ++i) {
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002053 hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) {
2054 port = inet_sk(vs->sock->sk)->inet_sport;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002055 sa_family = vs->sock->sk->sk_family;
2056 dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
2057 port);
2058 }
2059 }
2060 spin_unlock(&vn->sock_lock);
2061}
2062EXPORT_SYMBOL_GPL(vxlan_get_rx_port);
2063
stephen hemmingerd3428942012-10-01 12:32:35 +00002064/* Initialize the device structure. */
2065static void vxlan_setup(struct net_device *dev)
2066{
2067 struct vxlan_dev *vxlan = netdev_priv(dev);
Cong Wang31fec5a2013-05-27 22:35:52 +00002068 unsigned int h;
stephen hemminger05f47d62012-10-09 20:35:50 +00002069 int low, high;
stephen hemmingerd3428942012-10-01 12:32:35 +00002070
2071 eth_hw_addr_random(dev);
2072 ether_setup(dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08002073 if (vxlan->default_dst.remote_ip.sa.sa_family == AF_INET6)
2074 dev->hard_header_len = ETH_HLEN + VXLAN6_HEADROOM;
2075 else
2076 dev->hard_header_len = ETH_HLEN + VXLAN_HEADROOM;
stephen hemmingerd3428942012-10-01 12:32:35 +00002077
2078 dev->netdev_ops = &vxlan_netdev_ops;
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002079 dev->destructor = free_netdev;
stephen hemmingerd3428942012-10-01 12:32:35 +00002080 SET_NETDEV_DEVTYPE(dev, &vxlan_type);
2081
2082 dev->tx_queue_len = 0;
2083 dev->features |= NETIF_F_LLTX;
2084 dev->features |= NETIF_F_NETNS_LOCAL;
Joseph Gasparakisd6727fe2012-12-07 14:14:16 +00002085 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002086 dev->features |= NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002087 dev->features |= NETIF_F_GSO_SOFTWARE;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002088
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002089 dev->vlan_features = dev->features;
2090 dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002091 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002092 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002093 dev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
stephen hemmingerd3428942012-10-01 12:32:35 +00002094 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
stephen hemminger6602d002012-12-31 12:00:21 +00002095 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
stephen hemmingerd3428942012-10-01 12:32:35 +00002096
stephen hemminger553675f2013-05-16 11:35:20 +00002097 INIT_LIST_HEAD(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00002098 spin_lock_init(&vxlan->hash_lock);
stephen hemminger3fc2de22013-07-18 08:40:15 -07002099 INIT_WORK(&vxlan->igmp_join, vxlan_igmp_join);
2100 INIT_WORK(&vxlan->igmp_leave, vxlan_igmp_leave);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002101 INIT_WORK(&vxlan->sock_work, vxlan_sock_work);
stephen hemmingerd3428942012-10-01 12:32:35 +00002102
2103 init_timer_deferrable(&vxlan->age_timer);
2104 vxlan->age_timer.function = vxlan_cleanup;
2105 vxlan->age_timer.data = (unsigned long) vxlan;
2106
Eric W. Biederman0bbf87d2013-09-28 14:10:59 -07002107 inet_get_local_port_range(dev_net(dev), &low, &high);
stephen hemminger05f47d62012-10-09 20:35:50 +00002108 vxlan->port_min = low;
2109 vxlan->port_max = high;
stephen hemminger823aa872013-04-27 11:31:57 +00002110 vxlan->dst_port = htons(vxlan_port);
stephen hemminger05f47d62012-10-09 20:35:50 +00002111
stephen hemmingerd3428942012-10-01 12:32:35 +00002112 vxlan->dev = dev;
2113
2114 for (h = 0; h < FDB_HASH_SIZE; ++h)
2115 INIT_HLIST_HEAD(&vxlan->fdb_head[h]);
2116}
2117
2118static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = {
2119 [IFLA_VXLAN_ID] = { .type = NLA_U32 },
stephen hemminger5d174dd2013-04-27 11:31:55 +00002120 [IFLA_VXLAN_GROUP] = { .len = FIELD_SIZEOF(struct iphdr, daddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002121 [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002122 [IFLA_VXLAN_LINK] = { .type = NLA_U32 },
2123 [IFLA_VXLAN_LOCAL] = { .len = FIELD_SIZEOF(struct iphdr, saddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002124 [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002125 [IFLA_VXLAN_TOS] = { .type = NLA_U8 },
2126 [IFLA_VXLAN_TTL] = { .type = NLA_U8 },
2127 [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 },
2128 [IFLA_VXLAN_AGEING] = { .type = NLA_U32 },
2129 [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 },
stephen hemminger05f47d62012-10-09 20:35:50 +00002130 [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) },
David Stevense4f67ad2012-11-20 02:50:14 +00002131 [IFLA_VXLAN_PROXY] = { .type = NLA_U8 },
2132 [IFLA_VXLAN_RSC] = { .type = NLA_U8 },
2133 [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 },
2134 [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 },
stephen hemminger823aa872013-04-27 11:31:57 +00002135 [IFLA_VXLAN_PORT] = { .type = NLA_U16 },
stephen hemmingerd3428942012-10-01 12:32:35 +00002136};
2137
2138static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[])
2139{
2140 if (tb[IFLA_ADDRESS]) {
2141 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) {
2142 pr_debug("invalid link address (not ethernet)\n");
2143 return -EINVAL;
2144 }
2145
2146 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) {
2147 pr_debug("invalid all zero ethernet address\n");
2148 return -EADDRNOTAVAIL;
2149 }
2150 }
2151
2152 if (!data)
2153 return -EINVAL;
2154
2155 if (data[IFLA_VXLAN_ID]) {
2156 __u32 id = nla_get_u32(data[IFLA_VXLAN_ID]);
2157 if (id >= VXLAN_VID_MASK)
2158 return -ERANGE;
2159 }
2160
stephen hemminger05f47d62012-10-09 20:35:50 +00002161 if (data[IFLA_VXLAN_PORT_RANGE]) {
2162 const struct ifla_vxlan_port_range *p
2163 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2164
2165 if (ntohs(p->high) < ntohs(p->low)) {
2166 pr_debug("port range %u .. %u not valid\n",
2167 ntohs(p->low), ntohs(p->high));
2168 return -EINVAL;
2169 }
2170 }
2171
stephen hemmingerd3428942012-10-01 12:32:35 +00002172 return 0;
2173}
2174
Yan Burman1b13c972013-01-29 23:43:07 +00002175static void vxlan_get_drvinfo(struct net_device *netdev,
2176 struct ethtool_drvinfo *drvinfo)
2177{
2178 strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version));
2179 strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver));
2180}
2181
2182static const struct ethtool_ops vxlan_ethtool_ops = {
2183 .get_drvinfo = vxlan_get_drvinfo,
2184 .get_link = ethtool_op_get_link,
2185};
2186
stephen hemminger553675f2013-05-16 11:35:20 +00002187static void vxlan_del_work(struct work_struct *work)
2188{
2189 struct vxlan_sock *vs = container_of(work, struct vxlan_sock, del_work);
2190
2191 sk_release_kernel(vs->sock->sk);
2192 kfree_rcu(vs, rcu);
2193}
2194
Cong Wange4c7ed42013-08-31 13:44:33 +08002195#if IS_ENABLED(CONFIG_IPV6)
2196/* Create UDP socket for encapsulation receive. AF_INET6 socket
2197 * could be used for both IPv4 and IPv6 communications, but
2198 * users may set bindv6only=1.
2199 */
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002200static struct socket *create_v6_sock(struct net *net, __be16 port)
stephen hemminger553675f2013-05-16 11:35:20 +00002201{
stephen hemminger553675f2013-05-16 11:35:20 +00002202 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08002203 struct socket *sock;
2204 struct sockaddr_in6 vxlan_addr = {
2205 .sin6_family = AF_INET6,
2206 .sin6_port = port,
2207 };
2208 int rc, val = 1;
2209
2210 rc = sock_create_kern(AF_INET6, SOCK_DGRAM, IPPROTO_UDP, &sock);
2211 if (rc < 0) {
2212 pr_debug("UDPv6 socket create failed\n");
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002213 return ERR_PTR(rc);
Cong Wange4c7ed42013-08-31 13:44:33 +08002214 }
2215
2216 /* Put in proper namespace */
2217 sk = sock->sk;
2218 sk_change_net(sk, net);
2219
2220 kernel_setsockopt(sock, SOL_IPV6, IPV6_V6ONLY,
2221 (char *)&val, sizeof(val));
2222 rc = kernel_bind(sock, (struct sockaddr *)&vxlan_addr,
2223 sizeof(struct sockaddr_in6));
2224 if (rc < 0) {
2225 pr_debug("bind for UDPv6 socket %pI6:%u (%d)\n",
2226 &vxlan_addr.sin6_addr, ntohs(vxlan_addr.sin6_port), rc);
2227 sk_release_kernel(sk);
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002228 return ERR_PTR(rc);
Cong Wange4c7ed42013-08-31 13:44:33 +08002229 }
2230 /* At this point, IPv6 module should have been loaded in
2231 * sock_create_kern().
2232 */
2233 BUG_ON(!ipv6_stub);
2234
Cong Wange4c7ed42013-08-31 13:44:33 +08002235 /* Disable multicast loopback */
2236 inet_sk(sk)->mc_loop = 0;
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002237 return sock;
Cong Wange4c7ed42013-08-31 13:44:33 +08002238}
2239
2240#else
2241
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002242static struct socket *create_v6_sock(struct net *net, __be16 port)
Cong Wange4c7ed42013-08-31 13:44:33 +08002243{
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002244 return ERR_PTR(-EPFNOSUPPORT);
Cong Wange4c7ed42013-08-31 13:44:33 +08002245}
2246#endif
2247
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002248static struct socket *create_v4_sock(struct net *net, __be16 port)
Cong Wange4c7ed42013-08-31 13:44:33 +08002249{
2250 struct sock *sk;
2251 struct socket *sock;
stephen hemminger553675f2013-05-16 11:35:20 +00002252 struct sockaddr_in vxlan_addr = {
2253 .sin_family = AF_INET,
2254 .sin_addr.s_addr = htonl(INADDR_ANY),
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -07002255 .sin_port = port,
stephen hemminger553675f2013-05-16 11:35:20 +00002256 };
2257 int rc;
Cong Wange4c7ed42013-08-31 13:44:33 +08002258
2259 /* Create UDP socket for encapsulation receive. */
2260 rc = sock_create_kern(AF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock);
2261 if (rc < 0) {
2262 pr_debug("UDP socket create failed\n");
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002263 return ERR_PTR(rc);
Cong Wange4c7ed42013-08-31 13:44:33 +08002264 }
2265
2266 /* Put in proper namespace */
2267 sk = sock->sk;
2268 sk_change_net(sk, net);
2269
2270 rc = kernel_bind(sock, (struct sockaddr *) &vxlan_addr,
2271 sizeof(vxlan_addr));
2272 if (rc < 0) {
2273 pr_debug("bind for UDP socket %pI4:%u (%d)\n",
2274 &vxlan_addr.sin_addr, ntohs(vxlan_addr.sin_port), rc);
2275 sk_release_kernel(sk);
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002276 return ERR_PTR(rc);
Cong Wange4c7ed42013-08-31 13:44:33 +08002277 }
2278
Cong Wange4c7ed42013-08-31 13:44:33 +08002279 /* Disable multicast loopback */
2280 inet_sk(sk)->mc_loop = 0;
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002281 return sock;
Cong Wange4c7ed42013-08-31 13:44:33 +08002282}
2283
2284/* Create new listen socket if needed */
2285static struct vxlan_sock *vxlan_socket_create(struct net *net, __be16 port,
2286 vxlan_rcv_t *rcv, void *data, bool ipv6)
2287{
2288 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2289 struct vxlan_sock *vs;
2290 struct socket *sock;
2291 struct sock *sk;
Cong Wang31fec5a2013-05-27 22:35:52 +00002292 unsigned int h;
stephen hemminger553675f2013-05-16 11:35:20 +00002293
2294 vs = kmalloc(sizeof(*vs), GFP_KERNEL);
Cong Wange4c7ed42013-08-31 13:44:33 +08002295 if (!vs)
stephen hemminger553675f2013-05-16 11:35:20 +00002296 return ERR_PTR(-ENOMEM);
2297
2298 for (h = 0; h < VNI_HASH_SIZE; ++h)
2299 INIT_HLIST_HEAD(&vs->vni_list[h]);
2300
2301 INIT_WORK(&vs->del_work, vxlan_del_work);
2302
Cong Wange4c7ed42013-08-31 13:44:33 +08002303 if (ipv6)
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002304 sock = create_v6_sock(net, port);
Cong Wange4c7ed42013-08-31 13:44:33 +08002305 else
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002306 sock = create_v4_sock(net, port);
2307 if (IS_ERR(sock)) {
stephen hemminger553675f2013-05-16 11:35:20 +00002308 kfree(vs);
Duan Jionge50fddc2013-11-01 13:09:43 +08002309 return ERR_CAST(sock);
stephen hemminger553675f2013-05-16 11:35:20 +00002310 }
2311
Cong Wange4c7ed42013-08-31 13:44:33 +08002312 vs->sock = sock;
2313 sk = sock->sk;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002314 atomic_set(&vs->refcnt, 1);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07002315 vs->rcv = rcv;
Pravin B Shelar012a5722013-08-19 11:23:07 -07002316 vs->data = data;
Pravin B Shelar559835ea2013-09-24 10:25:40 -07002317 rcu_assign_sk_user_data(vs->sock->sk, vs);
stephen hemminger553675f2013-05-16 11:35:20 +00002318
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002319 spin_lock(&vn->sock_lock);
2320 hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002321 vxlan_notify_add_rx_port(sk);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002322 spin_unlock(&vn->sock_lock);
stephen hemminger553675f2013-05-16 11:35:20 +00002323
2324 /* Mark socket as an encapsulation socket. */
2325 udp_sk(sk)->encap_type = 1;
2326 udp_sk(sk)->encap_rcv = vxlan_udp_encap_recv;
Cong Wange4c7ed42013-08-31 13:44:33 +08002327#if IS_ENABLED(CONFIG_IPV6)
2328 if (ipv6)
2329 ipv6_stub->udpv6_encap_enable();
2330 else
2331#endif
2332 udp_encap_enable();
2333
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002334 return vs;
2335}
stephen hemminger553675f2013-05-16 11:35:20 +00002336
Pravin B Shelar012a5722013-08-19 11:23:07 -07002337struct vxlan_sock *vxlan_sock_add(struct net *net, __be16 port,
2338 vxlan_rcv_t *rcv, void *data,
Cong Wange4c7ed42013-08-31 13:44:33 +08002339 bool no_share, bool ipv6)
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002340{
2341 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2342 struct vxlan_sock *vs;
2343
Cong Wange4c7ed42013-08-31 13:44:33 +08002344 vs = vxlan_socket_create(net, port, rcv, data, ipv6);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002345 if (!IS_ERR(vs))
2346 return vs;
2347
Pravin B Shelar012a5722013-08-19 11:23:07 -07002348 if (no_share) /* Return error if sharing is not allowed. */
2349 return vs;
2350
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002351 spin_lock(&vn->sock_lock);
2352 vs = vxlan_find_sock(net, port);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07002353 if (vs) {
2354 if (vs->rcv == rcv)
2355 atomic_inc(&vs->refcnt);
2356 else
2357 vs = ERR_PTR(-EBUSY);
2358 }
2359 spin_unlock(&vn->sock_lock);
2360
2361 if (!vs)
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002362 vs = ERR_PTR(-EINVAL);
2363
stephen hemminger553675f2013-05-16 11:35:20 +00002364 return vs;
2365}
Pravin B Shelar012a5722013-08-19 11:23:07 -07002366EXPORT_SYMBOL_GPL(vxlan_sock_add);
stephen hemminger553675f2013-05-16 11:35:20 +00002367
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002368/* Scheduled at device creation to bind to a socket */
2369static void vxlan_sock_work(struct work_struct *work)
2370{
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002371 struct vxlan_dev *vxlan = container_of(work, struct vxlan_dev, sock_work);
2372 struct net *net = dev_net(vxlan->dev);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002373 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002374 __be16 port = vxlan->dst_port;
2375 struct vxlan_sock *nvs;
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002376
Cong Wange4c7ed42013-08-31 13:44:33 +08002377 nvs = vxlan_sock_add(net, port, vxlan_rcv, NULL, false, vxlan->flags & VXLAN_F_IPV6);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002378 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002379 if (!IS_ERR(nvs))
2380 vxlan_vs_add_dev(nvs, vxlan);
2381 spin_unlock(&vn->sock_lock);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002382
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002383 dev_put(vxlan->dev);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002384}
2385
stephen hemmingerd3428942012-10-01 12:32:35 +00002386static int vxlan_newlink(struct net *net, struct net_device *dev,
2387 struct nlattr *tb[], struct nlattr *data[])
2388{
stephen hemminger553675f2013-05-16 11:35:20 +00002389 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00002390 struct vxlan_dev *vxlan = netdev_priv(dev);
Atzm Watanabec7995c42013-04-16 02:50:52 +00002391 struct vxlan_rdst *dst = &vxlan->default_dst;
stephen hemmingerd3428942012-10-01 12:32:35 +00002392 __u32 vni;
2393 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002394 bool use_ipv6 = false;
stephen hemmingerd3428942012-10-01 12:32:35 +00002395
2396 if (!data[IFLA_VXLAN_ID])
2397 return -EINVAL;
2398
2399 vni = nla_get_u32(data[IFLA_VXLAN_ID]);
Atzm Watanabec7995c42013-04-16 02:50:52 +00002400 dst->remote_vni = vni;
stephen hemmingerd3428942012-10-01 12:32:35 +00002401
Cong Wange4c7ed42013-08-31 13:44:33 +08002402 if (data[IFLA_VXLAN_GROUP]) {
2403 dst->remote_ip.sin.sin_addr.s_addr = nla_get_be32(data[IFLA_VXLAN_GROUP]);
2404 dst->remote_ip.sa.sa_family = AF_INET;
2405 } else if (data[IFLA_VXLAN_GROUP6]) {
2406 if (!IS_ENABLED(CONFIG_IPV6))
2407 return -EPFNOSUPPORT;
stephen hemmingerd3428942012-10-01 12:32:35 +00002408
Cong Wange4c7ed42013-08-31 13:44:33 +08002409 nla_memcpy(&dst->remote_ip.sin6.sin6_addr, data[IFLA_VXLAN_GROUP6],
2410 sizeof(struct in6_addr));
2411 dst->remote_ip.sa.sa_family = AF_INET6;
2412 use_ipv6 = true;
2413 }
2414
2415 if (data[IFLA_VXLAN_LOCAL]) {
2416 vxlan->saddr.sin.sin_addr.s_addr = nla_get_be32(data[IFLA_VXLAN_LOCAL]);
2417 vxlan->saddr.sa.sa_family = AF_INET;
2418 } else if (data[IFLA_VXLAN_LOCAL6]) {
2419 if (!IS_ENABLED(CONFIG_IPV6))
2420 return -EPFNOSUPPORT;
2421
2422 /* TODO: respect scope id */
2423 nla_memcpy(&vxlan->saddr.sin6.sin6_addr, data[IFLA_VXLAN_LOCAL6],
2424 sizeof(struct in6_addr));
2425 vxlan->saddr.sa.sa_family = AF_INET6;
2426 use_ipv6 = true;
2427 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002428
stephen hemminger34e02aa2012-10-09 20:35:53 +00002429 if (data[IFLA_VXLAN_LINK] &&
Atzm Watanabec7995c42013-04-16 02:50:52 +00002430 (dst->remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]))) {
stephen hemminger34e02aa2012-10-09 20:35:53 +00002431 struct net_device *lowerdev
Atzm Watanabec7995c42013-04-16 02:50:52 +00002432 = __dev_get_by_index(net, dst->remote_ifindex);
stephen hemmingerd3428942012-10-01 12:32:35 +00002433
stephen hemminger34e02aa2012-10-09 20:35:53 +00002434 if (!lowerdev) {
Atzm Watanabec7995c42013-04-16 02:50:52 +00002435 pr_info("ifindex %d does not exist\n", dst->remote_ifindex);
stephen hemminger34e02aa2012-10-09 20:35:53 +00002436 return -ENODEV;
stephen hemmingerd3428942012-10-01 12:32:35 +00002437 }
stephen hemminger34e02aa2012-10-09 20:35:53 +00002438
Cong Wange4c7ed42013-08-31 13:44:33 +08002439#if IS_ENABLED(CONFIG_IPV6)
2440 if (use_ipv6) {
2441 struct inet6_dev *idev = __in6_dev_get(lowerdev);
2442 if (idev && idev->cnf.disable_ipv6) {
2443 pr_info("IPv6 is disabled via sysctl\n");
2444 return -EPERM;
2445 }
2446 vxlan->flags |= VXLAN_F_IPV6;
2447 }
2448#endif
2449
stephen hemminger34e02aa2012-10-09 20:35:53 +00002450 if (!tb[IFLA_MTU])
Cong Wange4c7ed42013-08-31 13:44:33 +08002451 dev->mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
Alexander Duyck1ba56fb2012-11-13 13:10:59 +00002452
2453 /* update header length based on lower device */
2454 dev->hard_header_len = lowerdev->hard_header_len +
Cong Wange4c7ed42013-08-31 13:44:33 +08002455 (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
stephen hemmingerd3428942012-10-01 12:32:35 +00002456 }
2457
2458 if (data[IFLA_VXLAN_TOS])
2459 vxlan->tos = nla_get_u8(data[IFLA_VXLAN_TOS]);
2460
Vincent Bernatafb97182012-10-30 10:27:16 +00002461 if (data[IFLA_VXLAN_TTL])
2462 vxlan->ttl = nla_get_u8(data[IFLA_VXLAN_TTL]);
2463
stephen hemmingerd3428942012-10-01 12:32:35 +00002464 if (!data[IFLA_VXLAN_LEARNING] || nla_get_u8(data[IFLA_VXLAN_LEARNING]))
David Stevense4f67ad2012-11-20 02:50:14 +00002465 vxlan->flags |= VXLAN_F_LEARN;
stephen hemmingerd3428942012-10-01 12:32:35 +00002466
2467 if (data[IFLA_VXLAN_AGEING])
2468 vxlan->age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]);
2469 else
2470 vxlan->age_interval = FDB_AGE_DEFAULT;
2471
David Stevense4f67ad2012-11-20 02:50:14 +00002472 if (data[IFLA_VXLAN_PROXY] && nla_get_u8(data[IFLA_VXLAN_PROXY]))
2473 vxlan->flags |= VXLAN_F_PROXY;
2474
2475 if (data[IFLA_VXLAN_RSC] && nla_get_u8(data[IFLA_VXLAN_RSC]))
2476 vxlan->flags |= VXLAN_F_RSC;
2477
2478 if (data[IFLA_VXLAN_L2MISS] && nla_get_u8(data[IFLA_VXLAN_L2MISS]))
2479 vxlan->flags |= VXLAN_F_L2MISS;
2480
2481 if (data[IFLA_VXLAN_L3MISS] && nla_get_u8(data[IFLA_VXLAN_L3MISS]))
2482 vxlan->flags |= VXLAN_F_L3MISS;
2483
stephen hemmingerd3428942012-10-01 12:32:35 +00002484 if (data[IFLA_VXLAN_LIMIT])
2485 vxlan->addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]);
2486
stephen hemminger05f47d62012-10-09 20:35:50 +00002487 if (data[IFLA_VXLAN_PORT_RANGE]) {
2488 const struct ifla_vxlan_port_range *p
2489 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2490 vxlan->port_min = ntohs(p->low);
2491 vxlan->port_max = ntohs(p->high);
2492 }
2493
stephen hemminger823aa872013-04-27 11:31:57 +00002494 if (data[IFLA_VXLAN_PORT])
2495 vxlan->dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]);
2496
stephen hemminger553675f2013-05-16 11:35:20 +00002497 if (vxlan_find_vni(net, vni, vxlan->dst_port)) {
2498 pr_info("duplicate VNI %u\n", vni);
2499 return -EEXIST;
2500 }
2501
Yan Burman1b13c972013-01-29 23:43:07 +00002502 SET_ETHTOOL_OPS(dev, &vxlan_ethtool_ops);
2503
Sridhar Samudrala2936b6a2013-09-17 12:12:40 -07002504 /* create an fdb entry for a valid default destination */
2505 if (!vxlan_addr_any(&vxlan->default_dst.remote_ip)) {
2506 err = vxlan_fdb_create(vxlan, all_zeros_mac,
2507 &vxlan->default_dst.remote_ip,
2508 NUD_REACHABLE|NUD_PERMANENT,
2509 NLM_F_EXCL|NLM_F_CREATE,
2510 vxlan->dst_port,
2511 vxlan->default_dst.remote_vni,
2512 vxlan->default_dst.remote_ifindex,
2513 NTF_SELF);
2514 if (err)
2515 return err;
2516 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002517
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002518 err = register_netdevice(dev);
2519 if (err) {
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002520 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002521 return err;
2522 }
2523
stephen hemminger553675f2013-05-16 11:35:20 +00002524 list_add(&vxlan->next, &vn->vxlan_list);
stephen hemminger553675f2013-05-16 11:35:20 +00002525
2526 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002527}
2528
2529static void vxlan_dellink(struct net_device *dev, struct list_head *head)
2530{
stephen hemmingerfe5c3562013-07-13 10:18:18 -07002531 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00002532 struct vxlan_dev *vxlan = netdev_priv(dev);
2533
stephen hemmingerfe5c3562013-07-13 10:18:18 -07002534 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002535 if (!hlist_unhashed(&vxlan->hlist))
2536 hlist_del_rcu(&vxlan->hlist);
stephen hemmingerfe5c3562013-07-13 10:18:18 -07002537 spin_unlock(&vn->sock_lock);
2538
stephen hemminger553675f2013-05-16 11:35:20 +00002539 list_del(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00002540 unregister_netdevice_queue(dev, head);
2541}
2542
2543static size_t vxlan_get_size(const struct net_device *dev)
2544{
2545
2546 return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */
Cong Wange4c7ed42013-08-31 13:44:33 +08002547 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00002548 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */
Cong Wange4c7ed42013-08-31 13:44:33 +08002549 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00002550 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */
2551 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */
2552 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */
David Stevense4f67ad2012-11-20 02:50:14 +00002553 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */
2554 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */
2555 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */
2556 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */
stephen hemmingerd3428942012-10-01 12:32:35 +00002557 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */
2558 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */
stephen hemminger05f47d62012-10-09 20:35:50 +00002559 nla_total_size(sizeof(struct ifla_vxlan_port_range)) +
stephen hemminger823aa872013-04-27 11:31:57 +00002560 nla_total_size(sizeof(__be16))+ /* IFLA_VXLAN_PORT */
stephen hemmingerd3428942012-10-01 12:32:35 +00002561 0;
2562}
2563
2564static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev)
2565{
2566 const struct vxlan_dev *vxlan = netdev_priv(dev);
Atzm Watanabec7995c42013-04-16 02:50:52 +00002567 const struct vxlan_rdst *dst = &vxlan->default_dst;
stephen hemminger05f47d62012-10-09 20:35:50 +00002568 struct ifla_vxlan_port_range ports = {
2569 .low = htons(vxlan->port_min),
2570 .high = htons(vxlan->port_max),
2571 };
stephen hemmingerd3428942012-10-01 12:32:35 +00002572
Atzm Watanabec7995c42013-04-16 02:50:52 +00002573 if (nla_put_u32(skb, IFLA_VXLAN_ID, dst->remote_vni))
stephen hemmingerd3428942012-10-01 12:32:35 +00002574 goto nla_put_failure;
2575
Cong Wange4c7ed42013-08-31 13:44:33 +08002576 if (!vxlan_addr_any(&dst->remote_ip)) {
2577 if (dst->remote_ip.sa.sa_family == AF_INET) {
2578 if (nla_put_be32(skb, IFLA_VXLAN_GROUP,
2579 dst->remote_ip.sin.sin_addr.s_addr))
2580 goto nla_put_failure;
2581#if IS_ENABLED(CONFIG_IPV6)
2582 } else {
2583 if (nla_put(skb, IFLA_VXLAN_GROUP6, sizeof(struct in6_addr),
2584 &dst->remote_ip.sin6.sin6_addr))
2585 goto nla_put_failure;
2586#endif
2587 }
2588 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002589
Atzm Watanabec7995c42013-04-16 02:50:52 +00002590 if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +00002591 goto nla_put_failure;
2592
Cong Wange4c7ed42013-08-31 13:44:33 +08002593 if (!vxlan_addr_any(&vxlan->saddr)) {
2594 if (vxlan->saddr.sa.sa_family == AF_INET) {
2595 if (nla_put_be32(skb, IFLA_VXLAN_LOCAL,
2596 vxlan->saddr.sin.sin_addr.s_addr))
2597 goto nla_put_failure;
2598#if IS_ENABLED(CONFIG_IPV6)
2599 } else {
2600 if (nla_put(skb, IFLA_VXLAN_LOCAL6, sizeof(struct in6_addr),
2601 &vxlan->saddr.sin6.sin6_addr))
2602 goto nla_put_failure;
2603#endif
2604 }
2605 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002606
2607 if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->ttl) ||
2608 nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->tos) ||
David Stevense4f67ad2012-11-20 02:50:14 +00002609 nla_put_u8(skb, IFLA_VXLAN_LEARNING,
2610 !!(vxlan->flags & VXLAN_F_LEARN)) ||
2611 nla_put_u8(skb, IFLA_VXLAN_PROXY,
2612 !!(vxlan->flags & VXLAN_F_PROXY)) ||
2613 nla_put_u8(skb, IFLA_VXLAN_RSC, !!(vxlan->flags & VXLAN_F_RSC)) ||
2614 nla_put_u8(skb, IFLA_VXLAN_L2MISS,
2615 !!(vxlan->flags & VXLAN_F_L2MISS)) ||
2616 nla_put_u8(skb, IFLA_VXLAN_L3MISS,
2617 !!(vxlan->flags & VXLAN_F_L3MISS)) ||
stephen hemmingerd3428942012-10-01 12:32:35 +00002618 nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->age_interval) ||
stephen hemminger823aa872013-04-27 11:31:57 +00002619 nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->addrmax) ||
2620 nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->dst_port))
stephen hemmingerd3428942012-10-01 12:32:35 +00002621 goto nla_put_failure;
2622
stephen hemminger05f47d62012-10-09 20:35:50 +00002623 if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports))
2624 goto nla_put_failure;
2625
stephen hemmingerd3428942012-10-01 12:32:35 +00002626 return 0;
2627
2628nla_put_failure:
2629 return -EMSGSIZE;
2630}
2631
2632static struct rtnl_link_ops vxlan_link_ops __read_mostly = {
2633 .kind = "vxlan",
2634 .maxtype = IFLA_VXLAN_MAX,
2635 .policy = vxlan_policy,
2636 .priv_size = sizeof(struct vxlan_dev),
2637 .setup = vxlan_setup,
2638 .validate = vxlan_validate,
2639 .newlink = vxlan_newlink,
2640 .dellink = vxlan_dellink,
2641 .get_size = vxlan_get_size,
2642 .fill_info = vxlan_fill_info,
2643};
2644
2645static __net_init int vxlan_init_net(struct net *net)
2646{
2647 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Cong Wang31fec5a2013-05-27 22:35:52 +00002648 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002649
stephen hemminger553675f2013-05-16 11:35:20 +00002650 INIT_LIST_HEAD(&vn->vxlan_list);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002651 spin_lock_init(&vn->sock_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002652
stephen hemminger553675f2013-05-16 11:35:20 +00002653 for (h = 0; h < PORT_HASH_SIZE; ++h)
2654 INIT_HLIST_HEAD(&vn->sock_list[h]);
stephen hemmingerd3428942012-10-01 12:32:35 +00002655
2656 return 0;
2657}
2658
2659static __net_exit void vxlan_exit_net(struct net *net)
2660{
2661 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Zang MingJie9cb6cb72013-03-06 04:37:37 +00002662 struct vxlan_dev *vxlan;
stephen hemminger372675a2013-07-18 08:38:26 -07002663 LIST_HEAD(list);
Zang MingJie9cb6cb72013-03-06 04:37:37 +00002664
2665 rtnl_lock();
stephen hemminger553675f2013-05-16 11:35:20 +00002666 list_for_each_entry(vxlan, &vn->vxlan_list, next)
stephen hemminger372675a2013-07-18 08:38:26 -07002667 unregister_netdevice_queue(vxlan->dev, &list);
2668 unregister_netdevice_many(&list);
Zang MingJie9cb6cb72013-03-06 04:37:37 +00002669 rtnl_unlock();
stephen hemmingerd3428942012-10-01 12:32:35 +00002670}
2671
2672static struct pernet_operations vxlan_net_ops = {
2673 .init = vxlan_init_net,
2674 .exit = vxlan_exit_net,
2675 .id = &vxlan_net_id,
2676 .size = sizeof(struct vxlan_net),
2677};
2678
2679static int __init vxlan_init_module(void)
2680{
2681 int rc;
2682
Stephen Hemminger758c57d2013-06-17 14:16:09 -07002683 vxlan_wq = alloc_workqueue("vxlan", 0, 0);
2684 if (!vxlan_wq)
2685 return -ENOMEM;
2686
stephen hemmingerd3428942012-10-01 12:32:35 +00002687 get_random_bytes(&vxlan_salt, sizeof(vxlan_salt));
2688
2689 rc = register_pernet_device(&vxlan_net_ops);
2690 if (rc)
2691 goto out1;
2692
2693 rc = rtnl_link_register(&vxlan_link_ops);
2694 if (rc)
2695 goto out2;
2696
2697 return 0;
2698
2699out2:
2700 unregister_pernet_device(&vxlan_net_ops);
2701out1:
Stephen Hemminger758c57d2013-06-17 14:16:09 -07002702 destroy_workqueue(vxlan_wq);
stephen hemmingerd3428942012-10-01 12:32:35 +00002703 return rc;
2704}
Cong Wang7332a132013-05-27 22:35:53 +00002705late_initcall(vxlan_init_module);
stephen hemmingerd3428942012-10-01 12:32:35 +00002706
2707static void __exit vxlan_cleanup_module(void)
2708{
Stephen Hemmingerb7153982013-06-17 14:16:09 -07002709 rtnl_link_unregister(&vxlan_link_ops);
Stephen Hemminger758c57d2013-06-17 14:16:09 -07002710 destroy_workqueue(vxlan_wq);
Pravin B Shelarf89e57c2013-07-11 11:38:06 -07002711 unregister_pernet_device(&vxlan_net_ops);
David Stevens66817122013-03-15 04:35:51 +00002712 rcu_barrier();
stephen hemmingerd3428942012-10-01 12:32:35 +00002713}
2714module_exit(vxlan_cleanup_module);
2715
2716MODULE_LICENSE("GPL");
2717MODULE_VERSION(VXLAN_VERSION);
stephen hemminger3b8df3c2013-04-27 11:31:52 +00002718MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>");
stephen hemmingerd3428942012-10-01 12:32:35 +00002719MODULE_ALIAS_RTNL_LINK("vxlan");