blob: 2ba01ca02c9c5ed7f8024cbe7370127a3cb33a4c [file] [log] [blame]
stephen hemmingerd3428942012-10-01 12:32:35 +00001/*
Rami Roseneb5ce432012-11-13 13:29:15 +00002 * VXLAN: Virtual eXtensible Local Area Network
stephen hemmingerd3428942012-10-01 12:32:35 +00003 *
stephen hemminger3b8df3c2013-04-27 11:31:52 +00004 * Copyright (c) 2012-2013 Vyatta Inc.
stephen hemmingerd3428942012-10-01 12:32:35 +00005 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
stephen hemmingerd3428942012-10-01 12:32:35 +00009 */
10
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/kernel.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000014#include <linux/module.h>
15#include <linux/errno.h>
16#include <linux/slab.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000017#include <linux/udp.h>
18#include <linux/igmp.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000019#include <linux/if_ether.h>
Yan Burman1b13c972013-01-29 23:43:07 +000020#include <linux/ethtool.h>
David Stevense4f67ad2012-11-20 02:50:14 +000021#include <net/arp.h>
22#include <net/ndisc.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000023#include <net/ip.h>
24#include <net/icmp.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000025#include <net/rtnetlink.h>
stephen hemmingerd3428942012-10-01 12:32:35 +000026#include <net/inet_ecn.h>
27#include <net/net_namespace.h>
28#include <net/netns/generic.h>
Pravin B Shelar012a5722013-08-19 11:23:07 -070029#include <net/vxlan.h>
stephen hemminger40d29af2016-02-09 22:07:29 -080030
Cong Wange4c7ed42013-08-31 13:44:33 +080031#if IS_ENABLED(CONFIG_IPV6)
Cong Wange4c7ed42013-08-31 13:44:33 +080032#include <net/ip6_tunnel.h>
Cong Wang660d98c2013-09-02 10:06:52 +080033#include <net/ip6_checksum.h>
Cong Wange4c7ed42013-08-31 13:44:33 +080034#endif
stephen hemmingerd3428942012-10-01 12:32:35 +000035
36#define VXLAN_VERSION "0.1"
37
stephen hemminger553675f2013-05-16 11:35:20 +000038#define PORT_HASH_BITS 8
39#define PORT_HASH_SIZE (1<<PORT_HASH_BITS)
stephen hemmingerd3428942012-10-01 12:32:35 +000040#define FDB_AGE_DEFAULT 300 /* 5 min */
41#define FDB_AGE_INTERVAL (10 * HZ) /* rescan interval */
42
stephen hemminger23c578b2013-04-27 11:31:53 +000043/* UDP port for VXLAN traffic.
44 * The IANA assigned port is 4789, but the Linux default is 8472
Stephen Hemminger234f5b72013-06-17 14:16:41 -070045 * for compatibility with early adopters.
stephen hemminger23c578b2013-04-27 11:31:53 +000046 */
Stephen Hemminger9daaa392013-06-17 14:16:12 -070047static unsigned short vxlan_port __read_mostly = 8472;
48module_param_named(udp_port, vxlan_port, ushort, 0444);
stephen hemmingerd3428942012-10-01 12:32:35 +000049MODULE_PARM_DESC(udp_port, "Destination UDP port");
50
51static bool log_ecn_error = true;
52module_param(log_ecn_error, bool, 0644);
53MODULE_PARM_DESC(log_ecn_error, "Log packets received with corrupted ECN");
54
Pravin B Shelar60d9d4c2013-06-20 00:26:31 -070055static int vxlan_net_id;
Thomas Graf0dfbdf42015-07-21 10:44:02 +020056static struct rtnl_link_ops vxlan_link_ops;
stephen hemminger553675f2013-05-16 11:35:20 +000057
Li RongQing7256eac2016-01-29 09:43:47 +080058static const u8 all_zeros_mac[ETH_ALEN + 2];
Mike Rapoportafbd8ba2013-06-25 16:01:51 +030059
Jiri Benc205f3562015-09-24 13:50:01 +020060static int vxlan_sock_add(struct vxlan_dev *vxlan);
Thomas Graf614732e2015-07-21 10:44:06 +020061
stephen hemminger553675f2013-05-16 11:35:20 +000062/* per-network namespace private data for this module */
63struct vxlan_net {
64 struct list_head vxlan_list;
65 struct hlist_head sock_list[PORT_HASH_SIZE];
Stephen Hemminger1c51a912013-06-17 14:16:11 -070066 spinlock_t sock_lock;
stephen hemminger553675f2013-05-16 11:35:20 +000067};
68
stephen hemmingerd3428942012-10-01 12:32:35 +000069/* Forwarding table entry */
70struct vxlan_fdb {
71 struct hlist_node hlist; /* linked list of entries */
72 struct rcu_head rcu;
73 unsigned long updated; /* jiffies */
74 unsigned long used;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -070075 struct list_head remotes;
Sowmini Varadhan7177a3b2015-07-20 09:54:50 +020076 u8 eth_addr[ETH_ALEN];
stephen hemmingerd3428942012-10-01 12:32:35 +000077 u16 state; /* see ndm_state */
David Stevensae884082013-04-19 00:36:26 +000078 u8 flags; /* see ndm_flags */
stephen hemmingerd3428942012-10-01 12:32:35 +000079};
80
stephen hemmingerd3428942012-10-01 12:32:35 +000081/* salt for hash table */
82static u32 vxlan_salt __read_mostly;
83
Thomas Grafee122c72015-07-21 10:43:58 +020084static inline bool vxlan_collect_metadata(struct vxlan_sock *vs)
85{
Thomas Grafe7030872015-07-21 10:44:01 +020086 return vs->flags & VXLAN_F_COLLECT_METADATA ||
87 ip_tunnel_collect_metadata();
Thomas Grafee122c72015-07-21 10:43:58 +020088}
89
Cong Wange4c7ed42013-08-31 13:44:33 +080090#if IS_ENABLED(CONFIG_IPV6)
91static inline
92bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
93{
Jiri Bencf0ef3122015-03-29 16:17:37 +020094 if (a->sa.sa_family != b->sa.sa_family)
95 return false;
96 if (a->sa.sa_family == AF_INET6)
97 return ipv6_addr_equal(&a->sin6.sin6_addr, &b->sin6.sin6_addr);
98 else
99 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
Cong Wange4c7ed42013-08-31 13:44:33 +0800100}
101
102static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
103{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200104 if (ipa->sa.sa_family == AF_INET6)
105 return ipv6_addr_any(&ipa->sin6.sin6_addr);
106 else
107 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
Cong Wange4c7ed42013-08-31 13:44:33 +0800108}
109
110static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
111{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200112 if (ipa->sa.sa_family == AF_INET6)
113 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
114 else
115 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
Cong Wange4c7ed42013-08-31 13:44:33 +0800116}
117
118static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
119{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200120 if (nla_len(nla) >= sizeof(struct in6_addr)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200121 ip->sin6.sin6_addr = nla_get_in6_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200122 ip->sa.sa_family = AF_INET6;
123 return 0;
124 } else if (nla_len(nla) >= sizeof(__be32)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200125 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200126 ip->sa.sa_family = AF_INET;
127 return 0;
128 } else {
129 return -EAFNOSUPPORT;
130 }
Cong Wange4c7ed42013-08-31 13:44:33 +0800131}
132
133static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
Jiri Bencf0ef3122015-03-29 16:17:37 +0200134 const union vxlan_addr *ip)
Cong Wange4c7ed42013-08-31 13:44:33 +0800135{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200136 if (ip->sa.sa_family == AF_INET6)
Jiri Benc930345e2015-03-29 16:59:25 +0200137 return nla_put_in6_addr(skb, attr, &ip->sin6.sin6_addr);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200138 else
Jiri Benc930345e2015-03-29 16:59:25 +0200139 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +0800140}
141
142#else /* !CONFIG_IPV6 */
143
144static inline
145bool vxlan_addr_equal(const union vxlan_addr *a, const union vxlan_addr *b)
146{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200147 return a->sin.sin_addr.s_addr == b->sin.sin_addr.s_addr;
Cong Wange4c7ed42013-08-31 13:44:33 +0800148}
149
150static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
151{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200152 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
Cong Wange4c7ed42013-08-31 13:44:33 +0800153}
154
155static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
156{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200157 return IN_MULTICAST(ntohl(ipa->sin.sin_addr.s_addr));
Cong Wange4c7ed42013-08-31 13:44:33 +0800158}
159
160static int vxlan_nla_get_addr(union vxlan_addr *ip, struct nlattr *nla)
161{
Jiri Bencf0ef3122015-03-29 16:17:37 +0200162 if (nla_len(nla) >= sizeof(struct in6_addr)) {
163 return -EAFNOSUPPORT;
164 } else if (nla_len(nla) >= sizeof(__be32)) {
Jiri Benc67b61f62015-03-29 16:59:26 +0200165 ip->sin.sin_addr.s_addr = nla_get_in_addr(nla);
Jiri Bencf0ef3122015-03-29 16:17:37 +0200166 ip->sa.sa_family = AF_INET;
167 return 0;
168 } else {
169 return -EAFNOSUPPORT;
170 }
Cong Wange4c7ed42013-08-31 13:44:33 +0800171}
172
173static int vxlan_nla_put_addr(struct sk_buff *skb, int attr,
Jiri Bencf0ef3122015-03-29 16:17:37 +0200174 const union vxlan_addr *ip)
Cong Wange4c7ed42013-08-31 13:44:33 +0800175{
Jiri Benc930345e2015-03-29 16:59:25 +0200176 return nla_put_in_addr(skb, attr, ip->sin.sin_addr.s_addr);
Cong Wange4c7ed42013-08-31 13:44:33 +0800177}
178#endif
179
stephen hemminger553675f2013-05-16 11:35:20 +0000180/* Virtual Network hash table head */
Jiri Benc54bfd872016-02-16 21:58:58 +0100181static inline struct hlist_head *vni_head(struct vxlan_sock *vs, __be32 vni)
stephen hemminger553675f2013-05-16 11:35:20 +0000182{
Jiri Benc54bfd872016-02-16 21:58:58 +0100183 return &vs->vni_list[hash_32((__force u32)vni, VNI_HASH_BITS)];
stephen hemminger553675f2013-05-16 11:35:20 +0000184}
185
186/* Socket hash table head */
187static inline struct hlist_head *vs_head(struct net *net, __be16 port)
stephen hemmingerd3428942012-10-01 12:32:35 +0000188{
189 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
190
stephen hemminger553675f2013-05-16 11:35:20 +0000191 return &vn->sock_list[hash_32(ntohs(port), PORT_HASH_BITS)];
192}
193
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700194/* First remote destination for a forwarding entry.
195 * Guaranteed to be non-NULL because remotes are never deleted.
196 */
stephen hemminger5ca54612013-08-04 17:17:39 -0700197static inline struct vxlan_rdst *first_remote_rcu(struct vxlan_fdb *fdb)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700198{
stephen hemminger5ca54612013-08-04 17:17:39 -0700199 return list_entry_rcu(fdb->remotes.next, struct vxlan_rdst, list);
200}
201
202static inline struct vxlan_rdst *first_remote_rtnl(struct vxlan_fdb *fdb)
203{
204 return list_first_entry(&fdb->remotes, struct vxlan_rdst, list);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700205}
206
Thomas Grafac5132d2015-01-15 03:53:56 +0100207/* Find VXLAN socket based on network namespace, address family and UDP port
208 * and enabled unshareable flags.
209 */
210static struct vxlan_sock *vxlan_find_sock(struct net *net, sa_family_t family,
211 __be16 port, u32 flags)
stephen hemminger553675f2013-05-16 11:35:20 +0000212{
213 struct vxlan_sock *vs;
Tom Herbertaf33c1a2015-01-20 11:23:05 -0800214
215 flags &= VXLAN_F_RCV_FLAGS;
stephen hemminger553675f2013-05-16 11:35:20 +0000216
217 hlist_for_each_entry_rcu(vs, vs_head(net, port), hlist) {
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -0200218 if (inet_sk(vs->sock->sk)->inet_sport == port &&
Jiri Benc705cc622015-08-20 13:56:28 +0200219 vxlan_get_sk_family(vs) == family &&
Tom Herbertaf33c1a2015-01-20 11:23:05 -0800220 vs->flags == flags)
stephen hemminger553675f2013-05-16 11:35:20 +0000221 return vs;
222 }
223 return NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000224}
225
Jiri Benc54bfd872016-02-16 21:58:58 +0100226static struct vxlan_dev *vxlan_vs_find_vni(struct vxlan_sock *vs, __be32 vni)
stephen hemmingerd3428942012-10-01 12:32:35 +0000227{
228 struct vxlan_dev *vxlan;
stephen hemmingerd3428942012-10-01 12:32:35 +0000229
Jiri Bencb9167b22016-02-16 21:59:03 +0100230 /* For flow based devices, map all packets to VNI 0 */
231 if (vs->flags & VXLAN_F_COLLECT_METADATA)
232 vni = 0;
233
Jiri Benc54bfd872016-02-16 21:58:58 +0100234 hlist_for_each_entry_rcu(vxlan, vni_head(vs, vni), hlist) {
235 if (vxlan->default_dst.remote_vni == vni)
stephen hemmingerd3428942012-10-01 12:32:35 +0000236 return vxlan;
237 }
238
239 return NULL;
240}
241
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700242/* Look up VNI in a per net namespace table */
Jiri Benc54bfd872016-02-16 21:58:58 +0100243static struct vxlan_dev *vxlan_find_vni(struct net *net, __be32 vni,
Thomas Grafac5132d2015-01-15 03:53:56 +0100244 sa_family_t family, __be16 port,
245 u32 flags)
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700246{
247 struct vxlan_sock *vs;
248
Thomas Grafac5132d2015-01-15 03:53:56 +0100249 vs = vxlan_find_sock(net, family, port, flags);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700250 if (!vs)
251 return NULL;
252
Jiri Benc54bfd872016-02-16 21:58:58 +0100253 return vxlan_vs_find_vni(vs, vni);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -0700254}
255
stephen hemmingerd3428942012-10-01 12:32:35 +0000256/* Fill in neighbour message in skbuff. */
257static int vxlan_fdb_info(struct sk_buff *skb, struct vxlan_dev *vxlan,
Stephen Hemminger234f5b72013-06-17 14:16:41 -0700258 const struct vxlan_fdb *fdb,
259 u32 portid, u32 seq, int type, unsigned int flags,
260 const struct vxlan_rdst *rdst)
stephen hemmingerd3428942012-10-01 12:32:35 +0000261{
262 unsigned long now = jiffies;
263 struct nda_cacheinfo ci;
264 struct nlmsghdr *nlh;
265 struct ndmsg *ndm;
David Stevense4f67ad2012-11-20 02:50:14 +0000266 bool send_ip, send_eth;
stephen hemmingerd3428942012-10-01 12:32:35 +0000267
268 nlh = nlmsg_put(skb, portid, seq, type, sizeof(*ndm), flags);
269 if (nlh == NULL)
270 return -EMSGSIZE;
271
272 ndm = nlmsg_data(nlh);
273 memset(ndm, 0, sizeof(*ndm));
David Stevense4f67ad2012-11-20 02:50:14 +0000274
275 send_eth = send_ip = true;
276
277 if (type == RTM_GETNEIGH) {
278 ndm->ndm_family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +0800279 send_ip = !vxlan_addr_any(&rdst->remote_ip);
David Stevense4f67ad2012-11-20 02:50:14 +0000280 send_eth = !is_zero_ether_addr(fdb->eth_addr);
281 } else
282 ndm->ndm_family = AF_BRIDGE;
stephen hemmingerd3428942012-10-01 12:32:35 +0000283 ndm->ndm_state = fdb->state;
284 ndm->ndm_ifindex = vxlan->dev->ifindex;
David Stevensae884082013-04-19 00:36:26 +0000285 ndm->ndm_flags = fdb->flags;
Jun Zhao545469f2014-07-26 00:38:59 +0800286 ndm->ndm_type = RTN_UNICAST;
stephen hemmingerd3428942012-10-01 12:32:35 +0000287
Nicolas Dichtel193523b2015-01-20 15:15:47 +0100288 if (!net_eq(dev_net(vxlan->dev), vxlan->net) &&
Nicolas Dichtel49670822015-01-26 14:10:53 +0100289 nla_put_s32(skb, NDA_LINK_NETNSID,
WANG Cong38f507f2016-09-01 21:53:44 -0700290 peernet2id(dev_net(vxlan->dev), vxlan->net)))
Nicolas Dichtel193523b2015-01-20 15:15:47 +0100291 goto nla_put_failure;
292
David Stevense4f67ad2012-11-20 02:50:14 +0000293 if (send_eth && nla_put(skb, NDA_LLADDR, ETH_ALEN, &fdb->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000294 goto nla_put_failure;
295
Cong Wange4c7ed42013-08-31 13:44:33 +0800296 if (send_ip && vxlan_nla_put_addr(skb, NDA_DST, &rdst->remote_ip))
David Stevens66817122013-03-15 04:35:51 +0000297 goto nla_put_failure;
298
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200299 if (rdst->remote_port && rdst->remote_port != vxlan->cfg.dst_port &&
David Stevens66817122013-03-15 04:35:51 +0000300 nla_put_be16(skb, NDA_PORT, rdst->remote_port))
301 goto nla_put_failure;
Atzm Watanabec7995c42013-04-16 02:50:52 +0000302 if (rdst->remote_vni != vxlan->default_dst.remote_vni &&
Jiri Benc54bfd872016-02-16 21:58:58 +0100303 nla_put_u32(skb, NDA_VNI, be32_to_cpu(rdst->remote_vni)))
David Stevens66817122013-03-15 04:35:51 +0000304 goto nla_put_failure;
305 if (rdst->remote_ifindex &&
306 nla_put_u32(skb, NDA_IFINDEX, rdst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +0000307 goto nla_put_failure;
308
309 ci.ndm_used = jiffies_to_clock_t(now - fdb->used);
310 ci.ndm_confirmed = 0;
311 ci.ndm_updated = jiffies_to_clock_t(now - fdb->updated);
312 ci.ndm_refcnt = 0;
313
314 if (nla_put(skb, NDA_CACHEINFO, sizeof(ci), &ci))
315 goto nla_put_failure;
316
Johannes Berg053c0952015-01-16 22:09:00 +0100317 nlmsg_end(skb, nlh);
318 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +0000319
320nla_put_failure:
321 nlmsg_cancel(skb, nlh);
322 return -EMSGSIZE;
323}
324
325static inline size_t vxlan_nlmsg_size(void)
326{
327 return NLMSG_ALIGN(sizeof(struct ndmsg))
328 + nla_total_size(ETH_ALEN) /* NDA_LLADDR */
Cong Wange4c7ed42013-08-31 13:44:33 +0800329 + nla_total_size(sizeof(struct in6_addr)) /* NDA_DST */
stephen hemminger73cf3312013-04-27 11:31:54 +0000330 + nla_total_size(sizeof(__be16)) /* NDA_PORT */
David Stevens66817122013-03-15 04:35:51 +0000331 + nla_total_size(sizeof(__be32)) /* NDA_VNI */
332 + nla_total_size(sizeof(__u32)) /* NDA_IFINDEX */
Nicolas Dichtel49670822015-01-26 14:10:53 +0100333 + nla_total_size(sizeof(__s32)) /* NDA_LINK_NETNSID */
stephen hemmingerd3428942012-10-01 12:32:35 +0000334 + nla_total_size(sizeof(struct nda_cacheinfo));
335}
336
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200337static void vxlan_fdb_notify(struct vxlan_dev *vxlan, struct vxlan_fdb *fdb,
338 struct vxlan_rdst *rd, int type)
stephen hemmingerd3428942012-10-01 12:32:35 +0000339{
340 struct net *net = dev_net(vxlan->dev);
341 struct sk_buff *skb;
342 int err = -ENOBUFS;
343
344 skb = nlmsg_new(vxlan_nlmsg_size(), GFP_ATOMIC);
345 if (skb == NULL)
346 goto errout;
347
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200348 err = vxlan_fdb_info(skb, vxlan, fdb, 0, 0, type, 0, rd);
stephen hemmingerd3428942012-10-01 12:32:35 +0000349 if (err < 0) {
350 /* -EMSGSIZE implies BUG in vxlan_nlmsg_size() */
351 WARN_ON(err == -EMSGSIZE);
352 kfree_skb(skb);
353 goto errout;
354 }
355
356 rtnl_notify(skb, net, 0, RTNLGRP_NEIGH, NULL, GFP_ATOMIC);
357 return;
358errout:
359 if (err < 0)
360 rtnl_set_sk_err(net, RTNLGRP_NEIGH, err);
361}
362
Cong Wange4c7ed42013-08-31 13:44:33 +0800363static void vxlan_ip_miss(struct net_device *dev, union vxlan_addr *ipa)
David Stevense4f67ad2012-11-20 02:50:14 +0000364{
365 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700366 struct vxlan_fdb f = {
367 .state = NUD_STALE,
368 };
369 struct vxlan_rdst remote = {
Cong Wange4c7ed42013-08-31 13:44:33 +0800370 .remote_ip = *ipa, /* goes to NDA_DST */
Jiri Benc54bfd872016-02-16 21:58:58 +0100371 .remote_vni = cpu_to_be32(VXLAN_N_VID),
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700372 };
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700373
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200374 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH);
David Stevense4f67ad2012-11-20 02:50:14 +0000375}
376
377static void vxlan_fdb_miss(struct vxlan_dev *vxlan, const u8 eth_addr[ETH_ALEN])
378{
Stephen Hemmingerbb3fd682013-06-17 14:16:40 -0700379 struct vxlan_fdb f = {
380 .state = NUD_STALE,
381 };
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200382 struct vxlan_rdst remote = { };
David Stevense4f67ad2012-11-20 02:50:14 +0000383
David Stevense4f67ad2012-11-20 02:50:14 +0000384 memcpy(f.eth_addr, eth_addr, ETH_ALEN);
385
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200386 vxlan_fdb_notify(vxlan, &f, &remote, RTM_GETNEIGH);
David Stevense4f67ad2012-11-20 02:50:14 +0000387}
388
stephen hemmingerd3428942012-10-01 12:32:35 +0000389/* Hash Ethernet address */
390static u32 eth_hash(const unsigned char *addr)
391{
392 u64 value = get_unaligned((u64 *)addr);
393
394 /* only want 6 bytes */
395#ifdef __BIG_ENDIAN
stephen hemmingerd3428942012-10-01 12:32:35 +0000396 value >>= 16;
stephen hemminger321fb992012-10-09 20:35:47 +0000397#else
398 value <<= 16;
stephen hemmingerd3428942012-10-01 12:32:35 +0000399#endif
400 return hash_64(value, FDB_HASH_BITS);
401}
402
403/* Hash chain to use given mac address */
404static inline struct hlist_head *vxlan_fdb_head(struct vxlan_dev *vxlan,
405 const u8 *mac)
406{
407 return &vxlan->fdb_head[eth_hash(mac)];
408}
409
410/* Look up Ethernet address in forwarding table */
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000411static struct vxlan_fdb *__vxlan_find_mac(struct vxlan_dev *vxlan,
stephen hemmingerd3428942012-10-01 12:32:35 +0000412 const u8 *mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000413{
414 struct hlist_head *head = vxlan_fdb_head(vxlan, mac);
415 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000416
Sasha Levinb67bfe02013-02-27 17:06:00 -0800417 hlist_for_each_entry_rcu(f, head, hlist) {
Joe Perches7367d0b2013-09-01 11:51:23 -0700418 if (ether_addr_equal(mac, f->eth_addr))
stephen hemmingerd3428942012-10-01 12:32:35 +0000419 return f;
420 }
421
422 return NULL;
423}
424
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000425static struct vxlan_fdb *vxlan_find_mac(struct vxlan_dev *vxlan,
426 const u8 *mac)
427{
428 struct vxlan_fdb *f;
429
430 f = __vxlan_find_mac(vxlan, mac);
431 if (f)
432 f->used = jiffies;
433
434 return f;
435}
436
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300437/* caller should hold vxlan->hash_lock */
438static struct vxlan_rdst *vxlan_fdb_find_rdst(struct vxlan_fdb *f,
Cong Wange4c7ed42013-08-31 13:44:33 +0800439 union vxlan_addr *ip, __be16 port,
Jiri Benc54bfd872016-02-16 21:58:58 +0100440 __be32 vni, __u32 ifindex)
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300441{
442 struct vxlan_rdst *rd;
443
444 list_for_each_entry(rd, &f->remotes, list) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800445 if (vxlan_addr_equal(&rd->remote_ip, ip) &&
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300446 rd->remote_port == port &&
447 rd->remote_vni == vni &&
448 rd->remote_ifindex == ifindex)
449 return rd;
450 }
451
452 return NULL;
453}
454
Thomas Richter906dc182013-07-19 17:20:07 +0200455/* Replace destination of unicast mac */
456static int vxlan_fdb_replace(struct vxlan_fdb *f,
Jiri Benc54bfd872016-02-16 21:58:58 +0100457 union vxlan_addr *ip, __be16 port, __be32 vni,
458 __u32 ifindex)
Thomas Richter906dc182013-07-19 17:20:07 +0200459{
460 struct vxlan_rdst *rd;
461
462 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
463 if (rd)
464 return 0;
465
466 rd = list_first_entry_or_null(&f->remotes, struct vxlan_rdst, list);
467 if (!rd)
468 return 0;
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100469
470 dst_cache_reset(&rd->dst_cache);
Cong Wange4c7ed42013-08-31 13:44:33 +0800471 rd->remote_ip = *ip;
Thomas Richter906dc182013-07-19 17:20:07 +0200472 rd->remote_port = port;
473 rd->remote_vni = vni;
474 rd->remote_ifindex = ifindex;
475 return 1;
476}
477
David Stevens66817122013-03-15 04:35:51 +0000478/* Add/update destinations for multicast */
479static int vxlan_fdb_append(struct vxlan_fdb *f,
Jiri Benc54bfd872016-02-16 21:58:58 +0100480 union vxlan_addr *ip, __be16 port, __be32 vni,
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200481 __u32 ifindex, struct vxlan_rdst **rdp)
David Stevens66817122013-03-15 04:35:51 +0000482{
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700483 struct vxlan_rdst *rd;
David Stevens66817122013-03-15 04:35:51 +0000484
Mike Rapoporta5e7c102013-06-25 16:01:52 +0300485 rd = vxlan_fdb_find_rdst(f, ip, port, vni, ifindex);
486 if (rd)
487 return 0;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700488
David Stevens66817122013-03-15 04:35:51 +0000489 rd = kmalloc(sizeof(*rd), GFP_ATOMIC);
490 if (rd == NULL)
491 return -ENOBUFS;
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100492
493 if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) {
494 kfree(rd);
495 return -ENOBUFS;
496 }
497
Cong Wange4c7ed42013-08-31 13:44:33 +0800498 rd->remote_ip = *ip;
David Stevens66817122013-03-15 04:35:51 +0000499 rd->remote_port = port;
500 rd->remote_vni = vni;
501 rd->remote_ifindex = ifindex;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700502
503 list_add_tail_rcu(&rd->list, &f->remotes);
504
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200505 *rdp = rd;
David Stevens66817122013-03-15 04:35:51 +0000506 return 1;
507}
508
Tom Herbertdfd86452015-01-12 17:00:38 -0800509static struct vxlanhdr *vxlan_gro_remcsum(struct sk_buff *skb,
510 unsigned int off,
511 struct vxlanhdr *vh, size_t hdrlen,
Jiri Benc54bfd872016-02-16 21:58:58 +0100512 __be32 vni_field,
513 struct gro_remcsum *grc,
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800514 bool nopartial)
Tom Herbertdfd86452015-01-12 17:00:38 -0800515{
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700516 size_t start, offset;
Tom Herbertdfd86452015-01-12 17:00:38 -0800517
518 if (skb->remcsum_offload)
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700519 return vh;
Tom Herbertdfd86452015-01-12 17:00:38 -0800520
521 if (!NAPI_GRO_CB(skb)->csum_valid)
522 return NULL;
523
Jiri Benc54bfd872016-02-16 21:58:58 +0100524 start = vxlan_rco_start(vni_field);
525 offset = start + vxlan_rco_offset(vni_field);
Tom Herbertdfd86452015-01-12 17:00:38 -0800526
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700527 vh = skb_gro_remcsum_process(skb, (void *)vh, off, hdrlen,
528 start, offset, grc, nopartial);
Tom Herbertdfd86452015-01-12 17:00:38 -0800529
530 skb->remcsum_offload = 1;
531
532 return vh;
533}
534
Tom Herbert5602c482016-04-05 08:22:53 -0700535static struct sk_buff **vxlan_gro_receive(struct sock *sk,
536 struct sk_buff **head,
537 struct sk_buff *skb)
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200538{
539 struct sk_buff *p, **pp = NULL;
540 struct vxlanhdr *vh, *vh2;
Jesse Gross9b174d82014-12-30 19:10:15 -0800541 unsigned int hlen, off_vx;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200542 int flush = 1;
Tom Herbert5602c482016-04-05 08:22:53 -0700543 struct vxlan_sock *vs = rcu_dereference_sk_user_data(sk);
Jiri Benc54bfd872016-02-16 21:58:58 +0100544 __be32 flags;
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800545 struct gro_remcsum grc;
546
547 skb_gro_remcsum_init(&grc);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200548
549 off_vx = skb_gro_offset(skb);
550 hlen = off_vx + sizeof(*vh);
551 vh = skb_gro_header_fast(skb, off_vx);
552 if (skb_gro_header_hard(skb, hlen)) {
553 vh = skb_gro_header_slow(skb, hlen, off_vx);
554 if (unlikely(!vh))
555 goto out;
556 }
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200557
Tom Herbertdfd86452015-01-12 17:00:38 -0800558 skb_gro_postpull_rcsum(skb, vh, sizeof(struct vxlanhdr));
559
Jiri Benc54bfd872016-02-16 21:58:58 +0100560 flags = vh->vx_flags;
Tom Herbertdfd86452015-01-12 17:00:38 -0800561
562 if ((flags & VXLAN_HF_RCO) && (vs->flags & VXLAN_F_REMCSUM_RX)) {
563 vh = vxlan_gro_remcsum(skb, off_vx, vh, sizeof(struct vxlanhdr),
Jiri Benc54bfd872016-02-16 21:58:58 +0100564 vh->vx_vni, &grc,
Tom Herbert0ace2ca2015-02-10 16:30:32 -0800565 !!(vs->flags &
566 VXLAN_F_REMCSUM_NOPARTIAL));
Tom Herbertdfd86452015-01-12 17:00:38 -0800567
568 if (!vh)
569 goto out;
570 }
571
Tom Herbertb7fe10e2015-08-19 17:07:32 -0700572 skb_gro_pull(skb, sizeof(struct vxlanhdr)); /* pull vxlan header */
573
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200574 for (p = *head; p; p = p->next) {
575 if (!NAPI_GRO_CB(p)->same_flow)
576 continue;
577
578 vh2 = (struct vxlanhdr *)(p->data + off_vx);
Thomas Graf35114942015-01-15 03:53:55 +0100579 if (vh->vx_flags != vh2->vx_flags ||
580 vh->vx_vni != vh2->vx_vni) {
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200581 NAPI_GRO_CB(p)->same_flow = 0;
582 continue;
583 }
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200584 }
585
Sabrina Dubrocafcd91dd2016-10-20 15:58:02 +0200586 pp = call_gro_receive(eth_gro_receive, head, skb);
Alexander Duyckc194cf92016-03-09 09:24:23 -0800587 flush = 0;
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200588
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200589out:
Tom Herbert26c4f7d2015-02-10 16:30:27 -0800590 skb_gro_remcsum_cleanup(skb, &grc);
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200591 NAPI_GRO_CB(skb)->flush |= flush;
592
593 return pp;
594}
595
Tom Herbert5602c482016-04-05 08:22:53 -0700596static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200597{
Jarno Rajahalme229740c2016-05-03 16:10:21 -0700598 /* Sets 'skb->inner_mac_header' since we are always called with
599 * 'skb->encapsulation' set.
600 */
Jesse Gross9b174d82014-12-30 19:10:15 -0800601 return eth_gro_complete(skb, nhoff + sizeof(struct vxlanhdr));
Or Gerlitzdc01e7d2014-01-20 13:59:21 +0200602}
603
stephen hemmingerd3428942012-10-01 12:32:35 +0000604/* Add new entry to forwarding table -- assumes lock held */
605static int vxlan_fdb_create(struct vxlan_dev *vxlan,
Cong Wange4c7ed42013-08-31 13:44:33 +0800606 const u8 *mac, union vxlan_addr *ip,
David Stevens66817122013-03-15 04:35:51 +0000607 __u16 state, __u16 flags,
Jiri Benc54bfd872016-02-16 21:58:58 +0100608 __be16 port, __be32 vni, __u32 ifindex,
David Stevensae884082013-04-19 00:36:26 +0000609 __u8 ndm_flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000610{
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200611 struct vxlan_rdst *rd = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +0000612 struct vxlan_fdb *f;
613 int notify = 0;
Haishuang Yan17b46362016-11-29 09:59:36 +0800614 int rc;
stephen hemmingerd3428942012-10-01 12:32:35 +0000615
Sridhar Samudrala014be2c2013-05-17 06:39:07 +0000616 f = __vxlan_find_mac(vxlan, mac);
stephen hemmingerd3428942012-10-01 12:32:35 +0000617 if (f) {
618 if (flags & NLM_F_EXCL) {
619 netdev_dbg(vxlan->dev,
620 "lost race to create %pM\n", mac);
621 return -EEXIST;
622 }
623 if (f->state != state) {
624 f->state = state;
625 f->updated = jiffies;
626 notify = 1;
627 }
David Stevensae884082013-04-19 00:36:26 +0000628 if (f->flags != ndm_flags) {
629 f->flags = ndm_flags;
630 f->updated = jiffies;
631 notify = 1;
632 }
Thomas Richter906dc182013-07-19 17:20:07 +0200633 if ((flags & NLM_F_REPLACE)) {
634 /* Only change unicasts */
635 if (!(is_multicast_ether_addr(f->eth_addr) ||
636 is_zero_ether_addr(f->eth_addr))) {
Li RongQing60840422015-04-22 15:49:10 +0800637 notify |= vxlan_fdb_replace(f, ip, port, vni,
Thomas Richter906dc182013-07-19 17:20:07 +0200638 ifindex);
Thomas Richter906dc182013-07-19 17:20:07 +0200639 } else
640 return -EOPNOTSUPP;
641 }
David Stevens66817122013-03-15 04:35:51 +0000642 if ((flags & NLM_F_APPEND) &&
Mike Rapoport58e4c762013-06-25 16:01:56 +0300643 (is_multicast_ether_addr(f->eth_addr) ||
644 is_zero_ether_addr(f->eth_addr))) {
Haishuang Yan17b46362016-11-29 09:59:36 +0800645 rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
David Stevens66817122013-03-15 04:35:51 +0000646
647 if (rc < 0)
648 return rc;
649 notify |= rc;
650 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000651 } else {
652 if (!(flags & NLM_F_CREATE))
653 return -ENOENT;
654
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200655 if (vxlan->cfg.addrmax &&
656 vxlan->addrcnt >= vxlan->cfg.addrmax)
stephen hemmingerd3428942012-10-01 12:32:35 +0000657 return -ENOSPC;
658
Thomas Richter906dc182013-07-19 17:20:07 +0200659 /* Disallow replace to add a multicast entry */
660 if ((flags & NLM_F_REPLACE) &&
661 (is_multicast_ether_addr(mac) || is_zero_ether_addr(mac)))
662 return -EOPNOTSUPP;
663
Cong Wange4c7ed42013-08-31 13:44:33 +0800664 netdev_dbg(vxlan->dev, "add %pM -> %pIS\n", mac, ip);
stephen hemmingerd3428942012-10-01 12:32:35 +0000665 f = kmalloc(sizeof(*f), GFP_ATOMIC);
666 if (!f)
667 return -ENOMEM;
668
669 notify = 1;
stephen hemmingerd3428942012-10-01 12:32:35 +0000670 f->state = state;
David Stevensae884082013-04-19 00:36:26 +0000671 f->flags = ndm_flags;
stephen hemmingerd3428942012-10-01 12:32:35 +0000672 f->updated = f->used = jiffies;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700673 INIT_LIST_HEAD(&f->remotes);
stephen hemmingerd3428942012-10-01 12:32:35 +0000674 memcpy(f->eth_addr, mac, ETH_ALEN);
675
Haishuang Yan17b46362016-11-29 09:59:36 +0800676 rc = vxlan_fdb_append(f, ip, port, vni, ifindex, &rd);
677 if (rc < 0) {
678 kfree(f);
679 return rc;
680 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700681
stephen hemmingerd3428942012-10-01 12:32:35 +0000682 ++vxlan->addrcnt;
683 hlist_add_head_rcu(&f->hlist,
684 vxlan_fdb_head(vxlan, mac));
685 }
686
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200687 if (notify) {
688 if (rd == NULL)
689 rd = first_remote_rtnl(f);
690 vxlan_fdb_notify(vxlan, f, rd, RTM_NEWNEIGH);
691 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000692
693 return 0;
694}
695
Wei Yongjun6706c822013-04-11 19:00:35 +0000696static void vxlan_fdb_free(struct rcu_head *head)
David Stevens66817122013-03-15 04:35:51 +0000697{
698 struct vxlan_fdb *f = container_of(head, struct vxlan_fdb, rcu);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700699 struct vxlan_rdst *rd, *nd;
David Stevens66817122013-03-15 04:35:51 +0000700
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100701 list_for_each_entry_safe(rd, nd, &f->remotes, list) {
702 dst_cache_destroy(&rd->dst_cache);
David Stevens66817122013-03-15 04:35:51 +0000703 kfree(rd);
Paolo Abeni0c1d70a2016-02-12 15:43:56 +0100704 }
David Stevens66817122013-03-15 04:35:51 +0000705 kfree(f);
706}
707
stephen hemmingerd3428942012-10-01 12:32:35 +0000708static void vxlan_fdb_destroy(struct vxlan_dev *vxlan, struct vxlan_fdb *f)
709{
710 netdev_dbg(vxlan->dev,
711 "delete %pM\n", f->eth_addr);
712
713 --vxlan->addrcnt;
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200714 vxlan_fdb_notify(vxlan, f, first_remote_rtnl(f), RTM_DELNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +0000715
716 hlist_del_rcu(&f->hlist);
David Stevens66817122013-03-15 04:35:51 +0000717 call_rcu(&f->rcu, vxlan_fdb_free);
stephen hemmingerd3428942012-10-01 12:32:35 +0000718}
719
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300720static int vxlan_fdb_parse(struct nlattr *tb[], struct vxlan_dev *vxlan,
Jiri Benc54bfd872016-02-16 21:58:58 +0100721 union vxlan_addr *ip, __be16 *port, __be32 *vni,
722 u32 *ifindex)
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300723{
724 struct net *net = dev_net(vxlan->dev);
Cong Wange4c7ed42013-08-31 13:44:33 +0800725 int err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300726
727 if (tb[NDA_DST]) {
Cong Wange4c7ed42013-08-31 13:44:33 +0800728 err = vxlan_nla_get_addr(ip, tb[NDA_DST]);
729 if (err)
730 return err;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300731 } else {
Cong Wange4c7ed42013-08-31 13:44:33 +0800732 union vxlan_addr *remote = &vxlan->default_dst.remote_ip;
733 if (remote->sa.sa_family == AF_INET) {
734 ip->sin.sin_addr.s_addr = htonl(INADDR_ANY);
735 ip->sa.sa_family = AF_INET;
736#if IS_ENABLED(CONFIG_IPV6)
737 } else {
738 ip->sin6.sin6_addr = in6addr_any;
739 ip->sa.sa_family = AF_INET6;
740#endif
741 }
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300742 }
743
744 if (tb[NDA_PORT]) {
745 if (nla_len(tb[NDA_PORT]) != sizeof(__be16))
746 return -EINVAL;
747 *port = nla_get_be16(tb[NDA_PORT]);
748 } else {
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200749 *port = vxlan->cfg.dst_port;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300750 }
751
752 if (tb[NDA_VNI]) {
753 if (nla_len(tb[NDA_VNI]) != sizeof(u32))
754 return -EINVAL;
Jiri Benc54bfd872016-02-16 21:58:58 +0100755 *vni = cpu_to_be32(nla_get_u32(tb[NDA_VNI]));
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300756 } else {
757 *vni = vxlan->default_dst.remote_vni;
758 }
759
760 if (tb[NDA_IFINDEX]) {
761 struct net_device *tdev;
762
763 if (nla_len(tb[NDA_IFINDEX]) != sizeof(u32))
764 return -EINVAL;
765 *ifindex = nla_get_u32(tb[NDA_IFINDEX]);
Ying Xue73763942014-01-15 10:23:41 +0800766 tdev = __dev_get_by_index(net, *ifindex);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300767 if (!tdev)
768 return -EADDRNOTAVAIL;
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300769 } else {
770 *ifindex = 0;
771 }
772
773 return 0;
774}
775
stephen hemmingerd3428942012-10-01 12:32:35 +0000776/* Add static entry (via netlink) */
777static int vxlan_fdb_add(struct ndmsg *ndm, struct nlattr *tb[],
778 struct net_device *dev,
Jiri Pirkof6f64242014-11-28 14:34:15 +0100779 const unsigned char *addr, u16 vid, u16 flags)
stephen hemmingerd3428942012-10-01 12:32:35 +0000780{
781 struct vxlan_dev *vxlan = netdev_priv(dev);
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300782 /* struct net *net = dev_net(vxlan->dev); */
Cong Wange4c7ed42013-08-31 13:44:33 +0800783 union vxlan_addr ip;
stephen hemminger73cf3312013-04-27 11:31:54 +0000784 __be16 port;
Jiri Benc54bfd872016-02-16 21:58:58 +0100785 __be32 vni;
786 u32 ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +0000787 int err;
788
789 if (!(ndm->ndm_state & (NUD_PERMANENT|NUD_REACHABLE))) {
790 pr_info("RTM_NEWNEIGH with invalid state %#x\n",
791 ndm->ndm_state);
792 return -EINVAL;
793 }
794
795 if (tb[NDA_DST] == NULL)
796 return -EINVAL;
797
Mike Rapoportf0b074b2013-06-25 16:01:53 +0300798 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
799 if (err)
800 return err;
David Stevens66817122013-03-15 04:35:51 +0000801
Mike Rapoport5933a7b2014-04-01 09:23:01 +0300802 if (vxlan->default_dst.remote_ip.sa.sa_family != ip.sa.sa_family)
803 return -EAFNOSUPPORT;
804
stephen hemmingerd3428942012-10-01 12:32:35 +0000805 spin_lock_bh(&vxlan->hash_lock);
Cong Wange4c7ed42013-08-31 13:44:33 +0800806 err = vxlan_fdb_create(vxlan, addr, &ip, ndm->ndm_state, flags,
stephen hemminger73cf3312013-04-27 11:31:54 +0000807 port, vni, ifindex, ndm->ndm_flags);
stephen hemmingerd3428942012-10-01 12:32:35 +0000808 spin_unlock_bh(&vxlan->hash_lock);
809
810 return err;
811}
812
813/* Delete entry (via netlink) */
Vlad Yasevich1690be62013-02-13 12:00:18 +0000814static int vxlan_fdb_delete(struct ndmsg *ndm, struct nlattr *tb[],
815 struct net_device *dev,
Jiri Pirkof6f64242014-11-28 14:34:15 +0100816 const unsigned char *addr, u16 vid)
stephen hemmingerd3428942012-10-01 12:32:35 +0000817{
818 struct vxlan_dev *vxlan = netdev_priv(dev);
819 struct vxlan_fdb *f;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300820 struct vxlan_rdst *rd = NULL;
Cong Wange4c7ed42013-08-31 13:44:33 +0800821 union vxlan_addr ip;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300822 __be16 port;
Jiri Benc54bfd872016-02-16 21:58:58 +0100823 __be32 vni;
824 u32 ifindex;
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300825 int err;
826
827 err = vxlan_fdb_parse(tb, vxlan, &ip, &port, &vni, &ifindex);
828 if (err)
829 return err;
830
831 err = -ENOENT;
stephen hemmingerd3428942012-10-01 12:32:35 +0000832
833 spin_lock_bh(&vxlan->hash_lock);
834 f = vxlan_find_mac(vxlan, addr);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300835 if (!f)
836 goto out;
837
Cong Wange4c7ed42013-08-31 13:44:33 +0800838 if (!vxlan_addr_any(&ip)) {
839 rd = vxlan_fdb_find_rdst(f, &ip, port, vni, ifindex);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300840 if (!rd)
841 goto out;
stephen hemmingerd3428942012-10-01 12:32:35 +0000842 }
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300843
844 err = 0;
845
846 /* remove a destination if it's not the only one on the list,
847 * otherwise destroy the fdb entry
848 */
849 if (rd && !list_is_singular(&f->remotes)) {
850 list_del_rcu(&rd->list);
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200851 vxlan_fdb_notify(vxlan, f, rd, RTM_DELNEIGH);
Wei Yongjundcdc7a52013-08-17 07:32:09 +0800852 kfree_rcu(rd, rcu);
Mike Rapoportbc7892b2013-06-25 16:01:54 +0300853 goto out;
854 }
855
856 vxlan_fdb_destroy(vxlan, f);
857
858out:
stephen hemmingerd3428942012-10-01 12:32:35 +0000859 spin_unlock_bh(&vxlan->hash_lock);
860
861 return err;
862}
863
864/* Dump forwarding table */
865static int vxlan_fdb_dump(struct sk_buff *skb, struct netlink_callback *cb,
Jamal Hadi Salim5d5eacb2014-07-10 07:01:58 -0400866 struct net_device *dev,
Roopa Prabhud2976532016-08-30 21:56:45 -0700867 struct net_device *filter_dev, int *idx)
stephen hemmingerd3428942012-10-01 12:32:35 +0000868{
869 struct vxlan_dev *vxlan = netdev_priv(dev);
870 unsigned int h;
Roopa Prabhud2976532016-08-30 21:56:45 -0700871 int err = 0;
stephen hemmingerd3428942012-10-01 12:32:35 +0000872
873 for (h = 0; h < FDB_HASH_SIZE; ++h) {
874 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000875
Sasha Levinb67bfe02013-02-27 17:06:00 -0800876 hlist_for_each_entry_rcu(f, &vxlan->fdb_head[h], hlist) {
David Stevens66817122013-03-15 04:35:51 +0000877 struct vxlan_rdst *rd;
stephen hemmingerd3428942012-10-01 12:32:35 +0000878
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700879 list_for_each_entry_rcu(rd, &f->remotes, list) {
Roopa Prabhud2976532016-08-30 21:56:45 -0700880 if (*idx < cb->args[2])
Atzm Watanabe07a51cd2015-08-10 23:39:09 +0900881 goto skip;
882
David Stevens66817122013-03-15 04:35:51 +0000883 err = vxlan_fdb_info(skb, vxlan, f,
884 NETLINK_CB(cb->skb).portid,
885 cb->nlh->nlmsg_seq,
886 RTM_NEWNEIGH,
887 NLM_F_MULTI, rd);
Roopa Prabhud2976532016-08-30 21:56:45 -0700888 if (err < 0)
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700889 goto out;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700890skip:
Roopa Prabhud2976532016-08-30 21:56:45 -0700891 *idx += 1;
Atzm Watanabe07a51cd2015-08-10 23:39:09 +0900892 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000893 }
894 }
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700895out:
Roopa Prabhud2976532016-08-30 21:56:45 -0700896 return err;
stephen hemmingerd3428942012-10-01 12:32:35 +0000897}
898
899/* Watch incoming packets to learn mapping between Ethernet address
900 * and Tunnel endpoint.
Simon Hormanc4b49512015-04-02 11:17:58 +0900901 * Return true if packet is bogus and should be dropped.
stephen hemmingerd3428942012-10-01 12:32:35 +0000902 */
stephen hemminger26a41ae62013-06-17 12:09:58 -0700903static bool vxlan_snoop(struct net_device *dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800904 union vxlan_addr *src_ip, const u8 *src_mac)
stephen hemmingerd3428942012-10-01 12:32:35 +0000905{
906 struct vxlan_dev *vxlan = netdev_priv(dev);
907 struct vxlan_fdb *f;
stephen hemmingerd3428942012-10-01 12:32:35 +0000908
909 f = vxlan_find_mac(vxlan, src_mac);
910 if (likely(f)) {
stephen hemminger5ca54612013-08-04 17:17:39 -0700911 struct vxlan_rdst *rdst = first_remote_rcu(f);
Stephen Hemminger3e61aa82013-06-17 14:16:12 -0700912
Cong Wange4c7ed42013-08-31 13:44:33 +0800913 if (likely(vxlan_addr_equal(&rdst->remote_ip, src_ip)))
stephen hemminger26a41ae62013-06-17 12:09:58 -0700914 return false;
915
916 /* Don't migrate static entries, drop packets */
stephen hemmingereb064c32013-06-18 14:27:01 -0700917 if (f->state & NUD_NOARP)
stephen hemminger26a41ae62013-06-17 12:09:58 -0700918 return true;
stephen hemmingerd3428942012-10-01 12:32:35 +0000919
920 if (net_ratelimit())
921 netdev_info(dev,
Cong Wange4c7ed42013-08-31 13:44:33 +0800922 "%pM migrated from %pIS to %pIS\n",
Rasmus Villemoesa4870f72015-02-07 03:17:31 +0100923 src_mac, &rdst->remote_ip.sa, &src_ip->sa);
stephen hemmingerd3428942012-10-01 12:32:35 +0000924
Cong Wange4c7ed42013-08-31 13:44:33 +0800925 rdst->remote_ip = *src_ip;
stephen hemmingerd3428942012-10-01 12:32:35 +0000926 f->updated = jiffies;
Nicolas Dichtel9e4b93f2014-04-22 15:01:30 +0200927 vxlan_fdb_notify(vxlan, f, rdst, RTM_NEWNEIGH);
stephen hemmingerd3428942012-10-01 12:32:35 +0000928 } else {
929 /* learned new entry */
930 spin_lock(&vxlan->hash_lock);
stephen hemminger3bf74b12013-06-17 12:09:57 -0700931
932 /* close off race between vxlan_flush and incoming packets */
933 if (netif_running(dev))
934 vxlan_fdb_create(vxlan, src_mac, src_ip,
935 NUD_REACHABLE,
936 NLM_F_EXCL|NLM_F_CREATE,
Thomas Graf0dfbdf42015-07-21 10:44:02 +0200937 vxlan->cfg.dst_port,
stephen hemminger3bf74b12013-06-17 12:09:57 -0700938 vxlan->default_dst.remote_vni,
939 0, NTF_SELF);
stephen hemmingerd3428942012-10-01 12:32:35 +0000940 spin_unlock(&vxlan->hash_lock);
941 }
stephen hemminger26a41ae62013-06-17 12:09:58 -0700942
943 return false;
stephen hemmingerd3428942012-10-01 12:32:35 +0000944}
945
stephen hemmingerd3428942012-10-01 12:32:35 +0000946/* See if multicast group is already in use by other ID */
Gao feng95ab0992013-12-10 16:37:33 +0800947static bool vxlan_group_used(struct vxlan_net *vn, struct vxlan_dev *dev)
stephen hemmingerd3428942012-10-01 12:32:35 +0000948{
stephen hemminger553675f2013-05-16 11:35:20 +0000949 struct vxlan_dev *vxlan;
pravin shelarc6fcc4f2016-10-28 09:59:15 -0700950 struct vxlan_sock *sock4;
Arnd Bergmann4053ab12016-11-07 22:09:07 +0100951#if IS_ENABLED(CONFIG_IPV6)
952 struct vxlan_sock *sock6;
953#endif
Jiri Bencb1be00a2015-09-24 13:50:02 +0200954 unsigned short family = dev->default_dst.remote_ip.sa.sa_family;
stephen hemmingerd3428942012-10-01 12:32:35 +0000955
pravin shelarc6fcc4f2016-10-28 09:59:15 -0700956 sock4 = rtnl_dereference(dev->vn4_sock);
957
Gao feng95ab0992013-12-10 16:37:33 +0800958 /* The vxlan_sock is only used by dev, leaving group has
959 * no effect on other vxlan devices.
960 */
pravin shelarc6fcc4f2016-10-28 09:59:15 -0700961 if (family == AF_INET && sock4 && atomic_read(&sock4->refcnt) == 1)
Gao feng95ab0992013-12-10 16:37:33 +0800962 return false;
Jiri Bencb1be00a2015-09-24 13:50:02 +0200963#if IS_ENABLED(CONFIG_IPV6)
pravin shelarc6fcc4f2016-10-28 09:59:15 -0700964 sock6 = rtnl_dereference(dev->vn6_sock);
965 if (family == AF_INET6 && sock6 && atomic_read(&sock6->refcnt) == 1)
Jiri Bencb1be00a2015-09-24 13:50:02 +0200966 return false;
967#endif
Gao feng95ab0992013-12-10 16:37:33 +0800968
stephen hemminger553675f2013-05-16 11:35:20 +0000969 list_for_each_entry(vxlan, &vn->vxlan_list, next) {
Gao feng95ab0992013-12-10 16:37:33 +0800970 if (!netif_running(vxlan->dev) || vxlan == dev)
stephen hemminger553675f2013-05-16 11:35:20 +0000971 continue;
stephen hemmingerd3428942012-10-01 12:32:35 +0000972
pravin shelarc6fcc4f2016-10-28 09:59:15 -0700973 if (family == AF_INET &&
974 rtnl_dereference(vxlan->vn4_sock) != sock4)
Gao feng95ab0992013-12-10 16:37:33 +0800975 continue;
Jiri Bencb1be00a2015-09-24 13:50:02 +0200976#if IS_ENABLED(CONFIG_IPV6)
pravin shelarc6fcc4f2016-10-28 09:59:15 -0700977 if (family == AF_INET6 &&
978 rtnl_dereference(vxlan->vn6_sock) != sock6)
Jiri Bencb1be00a2015-09-24 13:50:02 +0200979 continue;
980#endif
Gao feng95ab0992013-12-10 16:37:33 +0800981
982 if (!vxlan_addr_equal(&vxlan->default_dst.remote_ip,
983 &dev->default_dst.remote_ip))
984 continue;
985
986 if (vxlan->default_dst.remote_ifindex !=
987 dev->default_dst.remote_ifindex)
988 continue;
989
990 return true;
stephen hemminger553675f2013-05-16 11:35:20 +0000991 }
stephen hemmingerd3428942012-10-01 12:32:35 +0000992
993 return false;
994}
995
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +0200996static bool __vxlan_sock_release_prep(struct vxlan_sock *vs)
stephen hemmingerd3428942012-10-01 12:32:35 +0000997{
Jiri Bencb1be00a2015-09-24 13:50:02 +0200998 struct vxlan_net *vn;
Pravin B Shelar012a5722013-08-19 11:23:07 -0700999
Jiri Bencb1be00a2015-09-24 13:50:02 +02001000 if (!vs)
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001001 return false;
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001002 if (!atomic_dec_and_test(&vs->refcnt))
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001003 return false;
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001004
Jiri Bencb1be00a2015-09-24 13:50:02 +02001005 vn = net_generic(sock_net(vs->sock->sk), vxlan_net_id);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001006 spin_lock(&vn->sock_lock);
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001007 hlist_del_rcu(&vs->hlist);
Alexander Duycke7b3db52016-06-16 12:20:52 -07001008 udp_tunnel_notify_del_rx_port(vs->sock,
Alexander Duyckb9adcd692016-06-16 12:23:19 -07001009 (vs->flags & VXLAN_F_GPE) ?
1010 UDP_TUNNEL_TYPE_VXLAN_GPE :
Alexander Duycke7b3db52016-06-16 12:20:52 -07001011 UDP_TUNNEL_TYPE_VXLAN);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07001012 spin_unlock(&vn->sock_lock);
1013
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001014 return true;
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001015}
1016
Jiri Bencb1be00a2015-09-24 13:50:02 +02001017static void vxlan_sock_release(struct vxlan_dev *vxlan)
1018{
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001019 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001020#if IS_ENABLED(CONFIG_IPV6)
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001021 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1022
1023 rcu_assign_pointer(vxlan->vn6_sock, NULL);
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001024#endif
1025
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001026 rcu_assign_pointer(vxlan->vn4_sock, NULL);
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001027 synchronize_net();
1028
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001029 if (__vxlan_sock_release_prep(sock4)) {
1030 udp_tunnel_sock_release(sock4->sock);
1031 kfree(sock4);
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001032 }
1033
1034#if IS_ENABLED(CONFIG_IPV6)
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001035 if (__vxlan_sock_release_prep(sock6)) {
1036 udp_tunnel_sock_release(sock6->sock);
1037 kfree(sock6);
Hannes Frederic Sowa544a7732016-04-09 12:46:23 +02001038 }
Jiri Bencb1be00a2015-09-24 13:50:02 +02001039#endif
1040}
1041
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001042/* Update multicast group membership when first VNI on
Simon Hormanc4b49512015-04-02 11:17:58 +09001043 * multicast address is brought up
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001044 */
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001045static int vxlan_igmp_join(struct vxlan_dev *vxlan)
Stephen Hemminger7c47ced2013-06-17 14:16:10 -07001046{
Jiri Bencb1be00a2015-09-24 13:50:02 +02001047 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001048 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1049 int ifindex = vxlan->default_dst.remote_ifindex;
Marcelo Ricardo Leitner149d7542015-03-20 10:26:21 -03001050 int ret = -EINVAL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001051
Cong Wange4c7ed42013-08-31 13:44:33 +08001052 if (ip->sa.sa_family == AF_INET) {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001053 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
Cong Wange4c7ed42013-08-31 13:44:33 +08001054 struct ip_mreqn mreq = {
1055 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1056 .imr_ifindex = ifindex,
1057 };
1058
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001059 sk = sock4->sock->sk;
Jiri Bencb1be00a2015-09-24 13:50:02 +02001060 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001061 ret = ip_mc_join_group(sk, &mreq);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001062 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001063#if IS_ENABLED(CONFIG_IPV6)
1064 } else {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001065 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1066
1067 sk = sock6->sock->sk;
Jiri Bencb1be00a2015-09-24 13:50:02 +02001068 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001069 ret = ipv6_stub->ipv6_sock_mc_join(sk, ifindex,
1070 &ip->sin6.sin6_addr);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001071 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001072#endif
1073 }
stephen hemminger3fc2de22013-07-18 08:40:15 -07001074
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001075 return ret;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001076}
1077
1078/* Inverse of vxlan_igmp_join when last VNI is brought down */
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001079static int vxlan_igmp_leave(struct vxlan_dev *vxlan)
stephen hemminger3fc2de22013-07-18 08:40:15 -07001080{
Jiri Bencb1be00a2015-09-24 13:50:02 +02001081 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001082 union vxlan_addr *ip = &vxlan->default_dst.remote_ip;
1083 int ifindex = vxlan->default_dst.remote_ifindex;
Marcelo Ricardo Leitner149d7542015-03-20 10:26:21 -03001084 int ret = -EINVAL;
stephen hemminger3fc2de22013-07-18 08:40:15 -07001085
Cong Wange4c7ed42013-08-31 13:44:33 +08001086 if (ip->sa.sa_family == AF_INET) {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001087 struct vxlan_sock *sock4 = rtnl_dereference(vxlan->vn4_sock);
Cong Wange4c7ed42013-08-31 13:44:33 +08001088 struct ip_mreqn mreq = {
1089 .imr_multiaddr.s_addr = ip->sin.sin_addr.s_addr,
1090 .imr_ifindex = ifindex,
1091 };
1092
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001093 sk = sock4->sock->sk;
Jiri Bencb1be00a2015-09-24 13:50:02 +02001094 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001095 ret = ip_mc_leave_group(sk, &mreq);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001096 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001097#if IS_ENABLED(CONFIG_IPV6)
1098 } else {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001099 struct vxlan_sock *sock6 = rtnl_dereference(vxlan->vn6_sock);
1100
1101 sk = sock6->sock->sk;
Jiri Bencb1be00a2015-09-24 13:50:02 +02001102 lock_sock(sk);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001103 ret = ipv6_stub->ipv6_sock_mc_drop(sk, ifindex,
1104 &ip->sin6.sin6_addr);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001105 release_sock(sk);
Cong Wange4c7ed42013-08-31 13:44:33 +08001106#endif
1107 }
stephen hemmingerd3428942012-10-01 12:32:35 +00001108
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03001109 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00001110}
1111
Jiri Bencf14eceb2016-02-16 21:59:01 +01001112static bool vxlan_remcsum(struct vxlanhdr *unparsed,
1113 struct sk_buff *skb, u32 vxflags)
Tom Herbertdfd86452015-01-12 17:00:38 -08001114{
Jiri Benc7d34fa72016-03-21 17:50:05 +01001115 size_t start, offset;
Tom Herbertdfd86452015-01-12 17:00:38 -08001116
Jiri Bencf14eceb2016-02-16 21:59:01 +01001117 if (!(unparsed->vx_flags & VXLAN_HF_RCO) || skb->remcsum_offload)
1118 goto out;
Tom Herbertb7fe10e2015-08-19 17:07:32 -07001119
Jiri Bencf14eceb2016-02-16 21:59:01 +01001120 start = vxlan_rco_start(unparsed->vx_vni);
1121 offset = start + vxlan_rco_offset(unparsed->vx_vni);
Tom Herbertdfd86452015-01-12 17:00:38 -08001122
Jiri Benc7d34fa72016-03-21 17:50:05 +01001123 if (!pskb_may_pull(skb, offset + sizeof(u16)))
Jiri Bencbe5cfea2016-02-16 21:58:59 +01001124 return false;
Tom Herbertdfd86452015-01-12 17:00:38 -08001125
Jiri Bencbe5cfea2016-02-16 21:58:59 +01001126 skb_remcsum_process(skb, (void *)(vxlan_hdr(skb) + 1), start, offset,
1127 !!(vxflags & VXLAN_F_REMCSUM_NOPARTIAL));
Jiri Bencf14eceb2016-02-16 21:59:01 +01001128out:
1129 unparsed->vx_flags &= ~VXLAN_HF_RCO;
1130 unparsed->vx_vni &= VXLAN_VNI_MASK;
Jiri Bencbe5cfea2016-02-16 21:58:59 +01001131 return true;
Tom Herbertdfd86452015-01-12 17:00:38 -08001132}
1133
Jiri Bencf14eceb2016-02-16 21:59:01 +01001134static void vxlan_parse_gbp_hdr(struct vxlanhdr *unparsed,
Jiri Benc64f87d32016-02-23 18:02:55 +01001135 struct sk_buff *skb, u32 vxflags,
Jiri Benc10a5af22016-02-23 18:02:59 +01001136 struct vxlan_metadata *md)
Jiri Benc3288af02016-02-16 21:59:00 +01001137{
Jiri Bencf14eceb2016-02-16 21:59:01 +01001138 struct vxlanhdr_gbp *gbp = (struct vxlanhdr_gbp *)unparsed;
Jiri Benc10a5af22016-02-23 18:02:59 +01001139 struct metadata_dst *tun_dst;
Jiri Benc3288af02016-02-16 21:59:00 +01001140
Jiri Bencf14eceb2016-02-16 21:59:01 +01001141 if (!(unparsed->vx_flags & VXLAN_HF_GBP))
1142 goto out;
1143
Jiri Benc3288af02016-02-16 21:59:00 +01001144 md->gbp = ntohs(gbp->policy_id);
1145
Jiri Benc10a5af22016-02-23 18:02:59 +01001146 tun_dst = (struct metadata_dst *)skb_dst(skb);
David S. Miller810813c2016-03-08 12:34:12 -05001147 if (tun_dst) {
Jiri Benc3288af02016-02-16 21:59:00 +01001148 tun_dst->u.tun_info.key.tun_flags |= TUNNEL_VXLAN_OPT;
David S. Miller810813c2016-03-08 12:34:12 -05001149 tun_dst->u.tun_info.options_len = sizeof(*md);
1150 }
Jiri Benc3288af02016-02-16 21:59:00 +01001151 if (gbp->dont_learn)
1152 md->gbp |= VXLAN_GBP_DONT_LEARN;
1153
1154 if (gbp->policy_applied)
1155 md->gbp |= VXLAN_GBP_POLICY_APPLIED;
Jiri Bencf14eceb2016-02-16 21:59:01 +01001156
Jiri Benc64f87d32016-02-23 18:02:55 +01001157 /* In flow-based mode, GBP is carried in dst_metadata */
1158 if (!(vxflags & VXLAN_F_COLLECT_METADATA))
1159 skb->mark = md->gbp;
Jiri Bencf14eceb2016-02-16 21:59:01 +01001160out:
1161 unparsed->vx_flags &= ~VXLAN_GBP_USED_BITS;
Jiri Benc3288af02016-02-16 21:59:00 +01001162}
1163
Jiri Bence1e53142016-04-05 14:47:13 +02001164static bool vxlan_parse_gpe_hdr(struct vxlanhdr *unparsed,
Jiri Benc61618ee2016-04-11 17:06:08 +02001165 __be16 *protocol,
Jiri Bence1e53142016-04-05 14:47:13 +02001166 struct sk_buff *skb, u32 vxflags)
1167{
1168 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)unparsed;
1169
1170 /* Need to have Next Protocol set for interfaces in GPE mode. */
1171 if (!gpe->np_applied)
1172 return false;
1173 /* "The initial version is 0. If a receiver does not support the
1174 * version indicated it MUST drop the packet.
1175 */
1176 if (gpe->version != 0)
1177 return false;
1178 /* "When the O bit is set to 1, the packet is an OAM packet and OAM
1179 * processing MUST occur." However, we don't implement OAM
1180 * processing, thus drop the packet.
1181 */
1182 if (gpe->oam_flag)
1183 return false;
1184
1185 switch (gpe->next_protocol) {
1186 case VXLAN_GPE_NP_IPV4:
1187 *protocol = htons(ETH_P_IP);
1188 break;
1189 case VXLAN_GPE_NP_IPV6:
1190 *protocol = htons(ETH_P_IPV6);
1191 break;
1192 case VXLAN_GPE_NP_ETHERNET:
1193 *protocol = htons(ETH_P_TEB);
1194 break;
1195 default:
1196 return false;
1197 }
1198
1199 unparsed->vx_flags &= ~VXLAN_GPE_USED_BITS;
1200 return true;
1201}
1202
Jiri Benc1ab016e2016-02-23 18:02:56 +01001203static bool vxlan_set_mac(struct vxlan_dev *vxlan,
1204 struct vxlan_sock *vs,
1205 struct sk_buff *skb)
Thomas Graf614732e2015-07-21 10:44:06 +02001206{
Thomas Graf614732e2015-07-21 10:44:06 +02001207 union vxlan_addr saddr;
Thomas Graf614732e2015-07-21 10:44:06 +02001208
Thomas Graf614732e2015-07-21 10:44:06 +02001209 skb_reset_mac_header(skb);
Thomas Graf614732e2015-07-21 10:44:06 +02001210 skb->protocol = eth_type_trans(skb, vxlan->dev);
1211 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1212
1213 /* Ignore packet loops (and multicast echo) */
1214 if (ether_addr_equal(eth_hdr(skb)->h_source, vxlan->dev->dev_addr))
Jiri Benc1ab016e2016-02-23 18:02:56 +01001215 return false;
Thomas Graf614732e2015-07-21 10:44:06 +02001216
Jiri Benc760c6802016-02-23 18:02:57 +01001217 /* Get address from the outer IP header */
Jiri Bencce212d02015-12-07 16:29:08 +01001218 if (vxlan_get_sk_family(vs) == AF_INET) {
Jiri Benc1ab016e2016-02-23 18:02:56 +01001219 saddr.sin.sin_addr.s_addr = ip_hdr(skb)->saddr;
Thomas Graf614732e2015-07-21 10:44:06 +02001220 saddr.sa.sa_family = AF_INET;
1221#if IS_ENABLED(CONFIG_IPV6)
1222 } else {
Jiri Benc1ab016e2016-02-23 18:02:56 +01001223 saddr.sin6.sin6_addr = ipv6_hdr(skb)->saddr;
Thomas Graf614732e2015-07-21 10:44:06 +02001224 saddr.sa.sa_family = AF_INET6;
1225#endif
1226 }
1227
Jiri Benc1ab016e2016-02-23 18:02:56 +01001228 if ((vxlan->flags & VXLAN_F_LEARN) &&
1229 vxlan_snoop(skb->dev, &saddr, eth_hdr(skb)->h_source))
1230 return false;
1231
1232 return true;
1233}
1234
Jiri Benc760c6802016-02-23 18:02:57 +01001235static bool vxlan_ecn_decapsulate(struct vxlan_sock *vs, void *oiph,
1236 struct sk_buff *skb)
1237{
1238 int err = 0;
1239
1240 if (vxlan_get_sk_family(vs) == AF_INET)
1241 err = IP_ECN_decapsulate(oiph, skb);
1242#if IS_ENABLED(CONFIG_IPV6)
1243 else
1244 err = IP6_ECN_decapsulate(oiph, skb);
1245#endif
1246
1247 if (unlikely(err) && log_ecn_error) {
1248 if (vxlan_get_sk_family(vs) == AF_INET)
1249 net_info_ratelimited("non-ECT from %pI4 with TOS=%#x\n",
1250 &((struct iphdr *)oiph)->saddr,
1251 ((struct iphdr *)oiph)->tos);
1252 else
1253 net_info_ratelimited("non-ECT from %pI6\n",
1254 &((struct ipv6hdr *)oiph)->saddr);
1255 }
1256 return err <= 1;
1257}
1258
stephen hemmingerd3428942012-10-01 12:32:35 +00001259/* Callback from net/ipv4/udp.c to receive packets */
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001260static int vxlan_rcv(struct sock *sk, struct sk_buff *skb)
stephen hemmingerd3428942012-10-01 12:32:35 +00001261{
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001262 struct pcpu_sw_netstats *stats;
Jiri Bencc9e78ef2016-02-18 11:22:51 +01001263 struct vxlan_dev *vxlan;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001264 struct vxlan_sock *vs;
Jiri Bencf14eceb2016-02-16 21:59:01 +01001265 struct vxlanhdr unparsed;
Thomas Grafee122c72015-07-21 10:43:58 +02001266 struct vxlan_metadata _md;
1267 struct vxlan_metadata *md = &_md;
Jiri Benc61618ee2016-04-11 17:06:08 +02001268 __be16 protocol = htons(ETH_P_TEB);
Jiri Bence1e53142016-04-05 14:47:13 +02001269 bool raw_proto = false;
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001270 void *oiph;
stephen hemmingerd3428942012-10-01 12:32:35 +00001271
Jiri Bence1e53142016-04-05 14:47:13 +02001272 /* Need UDP and VXLAN header to be present */
Pravin B Shelar7ce04752013-08-19 11:22:54 -07001273 if (!pskb_may_pull(skb, VXLAN_HLEN))
Hannes Frederic Sowae5aed002016-05-19 15:58:33 +02001274 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001275
Jiri Bencf14eceb2016-02-16 21:59:01 +01001276 unparsed = *vxlan_hdr(skb);
Jiri Benc288b01c2016-02-16 21:59:02 +01001277 /* VNI flag always required to be set */
1278 if (!(unparsed.vx_flags & VXLAN_HF_VNI)) {
1279 netdev_dbg(skb->dev, "invalid vxlan flags=%#x vni=%#x\n",
1280 ntohl(vxlan_hdr(skb)->vx_flags),
1281 ntohl(vxlan_hdr(skb)->vx_vni));
1282 /* Return non vxlan pkt */
Hannes Frederic Sowae5aed002016-05-19 15:58:33 +02001283 goto drop;
stephen hemmingerd3428942012-10-01 12:32:35 +00001284 }
Jiri Benc288b01c2016-02-16 21:59:02 +01001285 unparsed.vx_flags &= ~VXLAN_HF_VNI;
1286 unparsed.vx_vni &= ~VXLAN_VNI_MASK;
stephen hemmingerd3428942012-10-01 12:32:35 +00001287
Pravin B Shelar559835ea2013-09-24 10:25:40 -07001288 vs = rcu_dereference_sk_user_data(sk);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001289 if (!vs)
stephen hemmingerd3428942012-10-01 12:32:35 +00001290 goto drop;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001291
Jiri Bencc9e78ef2016-02-18 11:22:51 +01001292 vxlan = vxlan_vs_find_vni(vs, vxlan_vni(vxlan_hdr(skb)->vx_vni));
1293 if (!vxlan)
1294 goto drop;
1295
Jiri Bence1e53142016-04-05 14:47:13 +02001296 /* For backwards compatibility, only allow reserved fields to be
1297 * used by VXLAN extensions if explicitly requested.
1298 */
1299 if (vs->flags & VXLAN_F_GPE) {
1300 if (!vxlan_parse_gpe_hdr(&unparsed, &protocol, skb, vs->flags))
1301 goto drop;
1302 raw_proto = true;
1303 }
1304
1305 if (__iptunnel_pull_header(skb, VXLAN_HLEN, protocol, raw_proto,
1306 !net_eq(vxlan->net, dev_net(vxlan->dev))))
1307 goto drop;
Jiri Bencc9e78ef2016-02-18 11:22:51 +01001308
Thomas Grafee122c72015-07-21 10:43:58 +02001309 if (vxlan_collect_metadata(vs)) {
Jiri Benc07dabf22016-02-18 19:19:29 +01001310 __be32 vni = vxlan_vni(vxlan_hdr(skb)->vx_vni);
Jiri Benc10a5af22016-02-23 18:02:59 +01001311 struct metadata_dst *tun_dst;
Jiri Benc07dabf22016-02-18 19:19:29 +01001312
Pravin B Shelarc29a70d2015-08-26 23:46:50 -07001313 tun_dst = udp_tun_rx_dst(skb, vxlan_get_sk_family(vs), TUNNEL_KEY,
Amir Vadaid817f432016-09-08 16:23:45 +03001314 key32_to_tunnel_id(vni), sizeof(*md));
Pravin B Shelarc29a70d2015-08-26 23:46:50 -07001315
Thomas Grafee122c72015-07-21 10:43:58 +02001316 if (!tun_dst)
1317 goto drop;
1318
Geert Uytterhoeven0f1b7352015-09-04 12:49:32 +02001319 md = ip_tunnel_info_opts(&tun_dst->u.tun_info);
Jiri Benc10a5af22016-02-23 18:02:59 +01001320
1321 skb_dst_set(skb, (struct dst_entry *)tun_dst);
Thomas Grafee122c72015-07-21 10:43:58 +02001322 } else {
1323 memset(md, 0, sizeof(*md));
1324 }
1325
Jiri Bencf14eceb2016-02-16 21:59:01 +01001326 if (vs->flags & VXLAN_F_REMCSUM_RX)
1327 if (!vxlan_remcsum(&unparsed, skb, vs->flags))
1328 goto drop;
1329 if (vs->flags & VXLAN_F_GBP)
Jiri Benc10a5af22016-02-23 18:02:59 +01001330 vxlan_parse_gbp_hdr(&unparsed, skb, vs->flags, md);
Jiri Bence1e53142016-04-05 14:47:13 +02001331 /* Note that GBP and GPE can never be active together. This is
1332 * ensured in vxlan_dev_configure.
1333 */
Thomas Graf35114942015-01-15 03:53:55 +01001334
Jiri Bencf14eceb2016-02-16 21:59:01 +01001335 if (unparsed.vx_flags || unparsed.vx_vni) {
Tom Herbert3bf39472015-01-08 12:31:18 -08001336 /* If there are any unprocessed flags remaining treat
1337 * this as a malformed packet. This behavior diverges from
1338 * VXLAN RFC (RFC7348) which stipulates that bits in reserved
1339 * in reserved fields are to be ignored. The approach here
Simon Hormanc4b49512015-04-02 11:17:58 +09001340 * maintains compatibility with previous stack code, and also
Tom Herbert3bf39472015-01-08 12:31:18 -08001341 * is more robust and provides a little more security in
1342 * adding extensions to VXLAN.
1343 */
Jiri Benc288b01c2016-02-16 21:59:02 +01001344 goto drop;
Tom Herbert3bf39472015-01-08 12:31:18 -08001345 }
1346
Jiri Bence1e53142016-04-05 14:47:13 +02001347 if (!raw_proto) {
1348 if (!vxlan_set_mac(vxlan, vs, skb))
1349 goto drop;
1350 } else {
Jiri Benc8be0cfa2016-05-13 10:48:42 +02001351 skb_reset_mac_header(skb);
Jiri Bence1e53142016-04-05 14:47:13 +02001352 skb->dev = vxlan->dev;
1353 skb->pkt_type = PACKET_HOST;
1354 }
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001355
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01001356 oiph = skb_network_header(skb);
1357 skb_reset_network_header(skb);
1358
1359 if (!vxlan_ecn_decapsulate(vs, oiph, skb)) {
1360 ++vxlan->dev->stats.rx_frame_errors;
1361 ++vxlan->dev->stats.rx_errors;
1362 goto drop;
1363 }
1364
1365 stats = this_cpu_ptr(vxlan->dev->tstats);
1366 u64_stats_update_begin(&stats->syncp);
1367 stats->rx_packets++;
1368 stats->rx_bytes += skb->len;
1369 u64_stats_update_end(&stats->syncp);
1370
1371 gro_cells_receive(&vxlan->gro_cells, skb);
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001372 return 0;
1373
1374drop:
Jiri Benc288b01c2016-02-16 21:59:02 +01001375 /* Consume bad packet */
1376 kfree_skb(skb);
1377 return 0;
Pravin B Shelar5cfccc52013-08-19 11:23:02 -07001378}
1379
David Stevense4f67ad2012-11-20 02:50:14 +00001380static int arp_reduce(struct net_device *dev, struct sk_buff *skb)
1381{
1382 struct vxlan_dev *vxlan = netdev_priv(dev);
1383 struct arphdr *parp;
1384 u8 *arpptr, *sha;
1385 __be32 sip, tip;
1386 struct neighbour *n;
1387
1388 if (dev->flags & IFF_NOARP)
1389 goto out;
1390
1391 if (!pskb_may_pull(skb, arp_hdr_len(dev))) {
1392 dev->stats.tx_dropped++;
1393 goto out;
1394 }
1395 parp = arp_hdr(skb);
1396
1397 if ((parp->ar_hrd != htons(ARPHRD_ETHER) &&
1398 parp->ar_hrd != htons(ARPHRD_IEEE802)) ||
1399 parp->ar_pro != htons(ETH_P_IP) ||
1400 parp->ar_op != htons(ARPOP_REQUEST) ||
1401 parp->ar_hln != dev->addr_len ||
1402 parp->ar_pln != 4)
1403 goto out;
1404 arpptr = (u8 *)parp + sizeof(struct arphdr);
1405 sha = arpptr;
1406 arpptr += dev->addr_len; /* sha */
1407 memcpy(&sip, arpptr, sizeof(sip));
1408 arpptr += sizeof(sip);
1409 arpptr += dev->addr_len; /* tha */
1410 memcpy(&tip, arpptr, sizeof(tip));
1411
1412 if (ipv4_is_loopback(tip) ||
1413 ipv4_is_multicast(tip))
1414 goto out;
1415
1416 n = neigh_lookup(&arp_tbl, &tip, dev);
1417
1418 if (n) {
David Stevense4f67ad2012-11-20 02:50:14 +00001419 struct vxlan_fdb *f;
1420 struct sk_buff *reply;
1421
1422 if (!(n->nud_state & NUD_CONNECTED)) {
1423 neigh_release(n);
1424 goto out;
1425 }
1426
1427 f = vxlan_find_mac(vxlan, n->ha);
Cong Wange4c7ed42013-08-31 13:44:33 +08001428 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
David Stevense4f67ad2012-11-20 02:50:14 +00001429 /* bridge-local neighbor */
1430 neigh_release(n);
1431 goto out;
1432 }
1433
1434 reply = arp_create(ARPOP_REPLY, ETH_P_ARP, sip, dev, tip, sha,
1435 n->ha, sha);
1436
1437 neigh_release(n);
1438
David Stevens73461352014-03-18 12:32:29 -04001439 if (reply == NULL)
1440 goto out;
1441
David Stevense4f67ad2012-11-20 02:50:14 +00001442 skb_reset_mac_header(reply);
1443 __skb_pull(reply, skb_network_offset(reply));
1444 reply->ip_summed = CHECKSUM_UNNECESSARY;
1445 reply->pkt_type = PACKET_HOST;
1446
1447 if (netif_rx_ni(reply) == NET_RX_DROP)
1448 dev->stats.rx_dropped++;
Cong Wange4c7ed42013-08-31 13:44:33 +08001449 } else if (vxlan->flags & VXLAN_F_L3MISS) {
1450 union vxlan_addr ipa = {
1451 .sin.sin_addr.s_addr = tip,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001452 .sin.sin_family = AF_INET,
Cong Wange4c7ed42013-08-31 13:44:33 +08001453 };
1454
1455 vxlan_ip_miss(dev, &ipa);
1456 }
David Stevense4f67ad2012-11-20 02:50:14 +00001457out:
1458 consume_skb(skb);
1459 return NETDEV_TX_OK;
1460}
1461
Cong Wangf564f452013-08-31 13:44:36 +08001462#if IS_ENABLED(CONFIG_IPV6)
David Stevens4b29dba2014-03-24 10:39:58 -04001463static struct sk_buff *vxlan_na_create(struct sk_buff *request,
1464 struct neighbour *n, bool isrouter)
1465{
1466 struct net_device *dev = request->dev;
1467 struct sk_buff *reply;
1468 struct nd_msg *ns, *na;
1469 struct ipv6hdr *pip6;
1470 u8 *daddr;
1471 int na_olen = 8; /* opt hdr + ETH_ALEN for target */
1472 int ns_olen;
1473 int i, len;
1474
1475 if (dev == NULL)
1476 return NULL;
1477
1478 len = LL_RESERVED_SPACE(dev) + sizeof(struct ipv6hdr) +
1479 sizeof(*na) + na_olen + dev->needed_tailroom;
1480 reply = alloc_skb(len, GFP_ATOMIC);
1481 if (reply == NULL)
1482 return NULL;
1483
1484 reply->protocol = htons(ETH_P_IPV6);
1485 reply->dev = dev;
1486 skb_reserve(reply, LL_RESERVED_SPACE(request->dev));
1487 skb_push(reply, sizeof(struct ethhdr));
Zhang Shengju6297b912016-03-03 01:16:54 +00001488 skb_reset_mac_header(reply);
David Stevens4b29dba2014-03-24 10:39:58 -04001489
1490 ns = (struct nd_msg *)skb_transport_header(request);
1491
1492 daddr = eth_hdr(request)->h_source;
1493 ns_olen = request->len - skb_transport_offset(request) - sizeof(*ns);
1494 for (i = 0; i < ns_olen-1; i += (ns->opt[i+1]<<3)) {
1495 if (ns->opt[i] == ND_OPT_SOURCE_LL_ADDR) {
1496 daddr = ns->opt + i + sizeof(struct nd_opt_hdr);
1497 break;
1498 }
1499 }
1500
1501 /* Ethernet header */
1502 ether_addr_copy(eth_hdr(reply)->h_dest, daddr);
1503 ether_addr_copy(eth_hdr(reply)->h_source, n->ha);
1504 eth_hdr(reply)->h_proto = htons(ETH_P_IPV6);
1505 reply->protocol = htons(ETH_P_IPV6);
1506
1507 skb_pull(reply, sizeof(struct ethhdr));
Zhang Shengju6297b912016-03-03 01:16:54 +00001508 skb_reset_network_header(reply);
David Stevens4b29dba2014-03-24 10:39:58 -04001509 skb_put(reply, sizeof(struct ipv6hdr));
1510
1511 /* IPv6 header */
1512
1513 pip6 = ipv6_hdr(reply);
1514 memset(pip6, 0, sizeof(struct ipv6hdr));
1515 pip6->version = 6;
1516 pip6->priority = ipv6_hdr(request)->priority;
1517 pip6->nexthdr = IPPROTO_ICMPV6;
1518 pip6->hop_limit = 255;
1519 pip6->daddr = ipv6_hdr(request)->saddr;
1520 pip6->saddr = *(struct in6_addr *)n->primary_key;
1521
1522 skb_pull(reply, sizeof(struct ipv6hdr));
Zhang Shengju6297b912016-03-03 01:16:54 +00001523 skb_reset_transport_header(reply);
David Stevens4b29dba2014-03-24 10:39:58 -04001524
1525 na = (struct nd_msg *)skb_put(reply, sizeof(*na) + na_olen);
1526
1527 /* Neighbor Advertisement */
1528 memset(na, 0, sizeof(*na)+na_olen);
1529 na->icmph.icmp6_type = NDISC_NEIGHBOUR_ADVERTISEMENT;
1530 na->icmph.icmp6_router = isrouter;
1531 na->icmph.icmp6_override = 1;
1532 na->icmph.icmp6_solicited = 1;
1533 na->target = ns->target;
1534 ether_addr_copy(&na->opt[2], n->ha);
1535 na->opt[0] = ND_OPT_TARGET_LL_ADDR;
1536 na->opt[1] = na_olen >> 3;
1537
1538 na->icmph.icmp6_cksum = csum_ipv6_magic(&pip6->saddr,
1539 &pip6->daddr, sizeof(*na)+na_olen, IPPROTO_ICMPV6,
1540 csum_partial(na, sizeof(*na)+na_olen, 0));
1541
1542 pip6->payload_len = htons(sizeof(*na)+na_olen);
1543
1544 skb_push(reply, sizeof(struct ipv6hdr));
1545
1546 reply->ip_summed = CHECKSUM_UNNECESSARY;
1547
1548 return reply;
1549}
1550
Cong Wangf564f452013-08-31 13:44:36 +08001551static int neigh_reduce(struct net_device *dev, struct sk_buff *skb)
1552{
1553 struct vxlan_dev *vxlan = netdev_priv(dev);
David Stevens4b29dba2014-03-24 10:39:58 -04001554 struct nd_msg *msg;
Cong Wangf564f452013-08-31 13:44:36 +08001555 const struct ipv6hdr *iphdr;
1556 const struct in6_addr *saddr, *daddr;
David Stevens4b29dba2014-03-24 10:39:58 -04001557 struct neighbour *n;
1558 struct inet6_dev *in6_dev;
Cong Wangf564f452013-08-31 13:44:36 +08001559
1560 in6_dev = __in6_dev_get(dev);
1561 if (!in6_dev)
1562 goto out;
1563
Cong Wangf564f452013-08-31 13:44:36 +08001564 iphdr = ipv6_hdr(skb);
1565 saddr = &iphdr->saddr;
1566 daddr = &iphdr->daddr;
1567
Cong Wangf564f452013-08-31 13:44:36 +08001568 msg = (struct nd_msg *)skb_transport_header(skb);
1569 if (msg->icmph.icmp6_code != 0 ||
1570 msg->icmph.icmp6_type != NDISC_NEIGHBOUR_SOLICITATION)
1571 goto out;
1572
David Stevens4b29dba2014-03-24 10:39:58 -04001573 if (ipv6_addr_loopback(daddr) ||
1574 ipv6_addr_is_multicast(&msg->target))
1575 goto out;
1576
1577 n = neigh_lookup(ipv6_stub->nd_tbl, &msg->target, dev);
Cong Wangf564f452013-08-31 13:44:36 +08001578
1579 if (n) {
1580 struct vxlan_fdb *f;
David Stevens4b29dba2014-03-24 10:39:58 -04001581 struct sk_buff *reply;
Cong Wangf564f452013-08-31 13:44:36 +08001582
1583 if (!(n->nud_state & NUD_CONNECTED)) {
1584 neigh_release(n);
1585 goto out;
1586 }
1587
1588 f = vxlan_find_mac(vxlan, n->ha);
1589 if (f && vxlan_addr_any(&(first_remote_rcu(f)->remote_ip))) {
1590 /* bridge-local neighbor */
1591 neigh_release(n);
1592 goto out;
1593 }
1594
David Stevens4b29dba2014-03-24 10:39:58 -04001595 reply = vxlan_na_create(skb, n,
1596 !!(f ? f->flags & NTF_ROUTER : 0));
1597
Cong Wangf564f452013-08-31 13:44:36 +08001598 neigh_release(n);
David Stevens4b29dba2014-03-24 10:39:58 -04001599
1600 if (reply == NULL)
1601 goto out;
1602
1603 if (netif_rx_ni(reply) == NET_RX_DROP)
1604 dev->stats.rx_dropped++;
1605
Cong Wangf564f452013-08-31 13:44:36 +08001606 } else if (vxlan->flags & VXLAN_F_L3MISS) {
David Stevens4b29dba2014-03-24 10:39:58 -04001607 union vxlan_addr ipa = {
1608 .sin6.sin6_addr = msg->target,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001609 .sin6.sin6_family = AF_INET6,
David Stevens4b29dba2014-03-24 10:39:58 -04001610 };
1611
Cong Wangf564f452013-08-31 13:44:36 +08001612 vxlan_ip_miss(dev, &ipa);
1613 }
1614
1615out:
1616 consume_skb(skb);
1617 return NETDEV_TX_OK;
1618}
1619#endif
1620
David Stevense4f67ad2012-11-20 02:50:14 +00001621static bool route_shortcircuit(struct net_device *dev, struct sk_buff *skb)
1622{
1623 struct vxlan_dev *vxlan = netdev_priv(dev);
1624 struct neighbour *n;
David Stevense4f67ad2012-11-20 02:50:14 +00001625
1626 if (is_multicast_ether_addr(eth_hdr(skb)->h_dest))
1627 return false;
1628
1629 n = NULL;
1630 switch (ntohs(eth_hdr(skb)->h_proto)) {
1631 case ETH_P_IP:
Cong Wange15a00a2013-08-31 13:44:34 +08001632 {
1633 struct iphdr *pip;
1634
David Stevense4f67ad2012-11-20 02:50:14 +00001635 if (!pskb_may_pull(skb, sizeof(struct iphdr)))
1636 return false;
1637 pip = ip_hdr(skb);
1638 n = neigh_lookup(&arp_tbl, &pip->daddr, dev);
Cong Wange4c7ed42013-08-31 13:44:33 +08001639 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1640 union vxlan_addr ipa = {
1641 .sin.sin_addr.s_addr = pip->daddr,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001642 .sin.sin_family = AF_INET,
Cong Wange4c7ed42013-08-31 13:44:33 +08001643 };
1644
1645 vxlan_ip_miss(dev, &ipa);
1646 return false;
1647 }
1648
David Stevense4f67ad2012-11-20 02:50:14 +00001649 break;
Cong Wange15a00a2013-08-31 13:44:34 +08001650 }
1651#if IS_ENABLED(CONFIG_IPV6)
1652 case ETH_P_IPV6:
1653 {
1654 struct ipv6hdr *pip6;
1655
1656 if (!pskb_may_pull(skb, sizeof(struct ipv6hdr)))
1657 return false;
1658 pip6 = ipv6_hdr(skb);
1659 n = neigh_lookup(ipv6_stub->nd_tbl, &pip6->daddr, dev);
1660 if (!n && (vxlan->flags & VXLAN_F_L3MISS)) {
1661 union vxlan_addr ipa = {
1662 .sin6.sin6_addr = pip6->daddr,
Gerhard Stenzela45e92a2014-08-22 21:34:16 +02001663 .sin6.sin6_family = AF_INET6,
Cong Wange15a00a2013-08-31 13:44:34 +08001664 };
1665
1666 vxlan_ip_miss(dev, &ipa);
1667 return false;
1668 }
1669
1670 break;
1671 }
1672#endif
David Stevense4f67ad2012-11-20 02:50:14 +00001673 default:
1674 return false;
1675 }
1676
1677 if (n) {
1678 bool diff;
1679
Joe Perches7367d0b2013-09-01 11:51:23 -07001680 diff = !ether_addr_equal(eth_hdr(skb)->h_dest, n->ha);
David Stevense4f67ad2012-11-20 02:50:14 +00001681 if (diff) {
1682 memcpy(eth_hdr(skb)->h_source, eth_hdr(skb)->h_dest,
1683 dev->addr_len);
1684 memcpy(eth_hdr(skb)->h_dest, n->ha, dev->addr_len);
1685 }
1686 neigh_release(n);
1687 return diff;
Cong Wange4c7ed42013-08-31 13:44:33 +08001688 }
1689
David Stevense4f67ad2012-11-20 02:50:14 +00001690 return false;
1691}
1692
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001693static void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, u32 vxflags,
Thomas Graf35114942015-01-15 03:53:55 +01001694 struct vxlan_metadata *md)
1695{
1696 struct vxlanhdr_gbp *gbp;
1697
Thomas Grafdb79a622015-02-04 17:00:04 +01001698 if (!md->gbp)
1699 return;
1700
Thomas Graf35114942015-01-15 03:53:55 +01001701 gbp = (struct vxlanhdr_gbp *)vxh;
Jiri Benc54bfd872016-02-16 21:58:58 +01001702 vxh->vx_flags |= VXLAN_HF_GBP;
Thomas Graf35114942015-01-15 03:53:55 +01001703
1704 if (md->gbp & VXLAN_GBP_DONT_LEARN)
1705 gbp->dont_learn = 1;
1706
1707 if (md->gbp & VXLAN_GBP_POLICY_APPLIED)
1708 gbp->policy_applied = 1;
1709
1710 gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK);
1711}
1712
Jiri Bence1e53142016-04-05 14:47:13 +02001713static int vxlan_build_gpe_hdr(struct vxlanhdr *vxh, u32 vxflags,
1714 __be16 protocol)
1715{
1716 struct vxlanhdr_gpe *gpe = (struct vxlanhdr_gpe *)vxh;
1717
1718 gpe->np_applied = 1;
1719
1720 switch (protocol) {
1721 case htons(ETH_P_IP):
1722 gpe->next_protocol = VXLAN_GPE_NP_IPV4;
1723 return 0;
1724 case htons(ETH_P_IPV6):
1725 gpe->next_protocol = VXLAN_GPE_NP_IPV6;
1726 return 0;
1727 case htons(ETH_P_TEB):
1728 gpe->next_protocol = VXLAN_GPE_NP_ETHERNET;
1729 return 0;
1730 }
1731 return -EPFNOSUPPORT;
1732}
1733
Jiri Bencf491e562016-02-02 18:09:16 +01001734static int vxlan_build_skb(struct sk_buff *skb, struct dst_entry *dst,
1735 int iphdr_len, __be32 vni,
1736 struct vxlan_metadata *md, u32 vxflags,
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01001737 bool udp_sum)
Cong Wange4c7ed42013-08-31 13:44:33 +08001738{
Cong Wange4c7ed42013-08-31 13:44:33 +08001739 struct vxlanhdr *vxh;
Cong Wange4c7ed42013-08-31 13:44:33 +08001740 int min_headroom;
1741 int err;
Tom Herbertdfd86452015-01-12 17:00:38 -08001742 int type = udp_sum ? SKB_GSO_UDP_TUNNEL_CSUM : SKB_GSO_UDP_TUNNEL;
Jiri Bence1e53142016-04-05 14:47:13 +02001743 __be16 inner_protocol = htons(ETH_P_TEB);
Cong Wange4c7ed42013-08-31 13:44:33 +08001744
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001745 if ((vxflags & VXLAN_F_REMCSUM_TX) &&
Tom Herbertdfd86452015-01-12 17:00:38 -08001746 skb->ip_summed == CHECKSUM_PARTIAL) {
1747 int csum_start = skb_checksum_start_offset(skb);
1748
1749 if (csum_start <= VXLAN_MAX_REMCSUM_START &&
1750 !(csum_start & VXLAN_RCO_SHIFT_MASK) &&
1751 (skb->csum_offset == offsetof(struct udphdr, check) ||
Edward Creeb5708502016-02-11 20:57:17 +00001752 skb->csum_offset == offsetof(struct tcphdr, check)))
Tom Herbertdfd86452015-01-12 17:00:38 -08001753 type |= SKB_GSO_TUNNEL_REMCSUM;
Tom Herbertdfd86452015-01-12 17:00:38 -08001754 }
1755
Cong Wange4c7ed42013-08-31 13:44:33 +08001756 min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len
Jiri Bencf491e562016-02-02 18:09:16 +01001757 + VXLAN_HLEN + iphdr_len
Jiri Pirkodf8a39d2015-01-13 17:13:44 +01001758 + (skb_vlan_tag_present(skb) ? VLAN_HLEN : 0);
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001759
1760 /* Need space for new headers (invalidates iph ptr) */
1761 err = skb_cow_head(skb, min_headroom);
Jiri Bence1e53142016-04-05 14:47:13 +02001762 if (unlikely(err))
1763 goto out_free;
Pravin B Shelar649c5b82013-08-19 11:23:22 -07001764
Jiri Pirko59682502014-11-19 14:04:59 +01001765 skb = vlan_hwaccel_push_inside(skb);
1766 if (WARN_ON(!skb))
1767 return -ENOMEM;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07001768
Alexander Duyckaed069d2016-04-14 15:33:37 -04001769 err = iptunnel_handle_offloads(skb, type);
1770 if (err)
1771 goto out_free;
Jesse Grossb736a622015-04-09 11:19:14 -07001772
Pravin B Shelar49560532013-08-19 11:23:17 -07001773 vxh = (struct vxlanhdr *) __skb_push(skb, sizeof(*vxh));
Jiri Benc54bfd872016-02-16 21:58:58 +01001774 vxh->vx_flags = VXLAN_HF_VNI;
1775 vxh->vx_vni = vxlan_vni_field(vni);
Pravin B Shelar49560532013-08-19 11:23:17 -07001776
Tom Herbertdfd86452015-01-12 17:00:38 -08001777 if (type & SKB_GSO_TUNNEL_REMCSUM) {
Jiri Benc54bfd872016-02-16 21:58:58 +01001778 unsigned int start;
Tom Herbertdfd86452015-01-12 17:00:38 -08001779
Jiri Benc54bfd872016-02-16 21:58:58 +01001780 start = skb_checksum_start_offset(skb) - sizeof(struct vxlanhdr);
1781 vxh->vx_vni |= vxlan_compute_rco(start, skb->csum_offset);
1782 vxh->vx_flags |= VXLAN_HF_RCO;
Tom Herbertdfd86452015-01-12 17:00:38 -08001783
1784 if (!skb_is_gso(skb)) {
1785 skb->ip_summed = CHECKSUM_NONE;
1786 skb->encapsulation = 0;
1787 }
1788 }
1789
Tom Herbertaf33c1a2015-01-20 11:23:05 -08001790 if (vxflags & VXLAN_F_GBP)
1791 vxlan_build_gbp_hdr(vxh, vxflags, md);
Jiri Bence1e53142016-04-05 14:47:13 +02001792 if (vxflags & VXLAN_F_GPE) {
1793 err = vxlan_build_gpe_hdr(vxh, vxflags, skb->protocol);
1794 if (err < 0)
1795 goto out_free;
1796 inner_protocol = skb->protocol;
1797 }
Thomas Graf35114942015-01-15 03:53:55 +01001798
Jiri Bence1e53142016-04-05 14:47:13 +02001799 skb_set_inner_protocol(skb, inner_protocol);
Pravin B Shelar039f5062015-12-24 14:34:54 -08001800 return 0;
Jiri Bence1e53142016-04-05 14:47:13 +02001801
1802out_free:
1803 kfree_skb(skb);
1804 return err;
Pravin B Shelar49560532013-08-19 11:23:17 -07001805}
Pravin B Shelar49560532013-08-19 11:23:17 -07001806
Jiri Benc1a8496b2016-02-02 18:09:14 +01001807static struct rtable *vxlan_get_route(struct vxlan_dev *vxlan,
1808 struct sk_buff *skb, int oif, u8 tos,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001809 __be32 daddr, __be32 *saddr,
1810 struct dst_cache *dst_cache,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001811 const struct ip_tunnel_info *info)
Jiri Benc1a8496b2016-02-02 18:09:14 +01001812{
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001813 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
Jiri Benc1a8496b2016-02-02 18:09:14 +01001814 struct rtable *rt = NULL;
1815 struct flowi4 fl4;
1816
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001817 if (tos && !info)
1818 use_cache = false;
1819 if (use_cache) {
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001820 rt = dst_cache_get_ip4(dst_cache, saddr);
1821 if (rt)
1822 return rt;
1823 }
1824
Jiri Benc1a8496b2016-02-02 18:09:14 +01001825 memset(&fl4, 0, sizeof(fl4));
1826 fl4.flowi4_oif = oif;
1827 fl4.flowi4_tos = RT_TOS(tos);
1828 fl4.flowi4_mark = skb->mark;
1829 fl4.flowi4_proto = IPPROTO_UDP;
1830 fl4.daddr = daddr;
pravin shelar272d96a2016-08-05 17:45:36 -07001831 fl4.saddr = *saddr;
Jiri Benc1a8496b2016-02-02 18:09:14 +01001832
1833 rt = ip_route_output_key(vxlan->net, &fl4);
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001834 if (!IS_ERR(rt)) {
Jiri Benc1a8496b2016-02-02 18:09:14 +01001835 *saddr = fl4.saddr;
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001836 if (use_cache)
1837 dst_cache_set_ip4(dst_cache, &rt->dst, fl4.saddr);
1838 }
Jiri Benc1a8496b2016-02-02 18:09:14 +01001839 return rt;
1840}
1841
Jiri Bence5d4b292015-12-07 13:04:30 +01001842#if IS_ENABLED(CONFIG_IPV6)
1843static struct dst_entry *vxlan6_get_route(struct vxlan_dev *vxlan,
Daniel Borkmann14006152016-03-04 15:15:08 +01001844 struct sk_buff *skb, int oif, u8 tos,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01001845 __be32 label,
Jiri Bence5d4b292015-12-07 13:04:30 +01001846 const struct in6_addr *daddr,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001847 struct in6_addr *saddr,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001848 struct dst_cache *dst_cache,
1849 const struct ip_tunnel_info *info)
Jiri Bence5d4b292015-12-07 13:04:30 +01001850{
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001851 struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock);
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001852 bool use_cache = ip_tunnel_dst_cache_usable(skb, info);
Jiri Bence5d4b292015-12-07 13:04:30 +01001853 struct dst_entry *ndst;
1854 struct flowi6 fl6;
1855 int err;
1856
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001857 if (!sock6)
1858 return ERR_PTR(-EIO);
1859
Daniel Borkmann14006152016-03-04 15:15:08 +01001860 if (tos && !info)
1861 use_cache = false;
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001862 if (use_cache) {
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001863 ndst = dst_cache_get_ip6(dst_cache, saddr);
1864 if (ndst)
1865 return ndst;
1866 }
1867
Jiri Bence5d4b292015-12-07 13:04:30 +01001868 memset(&fl6, 0, sizeof(fl6));
1869 fl6.flowi6_oif = oif;
1870 fl6.daddr = *daddr;
pravin shelar272d96a2016-08-05 17:45:36 -07001871 fl6.saddr = *saddr;
Daniel Borkmanneaa93bf2016-03-18 18:37:57 +01001872 fl6.flowlabel = ip6_make_flowinfo(RT_TOS(tos), label);
Jiri Bence5d4b292015-12-07 13:04:30 +01001873 fl6.flowi6_mark = skb->mark;
1874 fl6.flowi6_proto = IPPROTO_UDP;
1875
1876 err = ipv6_stub->ipv6_dst_lookup(vxlan->net,
pravin shelarc6fcc4f2016-10-28 09:59:15 -07001877 sock6->sock->sk,
Jiri Bence5d4b292015-12-07 13:04:30 +01001878 &ndst, &fl6);
1879 if (err < 0)
1880 return ERR_PTR(err);
1881
1882 *saddr = fl6.saddr;
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01001883 if (use_cache)
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01001884 dst_cache_set_ip6(dst_cache, ndst, saddr);
Jiri Bence5d4b292015-12-07 13:04:30 +01001885 return ndst;
1886}
1887#endif
1888
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001889/* Bypass encapsulation if the destination is local */
1890static void vxlan_encap_bypass(struct sk_buff *skb, struct vxlan_dev *src_vxlan,
1891 struct vxlan_dev *dst_vxlan)
1892{
Li RongQing8f849852014-01-04 13:57:59 +08001893 struct pcpu_sw_netstats *tx_stats, *rx_stats;
Cong Wange4c7ed42013-08-31 13:44:33 +08001894 union vxlan_addr loopback;
1895 union vxlan_addr *remote_ip = &dst_vxlan->default_dst.remote_ip;
Li RongQingce6502a2014-10-16 08:49:41 +08001896 struct net_device *dev = skb->dev;
1897 int len = skb->len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001898
Li RongQing8f849852014-01-04 13:57:59 +08001899 tx_stats = this_cpu_ptr(src_vxlan->dev->tstats);
1900 rx_stats = this_cpu_ptr(dst_vxlan->dev->tstats);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001901 skb->pkt_type = PACKET_HOST;
1902 skb->encapsulation = 0;
1903 skb->dev = dst_vxlan->dev;
1904 __skb_pull(skb, skb_network_offset(skb));
1905
Cong Wange4c7ed42013-08-31 13:44:33 +08001906 if (remote_ip->sa.sa_family == AF_INET) {
1907 loopback.sin.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1908 loopback.sa.sa_family = AF_INET;
1909#if IS_ENABLED(CONFIG_IPV6)
1910 } else {
1911 loopback.sin6.sin6_addr = in6addr_loopback;
1912 loopback.sa.sa_family = AF_INET6;
1913#endif
1914 }
1915
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001916 if (dst_vxlan->flags & VXLAN_F_LEARN)
Cong Wange4c7ed42013-08-31 13:44:33 +08001917 vxlan_snoop(skb->dev, &loopback, eth_hdr(skb)->h_source);
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001918
1919 u64_stats_update_begin(&tx_stats->syncp);
1920 tx_stats->tx_packets++;
Li RongQingce6502a2014-10-16 08:49:41 +08001921 tx_stats->tx_bytes += len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001922 u64_stats_update_end(&tx_stats->syncp);
1923
1924 if (netif_rx(skb) == NET_RX_SUCCESS) {
1925 u64_stats_update_begin(&rx_stats->syncp);
1926 rx_stats->rx_packets++;
Li RongQingce6502a2014-10-16 08:49:41 +08001927 rx_stats->rx_bytes += len;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001928 u64_stats_update_end(&rx_stats->syncp);
1929 } else {
Li RongQingce6502a2014-10-16 08:49:41 +08001930 dev->stats.rx_dropped++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001931 }
1932}
1933
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001934static void vxlan_xmit_one(struct sk_buff *skb, struct net_device *dev,
1935 struct vxlan_rdst *rdst, bool did_rsc)
stephen hemmingerd3428942012-10-01 12:32:35 +00001936{
Paolo Abenid71785f2016-02-12 15:43:57 +01001937 struct dst_cache *dst_cache;
Thomas Graf3093fbe2015-07-21 10:44:00 +02001938 struct ip_tunnel_info *info;
stephen hemmingerd3428942012-10-01 12:32:35 +00001939 struct vxlan_dev *vxlan = netdev_priv(dev);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001940 struct sock *sk;
Cong Wange4c7ed42013-08-31 13:44:33 +08001941 struct rtable *rt = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +00001942 const struct iphdr *old_iph;
Cong Wange4c7ed42013-08-31 13:44:33 +08001943 union vxlan_addr *dst;
pravin shelar272d96a2016-08-05 17:45:36 -07001944 union vxlan_addr remote_ip, local_ip;
1945 union vxlan_addr *src;
Thomas Grafee122c72015-07-21 10:43:58 +02001946 struct vxlan_metadata _md;
1947 struct vxlan_metadata *md = &_md;
Cong Wange4c7ed42013-08-31 13:44:33 +08001948 __be16 src_port = 0, dst_port;
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01001949 __be32 vni, label;
stephen hemmingerd3428942012-10-01 12:32:35 +00001950 __be16 df = 0;
1951 __u8 tos, ttl;
Pravin B Shelar0e6fbc52013-06-17 17:49:56 -07001952 int err;
Thomas Grafee122c72015-07-21 10:43:58 +02001953 u32 flags = vxlan->flags;
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01001954 bool udp_sum = false;
Jiri Bencf491e562016-02-02 18:09:16 +01001955 bool xnet = !net_eq(vxlan->net, dev_net(vxlan->dev));
stephen hemmingerd3428942012-10-01 12:32:35 +00001956
Jiri Benc61adedf2015-08-20 13:56:25 +02001957 info = skb_tunnel_info(skb);
Thomas Graf3093fbe2015-07-21 10:44:00 +02001958
Thomas Grafee122c72015-07-21 10:43:58 +02001959 if (rdst) {
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001960 dst_port = rdst->remote_port ? rdst->remote_port : vxlan->cfg.dst_port;
Thomas Grafee122c72015-07-21 10:43:58 +02001961 vni = rdst->remote_vni;
1962 dst = &rdst->remote_ip;
pravin shelar272d96a2016-08-05 17:45:36 -07001963 src = &vxlan->cfg.saddr;
Paolo Abenid71785f2016-02-12 15:43:57 +01001964 dst_cache = &rdst->dst_cache;
Thomas Grafee122c72015-07-21 10:43:58 +02001965 } else {
1966 if (!info) {
1967 WARN_ONCE(1, "%s: Missing encapsulation instructions\n",
1968 dev->name);
1969 goto drop;
1970 }
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001971 dst_port = info->key.tp_dst ? : vxlan->cfg.dst_port;
Amir Vadaid817f432016-09-08 16:23:45 +03001972 vni = tunnel_id_to_key32(info->key.tun_id);
Jiri Bencb1be00a2015-09-24 13:50:02 +02001973 remote_ip.sa.sa_family = ip_tunnel_info_af(info);
pravin shelar272d96a2016-08-05 17:45:36 -07001974 if (remote_ip.sa.sa_family == AF_INET) {
Jiri Benca725e512015-08-20 13:56:30 +02001975 remote_ip.sin.sin_addr.s_addr = info->key.u.ipv4.dst;
pravin shelar272d96a2016-08-05 17:45:36 -07001976 local_ip.sin.sin_addr.s_addr = info->key.u.ipv4.src;
1977 } else {
Jiri Benca725e512015-08-20 13:56:30 +02001978 remote_ip.sin6.sin6_addr = info->key.u.ipv6.dst;
pravin shelar272d96a2016-08-05 17:45:36 -07001979 local_ip.sin6.sin6_addr = info->key.u.ipv6.src;
1980 }
Thomas Grafee122c72015-07-21 10:43:58 +02001981 dst = &remote_ip;
pravin shelar272d96a2016-08-05 17:45:36 -07001982 src = &local_ip;
Paolo Abenid71785f2016-02-12 15:43:57 +01001983 dst_cache = &info->dst_cache;
Thomas Grafee122c72015-07-21 10:43:58 +02001984 }
David Stevense4f67ad2012-11-20 02:50:14 +00001985
Cong Wange4c7ed42013-08-31 13:44:33 +08001986 if (vxlan_addr_any(dst)) {
David Stevense4f67ad2012-11-20 02:50:14 +00001987 if (did_rsc) {
David Stevense4f67ad2012-11-20 02:50:14 +00001988 /* short-circuited back to local bridge */
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00001989 vxlan_encap_bypass(skb, vxlan, vxlan);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07001990 return;
David Stevense4f67ad2012-11-20 02:50:14 +00001991 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001992 goto drop;
David Stevense4f67ad2012-11-20 02:50:14 +00001993 }
stephen hemmingeref59feb2012-10-09 20:35:46 +00001994
stephen hemmingerd3428942012-10-01 12:32:35 +00001995 old_iph = ip_hdr(skb);
1996
Thomas Graf0dfbdf42015-07-21 10:44:02 +02001997 ttl = vxlan->cfg.ttl;
Cong Wange4c7ed42013-08-31 13:44:33 +08001998 if (!ttl && vxlan_addr_multicast(dst))
stephen hemmingerd3428942012-10-01 12:32:35 +00001999 ttl = 1;
2000
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002001 tos = vxlan->cfg.tos;
stephen hemmingerd3428942012-10-01 12:32:35 +00002002 if (tos == 1)
Pravin B Shelar206aaaf2013-03-25 14:49:53 +00002003 tos = ip_tunnel_get_dsfield(old_iph, skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00002004
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002005 label = vxlan->cfg.label;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002006 src_port = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
2007 vxlan->cfg.port_max, true);
stephen hemmingerd3428942012-10-01 12:32:35 +00002008
Jiri Benca725e512015-08-20 13:56:30 +02002009 if (info) {
Jiri Benca725e512015-08-20 13:56:30 +02002010 ttl = info->key.ttl;
2011 tos = info->key.tos;
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002012 label = info->key.label;
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01002013 udp_sum = !!(info->key.tun_flags & TUNNEL_CSUM);
Jiri Benca725e512015-08-20 13:56:30 +02002014
2015 if (info->options_len)
Pravin B Shelar4c222792015-08-30 18:09:38 -07002016 md = ip_tunnel_info_opts(info);
Jiri Benca725e512015-08-20 13:56:30 +02002017 } else {
2018 md->gbp = skb->mark;
2019 }
2020
Cong Wange4c7ed42013-08-31 13:44:33 +08002021 if (dst->sa.sa_family == AF_INET) {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002022 struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock);
2023
2024 if (!sock4)
Jiri Bencb1be00a2015-09-24 13:50:02 +02002025 goto drop;
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002026 sk = sock4->sock->sk;
Jiri Bencb1be00a2015-09-24 13:50:02 +02002027
Jiri Benc1a8496b2016-02-02 18:09:14 +01002028 rt = vxlan_get_route(vxlan, skb,
2029 rdst ? rdst->remote_ifindex : 0, tos,
pravin shelar272d96a2016-08-05 17:45:36 -07002030 dst->sin.sin_addr.s_addr,
2031 &src->sin.sin_addr.s_addr,
Paolo Abenid71785f2016-02-12 15:43:57 +01002032 dst_cache, info);
Cong Wange4c7ed42013-08-31 13:44:33 +08002033 if (IS_ERR(rt)) {
2034 netdev_dbg(dev, "no route to %pI4\n",
2035 &dst->sin.sin_addr.s_addr);
2036 dev->stats.tx_carrier_errors++;
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00002037 goto tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08002038 }
2039
2040 if (rt->dst.dev == dev) {
2041 netdev_dbg(dev, "circular route to %pI4\n",
2042 &dst->sin.sin_addr.s_addr);
2043 dev->stats.collisions++;
Fan Dufffc15a2013-12-09 10:33:53 +08002044 goto rt_tx_error;
Cong Wange4c7ed42013-08-31 13:44:33 +08002045 }
2046
2047 /* Bypass encapsulation if the destination is local */
pravin shelarbbec7802016-08-05 17:45:37 -07002048 if (!info && rt->rt_flags & RTCF_LOCAL &&
Cong Wange4c7ed42013-08-31 13:44:33 +08002049 !(rt->rt_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
2050 struct vxlan_dev *dst_vxlan;
2051
2052 ip_rt_put(rt);
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02002053 dst_vxlan = vxlan_find_vni(vxlan->net, vni,
Thomas Grafac5132d2015-01-15 03:53:56 +01002054 dst->sa.sa_family, dst_port,
2055 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08002056 if (!dst_vxlan)
2057 goto tx_error;
2058 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
2059 return;
2060 }
2061
Alexander Duyck6ceb31c2016-02-19 11:26:31 -08002062 if (!info)
2063 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM_TX);
2064 else if (info->key.tun_flags & TUNNEL_DONT_FRAGMENT)
2065 df = htons(IP_DF);
2066
Cong Wange4c7ed42013-08-31 13:44:33 +08002067 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
2068 ttl = ttl ? : ip4_dst_hoplimit(&rt->dst);
Jiri Bencf491e562016-02-02 18:09:16 +01002069 err = vxlan_build_skb(skb, &rt->dst, sizeof(struct iphdr),
Jiri Benc54bfd872016-02-16 21:58:58 +01002070 vni, md, flags, udp_sum);
Jiri Bencf491e562016-02-02 18:09:16 +01002071 if (err < 0)
2072 goto xmit_tx_error;
2073
pravin shelar272d96a2016-08-05 17:45:36 -07002074 udp_tunnel_xmit_skb(rt, sk, skb, src->sin.sin_addr.s_addr,
Jiri Bencf491e562016-02-02 18:09:16 +01002075 dst->sin.sin_addr.s_addr, tos, ttl, df,
2076 src_port, dst_port, xnet, !udp_sum);
Cong Wange4c7ed42013-08-31 13:44:33 +08002077#if IS_ENABLED(CONFIG_IPV6)
2078 } else {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002079 struct vxlan_sock *sock6 = rcu_dereference(vxlan->vn6_sock);
Cong Wange4c7ed42013-08-31 13:44:33 +08002080 struct dst_entry *ndst;
Jiri Benc6f264e42015-08-20 13:56:29 +02002081 u32 rt6i_flags;
Cong Wange4c7ed42013-08-31 13:44:33 +08002082
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002083 if (!sock6)
Jiri Bencb1be00a2015-09-24 13:50:02 +02002084 goto drop;
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002085 sk = sock6->sock->sk;
Jiri Bencb1be00a2015-09-24 13:50:02 +02002086
Jiri Bence5d4b292015-12-07 13:04:30 +01002087 ndst = vxlan6_get_route(vxlan, skb,
Daniel Borkmann14006152016-03-04 15:15:08 +01002088 rdst ? rdst->remote_ifindex : 0, tos,
pravin shelar272d96a2016-08-05 17:45:36 -07002089 label, &dst->sin6.sin6_addr,
2090 &src->sin6.sin6_addr,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01002091 dst_cache, info);
Jiri Bence5d4b292015-12-07 13:04:30 +01002092 if (IS_ERR(ndst)) {
Cong Wange4c7ed42013-08-31 13:44:33 +08002093 netdev_dbg(dev, "no route to %pI6\n",
2094 &dst->sin6.sin6_addr);
2095 dev->stats.tx_carrier_errors++;
2096 goto tx_error;
2097 }
2098
2099 if (ndst->dev == dev) {
2100 netdev_dbg(dev, "circular route to %pI6\n",
2101 &dst->sin6.sin6_addr);
2102 dst_release(ndst);
2103 dev->stats.collisions++;
2104 goto tx_error;
2105 }
2106
2107 /* Bypass encapsulation if the destination is local */
Jiri Benc6f264e42015-08-20 13:56:29 +02002108 rt6i_flags = ((struct rt6_info *)ndst)->rt6i_flags;
pravin shelarbbec7802016-08-05 17:45:37 -07002109 if (!info && rt6i_flags & RTF_LOCAL &&
Jiri Benc6f264e42015-08-20 13:56:29 +02002110 !(rt6i_flags & (RTCF_BROADCAST | RTCF_MULTICAST))) {
Cong Wange4c7ed42013-08-31 13:44:33 +08002111 struct vxlan_dev *dst_vxlan;
2112
2113 dst_release(ndst);
Marcelo Leitner19ca9fc2014-11-13 14:43:08 -02002114 dst_vxlan = vxlan_find_vni(vxlan->net, vni,
Thomas Grafac5132d2015-01-15 03:53:56 +01002115 dst->sa.sa_family, dst_port,
2116 vxlan->flags);
Cong Wange4c7ed42013-08-31 13:44:33 +08002117 if (!dst_vxlan)
2118 goto tx_error;
2119 vxlan_encap_bypass(skb, vxlan, dst_vxlan);
2120 return;
2121 }
2122
Jiri Bencb4ed5ca2016-02-02 18:09:15 +01002123 if (!info)
2124 udp_sum = !(flags & VXLAN_F_UDP_ZERO_CSUM6_TX);
Jesse Gross35e2d112016-01-20 16:22:47 -08002125
Daniel Borkmann14006152016-03-04 15:15:08 +01002126 tos = ip_tunnel_ecn_encap(tos, old_iph, skb);
Cong Wange4c7ed42013-08-31 13:44:33 +08002127 ttl = ttl ? : ip6_dst_hoplimit(ndst);
Jiri Bencf491e562016-02-02 18:09:16 +01002128 skb_scrub_packet(skb, xnet);
2129 err = vxlan_build_skb(skb, ndst, sizeof(struct ipv6hdr),
Jiri Benc54bfd872016-02-16 21:58:58 +01002130 vni, md, flags, udp_sum);
Jiri Bencf491e562016-02-02 18:09:16 +01002131 if (err < 0) {
2132 dst_release(ndst);
Haishuang Yan5e1e61a2016-09-04 18:52:51 +08002133 dev->stats.tx_errors++;
Jiri Bencf491e562016-02-02 18:09:16 +01002134 return;
2135 }
2136 udp_tunnel6_xmit_skb(ndst, sk, skb, dev,
pravin shelar272d96a2016-08-05 17:45:36 -07002137 &src->sin6.sin6_addr,
2138 &dst->sin6.sin6_addr, tos, ttl,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002139 label, src_port, dst_port, !udp_sum);
Cong Wange4c7ed42013-08-31 13:44:33 +08002140#endif
Sridhar Samudrala9dcc71e2013-04-02 12:31:52 +00002141 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002142
Stephen Hemminger4ad16932013-06-17 14:16:11 -07002143 return;
stephen hemmingerd3428942012-10-01 12:32:35 +00002144
2145drop:
2146 dev->stats.tx_dropped++;
2147 goto tx_free;
2148
Jiri Bencf491e562016-02-02 18:09:16 +01002149xmit_tx_error:
2150 /* skb is already freed. */
2151 skb = NULL;
Pravin B Shelar49560532013-08-19 11:23:17 -07002152rt_tx_error:
2153 ip_rt_put(rt);
stephen hemmingerd3428942012-10-01 12:32:35 +00002154tx_error:
2155 dev->stats.tx_errors++;
2156tx_free:
2157 dev_kfree_skb(skb);
stephen hemmingerd3428942012-10-01 12:32:35 +00002158}
2159
David Stevens66817122013-03-15 04:35:51 +00002160/* Transmit local packets over Vxlan
2161 *
2162 * Outer IP header inherits ECN and DF from inner header.
2163 * Outer UDP destination is the VXLAN assigned port.
2164 * source port is based on hash of flow
2165 */
2166static netdev_tx_t vxlan_xmit(struct sk_buff *skb, struct net_device *dev)
2167{
2168 struct vxlan_dev *vxlan = netdev_priv(dev);
Thomas Graf3093fbe2015-07-21 10:44:00 +02002169 const struct ip_tunnel_info *info;
David Stevens66817122013-03-15 04:35:51 +00002170 struct ethhdr *eth;
2171 bool did_rsc = false;
Eric Dumazet8f646c92014-01-06 09:54:31 -08002172 struct vxlan_rdst *rdst, *fdst = NULL;
David Stevens66817122013-03-15 04:35:51 +00002173 struct vxlan_fdb *f;
David Stevens66817122013-03-15 04:35:51 +00002174
Jiri Benc61adedf2015-08-20 13:56:25 +02002175 info = skb_tunnel_info(skb);
Thomas Graf3093fbe2015-07-21 10:44:00 +02002176
David Stevens66817122013-03-15 04:35:51 +00002177 skb_reset_mac_header(skb);
David Stevens66817122013-03-15 04:35:51 +00002178
Jiri Benc47e5d1b2016-04-05 14:47:11 +02002179 if (vxlan->flags & VXLAN_F_COLLECT_METADATA) {
2180 if (info && info->mode & IP_TUNNEL_INFO_TX)
2181 vxlan_xmit_one(skb, dev, NULL, false);
2182 else
2183 kfree_skb(skb);
2184 return NETDEV_TX_OK;
2185 }
2186
2187 if (vxlan->flags & VXLAN_F_PROXY) {
2188 eth = eth_hdr(skb);
Cong Wangf564f452013-08-31 13:44:36 +08002189 if (ntohs(eth->h_proto) == ETH_P_ARP)
2190 return arp_reduce(dev, skb);
2191#if IS_ENABLED(CONFIG_IPV6)
2192 else if (ntohs(eth->h_proto) == ETH_P_IPV6 &&
Li RongQing91269e32014-10-16 09:17:18 +08002193 pskb_may_pull(skb, sizeof(struct ipv6hdr)
2194 + sizeof(struct nd_msg)) &&
Cong Wangf564f452013-08-31 13:44:36 +08002195 ipv6_hdr(skb)->nexthdr == IPPROTO_ICMPV6) {
2196 struct nd_msg *msg;
2197
2198 msg = (struct nd_msg *)skb_transport_header(skb);
2199 if (msg->icmph.icmp6_code == 0 &&
2200 msg->icmph.icmp6_type == NDISC_NEIGHBOUR_SOLICITATION)
2201 return neigh_reduce(dev, skb);
2202 }
2203#endif
2204 }
David Stevens66817122013-03-15 04:35:51 +00002205
Jiri Benc47e5d1b2016-04-05 14:47:11 +02002206 eth = eth_hdr(skb);
David Stevens66817122013-03-15 04:35:51 +00002207 f = vxlan_find_mac(vxlan, eth->h_dest);
David Stevensae884082013-04-19 00:36:26 +00002208 did_rsc = false;
2209
2210 if (f && (f->flags & NTF_ROUTER) && (vxlan->flags & VXLAN_F_RSC) &&
Cong Wange15a00a2013-08-31 13:44:34 +08002211 (ntohs(eth->h_proto) == ETH_P_IP ||
2212 ntohs(eth->h_proto) == ETH_P_IPV6)) {
David Stevensae884082013-04-19 00:36:26 +00002213 did_rsc = route_shortcircuit(dev, skb);
2214 if (did_rsc)
2215 f = vxlan_find_mac(vxlan, eth->h_dest);
2216 }
2217
David Stevens66817122013-03-15 04:35:51 +00002218 if (f == NULL) {
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002219 f = vxlan_find_mac(vxlan, all_zeros_mac);
2220 if (f == NULL) {
2221 if ((vxlan->flags & VXLAN_F_L2MISS) &&
2222 !is_multicast_ether_addr(eth->h_dest))
2223 vxlan_fdb_miss(vxlan, eth->h_dest);
David Stevens66817122013-03-15 04:35:51 +00002224
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002225 dev->stats.tx_dropped++;
Eric Dumazet8f646c92014-01-06 09:54:31 -08002226 kfree_skb(skb);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002227 return NETDEV_TX_OK;
Stephen Hemminger3e61aa82013-06-17 14:16:12 -07002228 }
David Stevens66817122013-03-15 04:35:51 +00002229 }
2230
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002231 list_for_each_entry_rcu(rdst, &f->remotes, list) {
2232 struct sk_buff *skb1;
2233
Eric Dumazet8f646c92014-01-06 09:54:31 -08002234 if (!fdst) {
2235 fdst = rdst;
2236 continue;
2237 }
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002238 skb1 = skb_clone(skb, GFP_ATOMIC);
2239 if (skb1)
2240 vxlan_xmit_one(skb1, dev, rdst, did_rsc);
2241 }
2242
Eric Dumazet8f646c92014-01-06 09:54:31 -08002243 if (fdst)
2244 vxlan_xmit_one(skb, dev, fdst, did_rsc);
2245 else
2246 kfree_skb(skb);
Stephen Hemminger4ad16932013-06-17 14:16:11 -07002247 return NETDEV_TX_OK;
David Stevens66817122013-03-15 04:35:51 +00002248}
2249
stephen hemmingerd3428942012-10-01 12:32:35 +00002250/* Walk the forwarding table and purge stale entries */
2251static void vxlan_cleanup(unsigned long arg)
2252{
2253 struct vxlan_dev *vxlan = (struct vxlan_dev *) arg;
2254 unsigned long next_timer = jiffies + FDB_AGE_INTERVAL;
2255 unsigned int h;
2256
2257 if (!netif_running(vxlan->dev))
2258 return;
2259
stephen hemmingerd3428942012-10-01 12:32:35 +00002260 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2261 struct hlist_node *p, *n;
Sorin Dumitru14e1d0f2015-05-26 10:42:04 +03002262
2263 spin_lock_bh(&vxlan->hash_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002264 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2265 struct vxlan_fdb *f
2266 = container_of(p, struct vxlan_fdb, hlist);
2267 unsigned long timeout;
2268
stephen hemminger3c172862012-10-26 06:24:34 +00002269 if (f->state & NUD_PERMANENT)
stephen hemmingerd3428942012-10-01 12:32:35 +00002270 continue;
2271
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002272 timeout = f->used + vxlan->cfg.age_interval * HZ;
stephen hemmingerd3428942012-10-01 12:32:35 +00002273 if (time_before_eq(timeout, jiffies)) {
2274 netdev_dbg(vxlan->dev,
2275 "garbage collect %pM\n",
2276 f->eth_addr);
2277 f->state = NUD_STALE;
2278 vxlan_fdb_destroy(vxlan, f);
2279 } else if (time_before(timeout, next_timer))
2280 next_timer = timeout;
2281 }
Sorin Dumitru14e1d0f2015-05-26 10:42:04 +03002282 spin_unlock_bh(&vxlan->hash_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00002283 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002284
2285 mod_timer(&vxlan->age_timer, next_timer);
2286}
2287
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002288static void vxlan_vs_add_dev(struct vxlan_sock *vs, struct vxlan_dev *vxlan)
2289{
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002290 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
Jiri Benc54bfd872016-02-16 21:58:58 +01002291 __be32 vni = vxlan->default_dst.remote_vni;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002292
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002293 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002294 hlist_add_head_rcu(&vxlan->hlist, vni_head(vs, vni));
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002295 spin_unlock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002296}
2297
stephen hemmingerd3428942012-10-01 12:32:35 +00002298/* Setup stats when device is created */
2299static int vxlan_init(struct net_device *dev)
2300{
WANG Cong1c213bd2014-02-13 11:46:28 -08002301 dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
Pravin B Shelare8171042013-03-25 14:49:46 +00002302 if (!dev->tstats)
stephen hemmingerd3428942012-10-01 12:32:35 +00002303 return -ENOMEM;
2304
2305 return 0;
2306}
2307
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002308static void vxlan_fdb_delete_default(struct vxlan_dev *vxlan)
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002309{
2310 struct vxlan_fdb *f;
2311
2312 spin_lock_bh(&vxlan->hash_lock);
2313 f = __vxlan_find_mac(vxlan, all_zeros_mac);
2314 if (f)
2315 vxlan_fdb_destroy(vxlan, f);
2316 spin_unlock_bh(&vxlan->hash_lock);
2317}
2318
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002319static void vxlan_uninit(struct net_device *dev)
2320{
2321 struct vxlan_dev *vxlan = netdev_priv(dev);
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002322
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002323 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002324
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002325 free_percpu(dev->tstats);
2326}
2327
stephen hemmingerd3428942012-10-01 12:32:35 +00002328/* Start ageing timer and join group when device is brought up */
2329static int vxlan_open(struct net_device *dev)
2330{
2331 struct vxlan_dev *vxlan = netdev_priv(dev);
Jiri Benc205f3562015-09-24 13:50:01 +02002332 int ret;
Stephen Hemminger1c51a912013-06-17 14:16:11 -07002333
Jiri Benc205f3562015-09-24 13:50:01 +02002334 ret = vxlan_sock_add(vxlan);
2335 if (ret < 0)
2336 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002337
Gao feng79d4a942013-12-10 16:37:32 +08002338 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002339 ret = vxlan_igmp_join(vxlan);
Marcelo Ricardo Leitnerbef00572015-08-25 20:22:35 -03002340 if (ret == -EADDRINUSE)
2341 ret = 0;
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002342 if (ret) {
Jiri Benc205f3562015-09-24 13:50:01 +02002343 vxlan_sock_release(vxlan);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002344 return ret;
2345 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002346 }
2347
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002348 if (vxlan->cfg.age_interval)
stephen hemmingerd3428942012-10-01 12:32:35 +00002349 mod_timer(&vxlan->age_timer, jiffies + FDB_AGE_INTERVAL);
2350
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002351 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002352}
2353
2354/* Purge the forwarding table */
2355static void vxlan_flush(struct vxlan_dev *vxlan)
2356{
Cong Wang31fec5a2013-05-27 22:35:52 +00002357 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002358
2359 spin_lock_bh(&vxlan->hash_lock);
2360 for (h = 0; h < FDB_HASH_SIZE; ++h) {
2361 struct hlist_node *p, *n;
2362 hlist_for_each_safe(p, n, &vxlan->fdb_head[h]) {
2363 struct vxlan_fdb *f
2364 = container_of(p, struct vxlan_fdb, hlist);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002365 /* the all_zeros_mac entry is deleted at vxlan_uninit */
2366 if (!is_zero_ether_addr(f->eth_addr))
2367 vxlan_fdb_destroy(vxlan, f);
stephen hemmingerd3428942012-10-01 12:32:35 +00002368 }
2369 }
2370 spin_unlock_bh(&vxlan->hash_lock);
2371}
2372
2373/* Cleanup timer and forwarding table on shutdown */
2374static int vxlan_stop(struct net_device *dev)
2375{
2376 struct vxlan_dev *vxlan = netdev_priv(dev);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002377 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002378 int ret = 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002379
Marcelo Ricardo Leitner24c0e682015-03-23 16:23:12 -03002380 if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip) &&
WANG Congf13b1682015-04-08 14:48:30 -07002381 !vxlan_group_used(vn, vxlan))
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002382 ret = vxlan_igmp_leave(vxlan);
stephen hemmingerd3428942012-10-01 12:32:35 +00002383
2384 del_timer_sync(&vxlan->age_timer);
2385
2386 vxlan_flush(vxlan);
Jiri Benc205f3562015-09-24 13:50:01 +02002387 vxlan_sock_release(vxlan);
stephen hemmingerd3428942012-10-01 12:32:35 +00002388
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002389 return ret;
stephen hemmingerd3428942012-10-01 12:32:35 +00002390}
2391
stephen hemmingerd3428942012-10-01 12:32:35 +00002392/* Stub, nothing needs to be done. */
2393static void vxlan_set_multicast_list(struct net_device *dev)
2394{
2395}
2396
David Wragg72564b52016-02-10 00:05:55 +00002397static int __vxlan_change_mtu(struct net_device *dev,
2398 struct net_device *lowerdev,
2399 struct vxlan_rdst *dst, int new_mtu, bool strict)
2400{
2401 int max_mtu = IP_MAX_MTU;
2402
2403 if (lowerdev)
2404 max_mtu = lowerdev->mtu;
2405
2406 if (dst->remote_ip.sa.sa_family == AF_INET6)
2407 max_mtu -= VXLAN6_HEADROOM;
2408 else
2409 max_mtu -= VXLAN_HEADROOM;
2410
2411 if (new_mtu < 68)
2412 return -EINVAL;
2413
2414 if (new_mtu > max_mtu) {
2415 if (strict)
2416 return -EINVAL;
2417
2418 new_mtu = max_mtu;
2419 }
2420
2421 dev->mtu = new_mtu;
2422 return 0;
2423}
2424
Daniel Borkmann345010b2013-12-18 00:21:08 +01002425static int vxlan_change_mtu(struct net_device *dev, int new_mtu)
2426{
2427 struct vxlan_dev *vxlan = netdev_priv(dev);
2428 struct vxlan_rdst *dst = &vxlan->default_dst;
David Wragg72564b52016-02-10 00:05:55 +00002429 struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
2430 dst->remote_ifindex);
2431 return __vxlan_change_mtu(dev, lowerdev, dst, new_mtu, true);
Daniel Borkmann345010b2013-12-18 00:21:08 +01002432}
2433
Pravin B Shelarfc4099f2015-10-22 18:17:16 -07002434static int vxlan_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
2435{
2436 struct vxlan_dev *vxlan = netdev_priv(dev);
2437 struct ip_tunnel_info *info = skb_tunnel_info(skb);
2438 __be16 sport, dport;
2439
2440 sport = udp_flow_src_port(dev_net(dev), skb, vxlan->cfg.port_min,
2441 vxlan->cfg.port_max, true);
2442 dport = info->key.tp_dst ? : vxlan->cfg.dst_port;
2443
Jiri Benc239e9442015-12-07 13:04:31 +01002444 if (ip_tunnel_info_af(info) == AF_INET) {
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002445 struct vxlan_sock *sock4 = rcu_dereference(vxlan->vn4_sock);
Jiri Benc1a8496b2016-02-02 18:09:14 +01002446 struct rtable *rt;
2447
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002448 if (!sock4)
Jiri Benc239e9442015-12-07 13:04:31 +01002449 return -EINVAL;
Jiri Benc1a8496b2016-02-02 18:09:14 +01002450 rt = vxlan_get_route(vxlan, skb, 0, info->key.tos,
2451 info->key.u.ipv4.dst,
Paolo Abeni0c1d70a2016-02-12 15:43:56 +01002452 &info->key.u.ipv4.src, NULL, info);
Jiri Benc1a8496b2016-02-02 18:09:14 +01002453 if (IS_ERR(rt))
2454 return PTR_ERR(rt);
2455 ip_rt_put(rt);
Jiri Benc239e9442015-12-07 13:04:31 +01002456 } else {
2457#if IS_ENABLED(CONFIG_IPV6)
2458 struct dst_entry *ndst;
2459
Daniel Borkmann14006152016-03-04 15:15:08 +01002460 ndst = vxlan6_get_route(vxlan, skb, 0, info->key.tos,
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002461 info->key.label, &info->key.u.ipv6.dst,
Daniel Borkmanndb3c6132016-03-04 15:15:07 +01002462 &info->key.u.ipv6.src, NULL, info);
Jiri Benc239e9442015-12-07 13:04:31 +01002463 if (IS_ERR(ndst))
2464 return PTR_ERR(ndst);
2465 dst_release(ndst);
Jiri Benc239e9442015-12-07 13:04:31 +01002466#else /* !CONFIG_IPV6 */
2467 return -EPFNOSUPPORT;
2468#endif
2469 }
Jiri Benc1a8496b2016-02-02 18:09:14 +01002470 info->key.tp_src = sport;
2471 info->key.tp_dst = dport;
Jiri Benc239e9442015-12-07 13:04:31 +01002472 return 0;
Pravin B Shelarfc4099f2015-10-22 18:17:16 -07002473}
2474
Jiri Benc0c867c92016-04-05 14:47:10 +02002475static const struct net_device_ops vxlan_netdev_ether_ops = {
stephen hemmingerd3428942012-10-01 12:32:35 +00002476 .ndo_init = vxlan_init,
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002477 .ndo_uninit = vxlan_uninit,
stephen hemmingerd3428942012-10-01 12:32:35 +00002478 .ndo_open = vxlan_open,
2479 .ndo_stop = vxlan_stop,
2480 .ndo_start_xmit = vxlan_xmit,
Pravin B Shelare8171042013-03-25 14:49:46 +00002481 .ndo_get_stats64 = ip_tunnel_get_stats64,
stephen hemmingerd3428942012-10-01 12:32:35 +00002482 .ndo_set_rx_mode = vxlan_set_multicast_list,
Daniel Borkmann345010b2013-12-18 00:21:08 +01002483 .ndo_change_mtu = vxlan_change_mtu,
stephen hemmingerd3428942012-10-01 12:32:35 +00002484 .ndo_validate_addr = eth_validate_addr,
2485 .ndo_set_mac_address = eth_mac_addr,
2486 .ndo_fdb_add = vxlan_fdb_add,
2487 .ndo_fdb_del = vxlan_fdb_delete,
2488 .ndo_fdb_dump = vxlan_fdb_dump,
Pravin B Shelarfc4099f2015-10-22 18:17:16 -07002489 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
stephen hemmingerd3428942012-10-01 12:32:35 +00002490};
2491
Jiri Bence1e53142016-04-05 14:47:13 +02002492static const struct net_device_ops vxlan_netdev_raw_ops = {
2493 .ndo_init = vxlan_init,
2494 .ndo_uninit = vxlan_uninit,
2495 .ndo_open = vxlan_open,
2496 .ndo_stop = vxlan_stop,
2497 .ndo_start_xmit = vxlan_xmit,
2498 .ndo_get_stats64 = ip_tunnel_get_stats64,
2499 .ndo_change_mtu = vxlan_change_mtu,
2500 .ndo_fill_metadata_dst = vxlan_fill_metadata_dst,
2501};
2502
stephen hemmingerd3428942012-10-01 12:32:35 +00002503/* Info for udev, that this is a virtual tunnel endpoint */
2504static struct device_type vxlan_type = {
2505 .name = "vxlan",
2506};
2507
Sabrina Dubrocae5de25d2016-07-11 13:12:28 +02002508/* Calls the ndo_udp_tunnel_add of the caller in order to
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002509 * supply the listening VXLAN udp ports. Callers are expected
Sabrina Dubrocae5de25d2016-07-11 13:12:28 +02002510 * to implement the ndo_udp_tunnel_add.
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002511 */
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02002512static void vxlan_push_rx_ports(struct net_device *dev)
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002513{
2514 struct vxlan_sock *vs;
2515 struct net *net = dev_net(dev);
2516 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Joseph Gasparakis35e42372013-09-13 07:34:13 -07002517 unsigned int i;
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002518
2519 spin_lock(&vn->sock_lock);
2520 for (i = 0; i < PORT_HASH_SIZE; ++i) {
Alexander Duycke7b3db52016-06-16 12:20:52 -07002521 hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist)
2522 udp_tunnel_push_rx_port(dev, vs->sock,
Alexander Duyckb9adcd692016-06-16 12:23:19 -07002523 (vs->flags & VXLAN_F_GPE) ?
2524 UDP_TUNNEL_TYPE_VXLAN_GPE :
Alexander Duycke7b3db52016-06-16 12:20:52 -07002525 UDP_TUNNEL_TYPE_VXLAN);
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002526 }
2527 spin_unlock(&vn->sock_lock);
2528}
Joseph Gasparakis53cf52752013-09-04 02:13:38 -07002529
stephen hemmingerd3428942012-10-01 12:32:35 +00002530/* Initialize the device structure. */
2531static void vxlan_setup(struct net_device *dev)
2532{
2533 struct vxlan_dev *vxlan = netdev_priv(dev);
Cong Wang31fec5a2013-05-27 22:35:52 +00002534 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00002535
Jiri Benc65226ef2016-04-28 16:36:30 +02002536 eth_hw_addr_random(dev);
2537 ether_setup(dev);
2538
Stephen Hemmingerebf40632013-06-17 14:16:11 -07002539 dev->destructor = free_netdev;
stephen hemmingerd3428942012-10-01 12:32:35 +00002540 SET_NETDEV_DEVTYPE(dev, &vxlan_type);
2541
stephen hemmingerd3428942012-10-01 12:32:35 +00002542 dev->features |= NETIF_F_LLTX;
Joseph Gasparakisd6727fe2012-12-07 14:14:16 +00002543 dev->features |= NETIF_F_SG | NETIF_F_HW_CSUM;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002544 dev->features |= NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002545 dev->features |= NETIF_F_GSO_SOFTWARE;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002546
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002547 dev->vlan_features = dev->features;
2548 dev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Joseph Gasparakis0afb1662012-12-07 14:14:18 +00002549 dev->hw_features |= NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_RXCSUM;
Pravin B Shelar05c0db02013-03-07 13:22:36 +00002550 dev->hw_features |= NETIF_F_GSO_SOFTWARE;
Pravin B Shelar1eaa8172013-08-19 11:23:29 -07002551 dev->hw_features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
Eric Dumazet02875872014-10-05 18:38:35 -07002552 netif_keep_dst(dev);
Jiri Benc0c867c92016-04-05 14:47:10 +02002553 dev->priv_flags |= IFF_NO_QUEUE;
stephen hemmingerd3428942012-10-01 12:32:35 +00002554
stephen hemminger553675f2013-05-16 11:35:20 +00002555 INIT_LIST_HEAD(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00002556 spin_lock_init(&vxlan->hash_lock);
2557
2558 init_timer_deferrable(&vxlan->age_timer);
2559 vxlan->age_timer.function = vxlan_cleanup;
2560 vxlan->age_timer.data = (unsigned long) vxlan;
2561
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002562 vxlan->cfg.dst_port = htons(vxlan_port);
stephen hemminger05f47d62012-10-09 20:35:50 +00002563
stephen hemmingerd3428942012-10-01 12:32:35 +00002564 vxlan->dev = dev;
2565
Tom Herbert58ce31c2015-08-19 17:07:33 -07002566 gro_cells_init(&vxlan->gro_cells, dev);
2567
stephen hemmingerd3428942012-10-01 12:32:35 +00002568 for (h = 0; h < FDB_HASH_SIZE; ++h)
2569 INIT_HLIST_HEAD(&vxlan->fdb_head[h]);
2570}
2571
Jiri Benc0c867c92016-04-05 14:47:10 +02002572static void vxlan_ether_setup(struct net_device *dev)
2573{
Jiri Benc0c867c92016-04-05 14:47:10 +02002574 dev->priv_flags &= ~IFF_TX_SKB_SHARING;
2575 dev->priv_flags |= IFF_LIVE_ADDR_CHANGE;
2576 dev->netdev_ops = &vxlan_netdev_ether_ops;
2577}
2578
Jiri Bence1e53142016-04-05 14:47:13 +02002579static void vxlan_raw_setup(struct net_device *dev)
2580{
Jiri Benc65226ef2016-04-28 16:36:30 +02002581 dev->header_ops = NULL;
Jiri Bence1e53142016-04-05 14:47:13 +02002582 dev->type = ARPHRD_NONE;
2583 dev->hard_header_len = 0;
2584 dev->addr_len = 0;
Jiri Bence1e53142016-04-05 14:47:13 +02002585 dev->flags = IFF_POINTOPOINT | IFF_NOARP | IFF_MULTICAST;
2586 dev->netdev_ops = &vxlan_netdev_raw_ops;
2587}
2588
stephen hemmingerd3428942012-10-01 12:32:35 +00002589static const struct nla_policy vxlan_policy[IFLA_VXLAN_MAX + 1] = {
2590 [IFLA_VXLAN_ID] = { .type = NLA_U32 },
stephen hemminger5d174dd2013-04-27 11:31:55 +00002591 [IFLA_VXLAN_GROUP] = { .len = FIELD_SIZEOF(struct iphdr, daddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002592 [IFLA_VXLAN_GROUP6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002593 [IFLA_VXLAN_LINK] = { .type = NLA_U32 },
2594 [IFLA_VXLAN_LOCAL] = { .len = FIELD_SIZEOF(struct iphdr, saddr) },
Cong Wange4c7ed42013-08-31 13:44:33 +08002595 [IFLA_VXLAN_LOCAL6] = { .len = sizeof(struct in6_addr) },
stephen hemmingerd3428942012-10-01 12:32:35 +00002596 [IFLA_VXLAN_TOS] = { .type = NLA_U8 },
2597 [IFLA_VXLAN_TTL] = { .type = NLA_U8 },
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002598 [IFLA_VXLAN_LABEL] = { .type = NLA_U32 },
stephen hemmingerd3428942012-10-01 12:32:35 +00002599 [IFLA_VXLAN_LEARNING] = { .type = NLA_U8 },
2600 [IFLA_VXLAN_AGEING] = { .type = NLA_U32 },
2601 [IFLA_VXLAN_LIMIT] = { .type = NLA_U32 },
stephen hemminger05f47d62012-10-09 20:35:50 +00002602 [IFLA_VXLAN_PORT_RANGE] = { .len = sizeof(struct ifla_vxlan_port_range) },
David Stevense4f67ad2012-11-20 02:50:14 +00002603 [IFLA_VXLAN_PROXY] = { .type = NLA_U8 },
2604 [IFLA_VXLAN_RSC] = { .type = NLA_U8 },
2605 [IFLA_VXLAN_L2MISS] = { .type = NLA_U8 },
2606 [IFLA_VXLAN_L3MISS] = { .type = NLA_U8 },
Alexei Starovoitovf8a9b1b2015-07-30 20:10:22 -07002607 [IFLA_VXLAN_COLLECT_METADATA] = { .type = NLA_U8 },
stephen hemminger823aa872013-04-27 11:31:57 +00002608 [IFLA_VXLAN_PORT] = { .type = NLA_U16 },
Tom Herbert5c91ae02014-11-06 18:06:01 -08002609 [IFLA_VXLAN_UDP_CSUM] = { .type = NLA_U8 },
2610 [IFLA_VXLAN_UDP_ZERO_CSUM6_TX] = { .type = NLA_U8 },
2611 [IFLA_VXLAN_UDP_ZERO_CSUM6_RX] = { .type = NLA_U8 },
Tom Herbertdfd86452015-01-12 17:00:38 -08002612 [IFLA_VXLAN_REMCSUM_TX] = { .type = NLA_U8 },
2613 [IFLA_VXLAN_REMCSUM_RX] = { .type = NLA_U8 },
Thomas Graf35114942015-01-15 03:53:55 +01002614 [IFLA_VXLAN_GBP] = { .type = NLA_FLAG, },
Jiri Bence1e53142016-04-05 14:47:13 +02002615 [IFLA_VXLAN_GPE] = { .type = NLA_FLAG, },
Tom Herbert0ace2ca2015-02-10 16:30:32 -08002616 [IFLA_VXLAN_REMCSUM_NOPARTIAL] = { .type = NLA_FLAG },
stephen hemmingerd3428942012-10-01 12:32:35 +00002617};
2618
2619static int vxlan_validate(struct nlattr *tb[], struct nlattr *data[])
2620{
2621 if (tb[IFLA_ADDRESS]) {
2622 if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN) {
2623 pr_debug("invalid link address (not ethernet)\n");
2624 return -EINVAL;
2625 }
2626
2627 if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS]))) {
2628 pr_debug("invalid all zero ethernet address\n");
2629 return -EADDRNOTAVAIL;
2630 }
2631 }
2632
2633 if (!data)
2634 return -EINVAL;
2635
2636 if (data[IFLA_VXLAN_ID]) {
2637 __u32 id = nla_get_u32(data[IFLA_VXLAN_ID]);
2638 if (id >= VXLAN_VID_MASK)
2639 return -ERANGE;
2640 }
2641
stephen hemminger05f47d62012-10-09 20:35:50 +00002642 if (data[IFLA_VXLAN_PORT_RANGE]) {
2643 const struct ifla_vxlan_port_range *p
2644 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
2645
2646 if (ntohs(p->high) < ntohs(p->low)) {
2647 pr_debug("port range %u .. %u not valid\n",
2648 ntohs(p->low), ntohs(p->high));
2649 return -EINVAL;
2650 }
2651 }
2652
stephen hemmingerd3428942012-10-01 12:32:35 +00002653 return 0;
2654}
2655
Yan Burman1b13c972013-01-29 23:43:07 +00002656static void vxlan_get_drvinfo(struct net_device *netdev,
2657 struct ethtool_drvinfo *drvinfo)
2658{
2659 strlcpy(drvinfo->version, VXLAN_VERSION, sizeof(drvinfo->version));
2660 strlcpy(drvinfo->driver, "vxlan", sizeof(drvinfo->driver));
2661}
2662
2663static const struct ethtool_ops vxlan_ethtool_ops = {
2664 .get_drvinfo = vxlan_get_drvinfo,
2665 .get_link = ethtool_op_get_link,
2666};
2667
Tom Herbert3ee64f32014-07-13 19:49:42 -07002668static struct socket *vxlan_create_sock(struct net *net, bool ipv6,
2669 __be16 port, u32 flags)
stephen hemminger553675f2013-05-16 11:35:20 +00002670{
Cong Wange4c7ed42013-08-31 13:44:33 +08002671 struct socket *sock;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002672 struct udp_port_cfg udp_conf;
2673 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002674
Tom Herbert3ee64f32014-07-13 19:49:42 -07002675 memset(&udp_conf, 0, sizeof(udp_conf));
2676
2677 if (ipv6) {
2678 udp_conf.family = AF_INET6;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002679 udp_conf.use_udp6_rx_checksums =
Alexander Duyck3dc2b6a2014-11-24 20:08:38 -08002680 !(flags & VXLAN_F_UDP_ZERO_CSUM6_RX);
Jiri Benca43a9ef2015-08-28 20:48:22 +02002681 udp_conf.ipv6_v6only = 1;
Tom Herbert3ee64f32014-07-13 19:49:42 -07002682 } else {
2683 udp_conf.family = AF_INET;
Cong Wange4c7ed42013-08-31 13:44:33 +08002684 }
2685
Tom Herbert3ee64f32014-07-13 19:49:42 -07002686 udp_conf.local_udp_port = port;
Cong Wange4c7ed42013-08-31 13:44:33 +08002687
Tom Herbert3ee64f32014-07-13 19:49:42 -07002688 /* Open UDP socket */
2689 err = udp_sock_create(net, &udp_conf, &sock);
2690 if (err < 0)
2691 return ERR_PTR(err);
Cong Wange4c7ed42013-08-31 13:44:33 +08002692
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002693 return sock;
Cong Wange4c7ed42013-08-31 13:44:33 +08002694}
2695
2696/* Create new listen socket if needed */
Jiri Bencb1be00a2015-09-24 13:50:02 +02002697static struct vxlan_sock *vxlan_socket_create(struct net *net, bool ipv6,
2698 __be16 port, u32 flags)
Cong Wange4c7ed42013-08-31 13:44:33 +08002699{
2700 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
2701 struct vxlan_sock *vs;
2702 struct socket *sock;
Cong Wang31fec5a2013-05-27 22:35:52 +00002703 unsigned int h;
Andy Zhouacbf74a2014-09-16 17:31:18 -07002704 struct udp_tunnel_sock_cfg tunnel_cfg;
stephen hemminger553675f2013-05-16 11:35:20 +00002705
Or Gerlitzdc01e7d2014-01-20 13:59:21 +02002706 vs = kzalloc(sizeof(*vs), GFP_KERNEL);
Cong Wange4c7ed42013-08-31 13:44:33 +08002707 if (!vs)
stephen hemminger553675f2013-05-16 11:35:20 +00002708 return ERR_PTR(-ENOMEM);
2709
2710 for (h = 0; h < VNI_HASH_SIZE; ++h)
2711 INIT_HLIST_HEAD(&vs->vni_list[h]);
2712
Tom Herbert3ee64f32014-07-13 19:49:42 -07002713 sock = vxlan_create_sock(net, ipv6, port, flags);
Zhi Yong Wu39deb2c2013-10-28 14:01:48 +08002714 if (IS_ERR(sock)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002715 pr_info("Cannot bind port %d, err=%ld\n", ntohs(port),
2716 PTR_ERR(sock));
stephen hemminger553675f2013-05-16 11:35:20 +00002717 kfree(vs);
Duan Jionge50fddc2013-11-01 13:09:43 +08002718 return ERR_CAST(sock);
stephen hemminger553675f2013-05-16 11:35:20 +00002719 }
2720
Cong Wange4c7ed42013-08-31 13:44:33 +08002721 vs->sock = sock;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002722 atomic_set(&vs->refcnt, 1);
Tom Herbertaf33c1a2015-01-20 11:23:05 -08002723 vs->flags = (flags & VXLAN_F_RCV_FLAGS);
stephen hemminger553675f2013-05-16 11:35:20 +00002724
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002725 spin_lock(&vn->sock_lock);
2726 hlist_add_head_rcu(&vs->hlist, vs_head(net, port));
Alexander Duycke7b3db52016-06-16 12:20:52 -07002727 udp_tunnel_notify_add_rx_port(sock,
Alexander Duyckb9adcd692016-06-16 12:23:19 -07002728 (vs->flags & VXLAN_F_GPE) ?
2729 UDP_TUNNEL_TYPE_VXLAN_GPE :
Alexander Duycke7b3db52016-06-16 12:20:52 -07002730 UDP_TUNNEL_TYPE_VXLAN);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002731 spin_unlock(&vn->sock_lock);
stephen hemminger553675f2013-05-16 11:35:20 +00002732
2733 /* Mark socket as an encapsulation socket. */
Tom Herbert5602c482016-04-05 08:22:53 -07002734 memset(&tunnel_cfg, 0, sizeof(tunnel_cfg));
Andy Zhouacbf74a2014-09-16 17:31:18 -07002735 tunnel_cfg.sk_user_data = vs;
2736 tunnel_cfg.encap_type = 1;
Jiri Bencf2d1968ec2016-02-23 18:02:58 +01002737 tunnel_cfg.encap_rcv = vxlan_rcv;
Andy Zhouacbf74a2014-09-16 17:31:18 -07002738 tunnel_cfg.encap_destroy = NULL;
Tom Herbert5602c482016-04-05 08:22:53 -07002739 tunnel_cfg.gro_receive = vxlan_gro_receive;
2740 tunnel_cfg.gro_complete = vxlan_gro_complete;
Andy Zhouacbf74a2014-09-16 17:31:18 -07002741
2742 setup_udp_tunnel_sock(net, sock, &tunnel_cfg);
Cong Wange4c7ed42013-08-31 13:44:33 +08002743
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002744 return vs;
2745}
stephen hemminger553675f2013-05-16 11:35:20 +00002746
Jiri Bencb1be00a2015-09-24 13:50:02 +02002747static int __vxlan_sock_add(struct vxlan_dev *vxlan, bool ipv6)
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002748{
Jiri Benc205f3562015-09-24 13:50:01 +02002749 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
2750 struct vxlan_sock *vs = NULL;
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07002751
Jiri Benc205f3562015-09-24 13:50:01 +02002752 if (!vxlan->cfg.no_share) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002753 spin_lock(&vn->sock_lock);
Jiri Benc205f3562015-09-24 13:50:01 +02002754 vs = vxlan_find_sock(vxlan->net, ipv6 ? AF_INET6 : AF_INET,
2755 vxlan->cfg.dst_port, vxlan->flags);
2756 if (vs && !atomic_add_unless(&vs->refcnt, 1, 0)) {
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002757 spin_unlock(&vn->sock_lock);
Jiri Benc205f3562015-09-24 13:50:01 +02002758 return -EBUSY;
Marcelo Ricardo Leitner56ef9c92015-03-18 14:50:44 -03002759 }
2760 spin_unlock(&vn->sock_lock);
2761 }
Jiri Benc205f3562015-09-24 13:50:01 +02002762 if (!vs)
Jiri Bencb1be00a2015-09-24 13:50:02 +02002763 vs = vxlan_socket_create(vxlan->net, ipv6,
2764 vxlan->cfg.dst_port, vxlan->flags);
Jiri Benc205f3562015-09-24 13:50:01 +02002765 if (IS_ERR(vs))
2766 return PTR_ERR(vs);
Jiri Bencb1be00a2015-09-24 13:50:02 +02002767#if IS_ENABLED(CONFIG_IPV6)
2768 if (ipv6)
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002769 rcu_assign_pointer(vxlan->vn6_sock, vs);
Jiri Bencb1be00a2015-09-24 13:50:02 +02002770 else
2771#endif
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002772 rcu_assign_pointer(vxlan->vn4_sock, vs);
Jiri Benc205f3562015-09-24 13:50:01 +02002773 vxlan_vs_add_dev(vs, vxlan);
2774 return 0;
stephen hemminger553675f2013-05-16 11:35:20 +00002775}
2776
Jiri Bencb1be00a2015-09-24 13:50:02 +02002777static int vxlan_sock_add(struct vxlan_dev *vxlan)
2778{
2779 bool ipv6 = vxlan->flags & VXLAN_F_IPV6;
2780 bool metadata = vxlan->flags & VXLAN_F_COLLECT_METADATA;
2781 int ret = 0;
2782
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002783 RCU_INIT_POINTER(vxlan->vn4_sock, NULL);
Jiri Bencb1be00a2015-09-24 13:50:02 +02002784#if IS_ENABLED(CONFIG_IPV6)
pravin shelarc6fcc4f2016-10-28 09:59:15 -07002785 RCU_INIT_POINTER(vxlan->vn6_sock, NULL);
Jiri Bencb1be00a2015-09-24 13:50:02 +02002786 if (ipv6 || metadata)
2787 ret = __vxlan_sock_add(vxlan, true);
2788#endif
2789 if (!ret && (!ipv6 || metadata))
2790 ret = __vxlan_sock_add(vxlan, false);
2791 if (ret < 0)
2792 vxlan_sock_release(vxlan);
2793 return ret;
2794}
2795
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002796static int vxlan_dev_configure(struct net *src_net, struct net_device *dev,
2797 struct vxlan_config *conf)
stephen hemmingerd3428942012-10-01 12:32:35 +00002798{
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002799 struct vxlan_net *vn = net_generic(src_net, vxlan_net_id);
Nicolas Dichtel07b9b372016-01-07 11:26:53 +01002800 struct vxlan_dev *vxlan = netdev_priv(dev), *tmp;
Atzm Watanabec7995c42013-04-16 02:50:52 +00002801 struct vxlan_rdst *dst = &vxlan->default_dst;
Jiri Bencb1be00a2015-09-24 13:50:02 +02002802 unsigned short needed_headroom = ETH_HLEN;
stephen hemmingerd3428942012-10-01 12:32:35 +00002803 int err;
Cong Wange4c7ed42013-08-31 13:44:33 +08002804 bool use_ipv6 = false;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002805 __be16 default_port = vxlan->cfg.dst_port;
David Wragg7e059152016-02-10 00:05:58 +00002806 struct net_device *lowerdev = NULL;
stephen hemmingerd3428942012-10-01 12:32:35 +00002807
Jiri Bence1e53142016-04-05 14:47:13 +02002808 if (conf->flags & VXLAN_F_GPE) {
Jiri Bence1e53142016-04-05 14:47:13 +02002809 /* For now, allow GPE only together with COLLECT_METADATA.
2810 * This can be relaxed later; in such case, the other side
2811 * of the PtP link will have to be provided.
2812 */
Jiri Benc35556212016-09-02 13:37:12 +02002813 if ((conf->flags & ~VXLAN_F_ALLOWED_GPE) ||
2814 !(conf->flags & VXLAN_F_COLLECT_METADATA)) {
2815 pr_info("unsupported combination of extensions\n");
Jiri Bence1e53142016-04-05 14:47:13 +02002816 return -EINVAL;
Jiri Benc35556212016-09-02 13:37:12 +02002817 }
Jiri Bence1e53142016-04-05 14:47:13 +02002818
2819 vxlan_raw_setup(dev);
2820 } else {
2821 vxlan_ether_setup(dev);
2822 }
Jiri Benc0c867c92016-04-05 14:47:10 +02002823
Nicolas Dichtel33564bb2015-01-26 22:28:14 +01002824 vxlan->net = src_net;
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02002825
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002826 dst->remote_vni = conf->vni;
2827
2828 memcpy(&dst->remote_ip, &conf->remote_ip, sizeof(conf->remote_ip));
stephen hemmingerd3428942012-10-01 12:32:35 +00002829
Mike Rapoport5933a7b2014-04-01 09:23:01 +03002830 /* Unless IPv6 is explicitly requested, assume IPv4 */
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002831 if (!dst->remote_ip.sa.sa_family)
2832 dst->remote_ip.sa.sa_family = AF_INET;
stephen hemmingerd3428942012-10-01 12:32:35 +00002833
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002834 if (dst->remote_ip.sa.sa_family == AF_INET6 ||
Jiri Benc057ba292015-09-17 16:11:11 +02002835 vxlan->cfg.saddr.sa.sa_family == AF_INET6) {
2836 if (!IS_ENABLED(CONFIG_IPV6))
2837 return -EPFNOSUPPORT;
Cong Wange4c7ed42013-08-31 13:44:33 +08002838 use_ipv6 = true;
Jiri Bencb1be00a2015-09-24 13:50:02 +02002839 vxlan->flags |= VXLAN_F_IPV6;
Jiri Benc057ba292015-09-17 16:11:11 +02002840 }
Cong Wange4c7ed42013-08-31 13:44:33 +08002841
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002842 if (conf->label && !use_ipv6) {
2843 pr_info("label only supported in use with IPv6\n");
2844 return -EINVAL;
2845 }
2846
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002847 if (conf->remote_ifindex) {
David Wragg7e059152016-02-10 00:05:58 +00002848 lowerdev = __dev_get_by_index(src_net, conf->remote_ifindex);
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002849 dst->remote_ifindex = conf->remote_ifindex;
stephen hemmingerd3428942012-10-01 12:32:35 +00002850
stephen hemminger34e02aa2012-10-09 20:35:53 +00002851 if (!lowerdev) {
Atzm Watanabec7995c42013-04-16 02:50:52 +00002852 pr_info("ifindex %d does not exist\n", dst->remote_ifindex);
stephen hemminger34e02aa2012-10-09 20:35:53 +00002853 return -ENODEV;
stephen hemmingerd3428942012-10-01 12:32:35 +00002854 }
stephen hemminger34e02aa2012-10-09 20:35:53 +00002855
Cong Wange4c7ed42013-08-31 13:44:33 +08002856#if IS_ENABLED(CONFIG_IPV6)
2857 if (use_ipv6) {
2858 struct inet6_dev *idev = __in6_dev_get(lowerdev);
2859 if (idev && idev->cnf.disable_ipv6) {
2860 pr_info("IPv6 is disabled via sysctl\n");
2861 return -EPERM;
2862 }
Cong Wange4c7ed42013-08-31 13:44:33 +08002863 }
2864#endif
2865
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002866 if (!conf->mtu)
Cong Wange4c7ed42013-08-31 13:44:33 +08002867 dev->mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
Alexander Duyck1ba56fb2012-11-13 13:10:59 +00002868
Jiri Bencb1be00a2015-09-24 13:50:02 +02002869 needed_headroom = lowerdev->hard_header_len;
Jiri Benc9b4cdd52016-09-02 13:37:11 +02002870 } else if (vxlan_addr_multicast(&dst->remote_ip)) {
2871 pr_info("multicast destination requires interface to be specified\n");
2872 return -EINVAL;
Jiri Benc9dc2ad12015-09-17 16:11:10 +02002873 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002874
David Wragg7e059152016-02-10 00:05:58 +00002875 if (conf->mtu) {
2876 err = __vxlan_change_mtu(dev, lowerdev, dst, conf->mtu, false);
2877 if (err)
2878 return err;
2879 }
2880
Jiri Bencb1be00a2015-09-24 13:50:02 +02002881 if (use_ipv6 || conf->flags & VXLAN_F_COLLECT_METADATA)
2882 needed_headroom += VXLAN6_HEADROOM;
2883 else
2884 needed_headroom += VXLAN_HEADROOM;
2885 dev->needed_headroom = needed_headroom;
2886
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002887 memcpy(&vxlan->cfg, conf, sizeof(*conf));
Jiri Bence1e53142016-04-05 14:47:13 +02002888 if (!vxlan->cfg.dst_port) {
2889 if (conf->flags & VXLAN_F_GPE)
2890 vxlan->cfg.dst_port = 4790; /* IANA assigned VXLAN-GPE port */
2891 else
2892 vxlan->cfg.dst_port = default_port;
2893 }
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002894 vxlan->flags |= conf->flags;
stephen hemmingerd3428942012-10-01 12:32:35 +00002895
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002896 if (!vxlan->cfg.age_interval)
2897 vxlan->cfg.age_interval = FDB_AGE_DEFAULT;
Vincent Bernatafb97182012-10-30 10:27:16 +00002898
Nicolas Dichtel07b9b372016-01-07 11:26:53 +01002899 list_for_each_entry(tmp, &vn->vxlan_list, next) {
2900 if (tmp->cfg.vni == conf->vni &&
2901 (tmp->default_dst.remote_ip.sa.sa_family == AF_INET6 ||
2902 tmp->cfg.saddr.sa.sa_family == AF_INET6) == use_ipv6 &&
2903 tmp->cfg.dst_port == vxlan->cfg.dst_port &&
2904 (tmp->flags & VXLAN_F_RCV_FLAGS) ==
Jiri Benc35556212016-09-02 13:37:12 +02002905 (vxlan->flags & VXLAN_F_RCV_FLAGS)) {
2906 pr_info("duplicate VNI %u\n", be32_to_cpu(conf->vni));
2907 return -EEXIST;
2908 }
Nicolas Dichtel07b9b372016-01-07 11:26:53 +01002909 }
stephen hemminger553675f2013-05-16 11:35:20 +00002910
Wilfried Klaebe7ad24ea2014-05-11 00:12:32 +00002911 dev->ethtool_ops = &vxlan_ethtool_ops;
Yan Burman1b13c972013-01-29 23:43:07 +00002912
Sridhar Samudrala2936b6a2013-09-17 12:12:40 -07002913 /* create an fdb entry for a valid default destination */
2914 if (!vxlan_addr_any(&vxlan->default_dst.remote_ip)) {
2915 err = vxlan_fdb_create(vxlan, all_zeros_mac,
2916 &vxlan->default_dst.remote_ip,
2917 NUD_REACHABLE|NUD_PERMANENT,
2918 NLM_F_EXCL|NLM_F_CREATE,
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002919 vxlan->cfg.dst_port,
Sridhar Samudrala2936b6a2013-09-17 12:12:40 -07002920 vxlan->default_dst.remote_vni,
2921 vxlan->default_dst.remote_ifindex,
2922 NTF_SELF);
2923 if (err)
2924 return err;
2925 }
stephen hemmingerd3428942012-10-01 12:32:35 +00002926
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002927 err = register_netdevice(dev);
2928 if (err) {
Stephen Hemmingerba609e92013-06-25 17:06:01 -07002929 vxlan_fdb_delete_default(vxlan);
Mike Rapoportafbd8ba2013-06-25 16:01:51 +03002930 return err;
2931 }
2932
stephen hemminger553675f2013-05-16 11:35:20 +00002933 list_add(&vxlan->next, &vn->vxlan_list);
stephen hemminger553675f2013-05-16 11:35:20 +00002934
2935 return 0;
stephen hemmingerd3428942012-10-01 12:32:35 +00002936}
2937
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002938static int vxlan_newlink(struct net *src_net, struct net_device *dev,
2939 struct nlattr *tb[], struct nlattr *data[])
2940{
2941 struct vxlan_config conf;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002942
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002943 memset(&conf, 0, sizeof(conf));
Jesse Grosse277de52015-10-16 16:36:00 -07002944
2945 if (data[IFLA_VXLAN_ID])
Jiri Benc54bfd872016-02-16 21:58:58 +01002946 conf.vni = cpu_to_be32(nla_get_u32(data[IFLA_VXLAN_ID]));
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002947
2948 if (data[IFLA_VXLAN_GROUP]) {
2949 conf.remote_ip.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_GROUP]);
2950 } else if (data[IFLA_VXLAN_GROUP6]) {
2951 if (!IS_ENABLED(CONFIG_IPV6))
2952 return -EPFNOSUPPORT;
2953
2954 conf.remote_ip.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_GROUP6]);
2955 conf.remote_ip.sa.sa_family = AF_INET6;
2956 }
2957
2958 if (data[IFLA_VXLAN_LOCAL]) {
2959 conf.saddr.sin.sin_addr.s_addr = nla_get_in_addr(data[IFLA_VXLAN_LOCAL]);
2960 conf.saddr.sa.sa_family = AF_INET;
2961 } else if (data[IFLA_VXLAN_LOCAL6]) {
2962 if (!IS_ENABLED(CONFIG_IPV6))
2963 return -EPFNOSUPPORT;
2964
2965 /* TODO: respect scope id */
2966 conf.saddr.sin6.sin6_addr = nla_get_in6_addr(data[IFLA_VXLAN_LOCAL6]);
2967 conf.saddr.sa.sa_family = AF_INET6;
2968 }
2969
2970 if (data[IFLA_VXLAN_LINK])
2971 conf.remote_ifindex = nla_get_u32(data[IFLA_VXLAN_LINK]);
2972
2973 if (data[IFLA_VXLAN_TOS])
2974 conf.tos = nla_get_u8(data[IFLA_VXLAN_TOS]);
2975
2976 if (data[IFLA_VXLAN_TTL])
2977 conf.ttl = nla_get_u8(data[IFLA_VXLAN_TTL]);
2978
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01002979 if (data[IFLA_VXLAN_LABEL])
2980 conf.label = nla_get_be32(data[IFLA_VXLAN_LABEL]) &
2981 IPV6_FLOWLABEL_MASK;
2982
Thomas Graf0dfbdf42015-07-21 10:44:02 +02002983 if (!data[IFLA_VXLAN_LEARNING] || nla_get_u8(data[IFLA_VXLAN_LEARNING]))
2984 conf.flags |= VXLAN_F_LEARN;
2985
2986 if (data[IFLA_VXLAN_AGEING])
2987 conf.age_interval = nla_get_u32(data[IFLA_VXLAN_AGEING]);
2988
2989 if (data[IFLA_VXLAN_PROXY] && nla_get_u8(data[IFLA_VXLAN_PROXY]))
2990 conf.flags |= VXLAN_F_PROXY;
2991
2992 if (data[IFLA_VXLAN_RSC] && nla_get_u8(data[IFLA_VXLAN_RSC]))
2993 conf.flags |= VXLAN_F_RSC;
2994
2995 if (data[IFLA_VXLAN_L2MISS] && nla_get_u8(data[IFLA_VXLAN_L2MISS]))
2996 conf.flags |= VXLAN_F_L2MISS;
2997
2998 if (data[IFLA_VXLAN_L3MISS] && nla_get_u8(data[IFLA_VXLAN_L3MISS]))
2999 conf.flags |= VXLAN_F_L3MISS;
3000
3001 if (data[IFLA_VXLAN_LIMIT])
3002 conf.addrmax = nla_get_u32(data[IFLA_VXLAN_LIMIT]);
3003
Alexei Starovoitovf8a9b1b2015-07-30 20:10:22 -07003004 if (data[IFLA_VXLAN_COLLECT_METADATA] &&
3005 nla_get_u8(data[IFLA_VXLAN_COLLECT_METADATA]))
3006 conf.flags |= VXLAN_F_COLLECT_METADATA;
3007
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003008 if (data[IFLA_VXLAN_PORT_RANGE]) {
3009 const struct ifla_vxlan_port_range *p
3010 = nla_data(data[IFLA_VXLAN_PORT_RANGE]);
3011 conf.port_min = ntohs(p->low);
3012 conf.port_max = ntohs(p->high);
3013 }
3014
3015 if (data[IFLA_VXLAN_PORT])
3016 conf.dst_port = nla_get_be16(data[IFLA_VXLAN_PORT]);
3017
Alexander Duyck6ceb31c2016-02-19 11:26:31 -08003018 if (data[IFLA_VXLAN_UDP_CSUM] &&
3019 !nla_get_u8(data[IFLA_VXLAN_UDP_CSUM]))
3020 conf.flags |= VXLAN_F_UDP_ZERO_CSUM_TX;
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003021
3022 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX] &&
3023 nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_TX]))
3024 conf.flags |= VXLAN_F_UDP_ZERO_CSUM6_TX;
3025
3026 if (data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX] &&
3027 nla_get_u8(data[IFLA_VXLAN_UDP_ZERO_CSUM6_RX]))
3028 conf.flags |= VXLAN_F_UDP_ZERO_CSUM6_RX;
3029
3030 if (data[IFLA_VXLAN_REMCSUM_TX] &&
3031 nla_get_u8(data[IFLA_VXLAN_REMCSUM_TX]))
3032 conf.flags |= VXLAN_F_REMCSUM_TX;
3033
3034 if (data[IFLA_VXLAN_REMCSUM_RX] &&
3035 nla_get_u8(data[IFLA_VXLAN_REMCSUM_RX]))
3036 conf.flags |= VXLAN_F_REMCSUM_RX;
3037
3038 if (data[IFLA_VXLAN_GBP])
3039 conf.flags |= VXLAN_F_GBP;
3040
Jiri Bence1e53142016-04-05 14:47:13 +02003041 if (data[IFLA_VXLAN_GPE])
3042 conf.flags |= VXLAN_F_GPE;
3043
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003044 if (data[IFLA_VXLAN_REMCSUM_NOPARTIAL])
3045 conf.flags |= VXLAN_F_REMCSUM_NOPARTIAL;
3046
Chen Haiquance577662016-05-27 10:49:11 +08003047 if (tb[IFLA_MTU])
3048 conf.mtu = nla_get_u32(tb[IFLA_MTU]);
3049
Jiri Benc35556212016-09-02 13:37:12 +02003050 return vxlan_dev_configure(src_net, dev, &conf);
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003051}
3052
stephen hemmingerd3428942012-10-01 12:32:35 +00003053static void vxlan_dellink(struct net_device *dev, struct list_head *head)
3054{
3055 struct vxlan_dev *vxlan = netdev_priv(dev);
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003056 struct vxlan_net *vn = net_generic(vxlan->net, vxlan_net_id);
stephen hemmingerd3428942012-10-01 12:32:35 +00003057
stephen hemmingerfe5c3562013-07-13 10:18:18 -07003058 spin_lock(&vn->sock_lock);
Pravin B Shelar9c2e24e2013-08-19 11:22:48 -07003059 if (!hlist_unhashed(&vxlan->hlist))
3060 hlist_del_rcu(&vxlan->hlist);
stephen hemmingerfe5c3562013-07-13 10:18:18 -07003061 spin_unlock(&vn->sock_lock);
3062
Tom Herbert58ce31c2015-08-19 17:07:33 -07003063 gro_cells_destroy(&vxlan->gro_cells);
stephen hemminger553675f2013-05-16 11:35:20 +00003064 list_del(&vxlan->next);
stephen hemmingerd3428942012-10-01 12:32:35 +00003065 unregister_netdevice_queue(dev, head);
3066}
3067
3068static size_t vxlan_get_size(const struct net_device *dev)
3069{
3070
3071 return nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_ID */
Cong Wange4c7ed42013-08-31 13:44:33 +08003072 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_GROUP{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00003073 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LINK */
Cong Wange4c7ed42013-08-31 13:44:33 +08003074 nla_total_size(sizeof(struct in6_addr)) + /* IFLA_VXLAN_LOCAL{6} */
stephen hemmingerd3428942012-10-01 12:32:35 +00003075 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TTL */
3076 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_TOS */
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01003077 nla_total_size(sizeof(__be32)) + /* IFLA_VXLAN_LABEL */
stephen hemmingerd3428942012-10-01 12:32:35 +00003078 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_LEARNING */
David Stevense4f67ad2012-11-20 02:50:14 +00003079 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_PROXY */
3080 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_RSC */
3081 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L2MISS */
3082 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_L3MISS */
Alexei Starovoitovda8b43c2015-08-04 22:51:07 -07003083 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_COLLECT_METADATA */
stephen hemmingerd3428942012-10-01 12:32:35 +00003084 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_AGEING */
3085 nla_total_size(sizeof(__u32)) + /* IFLA_VXLAN_LIMIT */
stephen hemminger05f47d62012-10-09 20:35:50 +00003086 nla_total_size(sizeof(struct ifla_vxlan_port_range)) +
Tom Herbert359a0ea2014-06-04 17:20:29 -07003087 nla_total_size(sizeof(__be16)) + /* IFLA_VXLAN_PORT */
3088 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_CSUM */
3089 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_TX */
3090 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_UDP_ZERO_CSUM6_RX */
Tom Herbertdfd86452015-01-12 17:00:38 -08003091 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_TX */
3092 nla_total_size(sizeof(__u8)) + /* IFLA_VXLAN_REMCSUM_RX */
stephen hemmingerd3428942012-10-01 12:32:35 +00003093 0;
3094}
3095
3096static int vxlan_fill_info(struct sk_buff *skb, const struct net_device *dev)
3097{
3098 const struct vxlan_dev *vxlan = netdev_priv(dev);
Atzm Watanabec7995c42013-04-16 02:50:52 +00003099 const struct vxlan_rdst *dst = &vxlan->default_dst;
stephen hemminger05f47d62012-10-09 20:35:50 +00003100 struct ifla_vxlan_port_range ports = {
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003101 .low = htons(vxlan->cfg.port_min),
3102 .high = htons(vxlan->cfg.port_max),
stephen hemminger05f47d62012-10-09 20:35:50 +00003103 };
stephen hemmingerd3428942012-10-01 12:32:35 +00003104
Jiri Benc54bfd872016-02-16 21:58:58 +01003105 if (nla_put_u32(skb, IFLA_VXLAN_ID, be32_to_cpu(dst->remote_vni)))
stephen hemmingerd3428942012-10-01 12:32:35 +00003106 goto nla_put_failure;
3107
Cong Wange4c7ed42013-08-31 13:44:33 +08003108 if (!vxlan_addr_any(&dst->remote_ip)) {
3109 if (dst->remote_ip.sa.sa_family == AF_INET) {
Jiri Benc930345e2015-03-29 16:59:25 +02003110 if (nla_put_in_addr(skb, IFLA_VXLAN_GROUP,
3111 dst->remote_ip.sin.sin_addr.s_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003112 goto nla_put_failure;
3113#if IS_ENABLED(CONFIG_IPV6)
3114 } else {
Jiri Benc930345e2015-03-29 16:59:25 +02003115 if (nla_put_in6_addr(skb, IFLA_VXLAN_GROUP6,
3116 &dst->remote_ip.sin6.sin6_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003117 goto nla_put_failure;
3118#endif
3119 }
3120 }
stephen hemmingerd3428942012-10-01 12:32:35 +00003121
Atzm Watanabec7995c42013-04-16 02:50:52 +00003122 if (dst->remote_ifindex && nla_put_u32(skb, IFLA_VXLAN_LINK, dst->remote_ifindex))
stephen hemmingerd3428942012-10-01 12:32:35 +00003123 goto nla_put_failure;
3124
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003125 if (!vxlan_addr_any(&vxlan->cfg.saddr)) {
3126 if (vxlan->cfg.saddr.sa.sa_family == AF_INET) {
Jiri Benc930345e2015-03-29 16:59:25 +02003127 if (nla_put_in_addr(skb, IFLA_VXLAN_LOCAL,
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003128 vxlan->cfg.saddr.sin.sin_addr.s_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003129 goto nla_put_failure;
3130#if IS_ENABLED(CONFIG_IPV6)
3131 } else {
Jiri Benc930345e2015-03-29 16:59:25 +02003132 if (nla_put_in6_addr(skb, IFLA_VXLAN_LOCAL6,
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003133 &vxlan->cfg.saddr.sin6.sin6_addr))
Cong Wange4c7ed42013-08-31 13:44:33 +08003134 goto nla_put_failure;
3135#endif
3136 }
3137 }
stephen hemmingerd3428942012-10-01 12:32:35 +00003138
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003139 if (nla_put_u8(skb, IFLA_VXLAN_TTL, vxlan->cfg.ttl) ||
3140 nla_put_u8(skb, IFLA_VXLAN_TOS, vxlan->cfg.tos) ||
Daniel Borkmanne7f70af2016-03-09 03:00:03 +01003141 nla_put_be32(skb, IFLA_VXLAN_LABEL, vxlan->cfg.label) ||
David Stevense4f67ad2012-11-20 02:50:14 +00003142 nla_put_u8(skb, IFLA_VXLAN_LEARNING,
3143 !!(vxlan->flags & VXLAN_F_LEARN)) ||
3144 nla_put_u8(skb, IFLA_VXLAN_PROXY,
3145 !!(vxlan->flags & VXLAN_F_PROXY)) ||
3146 nla_put_u8(skb, IFLA_VXLAN_RSC, !!(vxlan->flags & VXLAN_F_RSC)) ||
3147 nla_put_u8(skb, IFLA_VXLAN_L2MISS,
3148 !!(vxlan->flags & VXLAN_F_L2MISS)) ||
3149 nla_put_u8(skb, IFLA_VXLAN_L3MISS,
3150 !!(vxlan->flags & VXLAN_F_L3MISS)) ||
Alexei Starovoitovda8b43c2015-08-04 22:51:07 -07003151 nla_put_u8(skb, IFLA_VXLAN_COLLECT_METADATA,
3152 !!(vxlan->flags & VXLAN_F_COLLECT_METADATA)) ||
Thomas Graf0dfbdf42015-07-21 10:44:02 +02003153 nla_put_u32(skb, IFLA_VXLAN_AGEING, vxlan->cfg.age_interval) ||
3154 nla_put_u32(skb, IFLA_VXLAN_LIMIT, vxlan->cfg.addrmax) ||
3155 nla_put_be16(skb, IFLA_VXLAN_PORT, vxlan->cfg.dst_port) ||
Tom Herbert359a0ea2014-06-04 17:20:29 -07003156 nla_put_u8(skb, IFLA_VXLAN_UDP_CSUM,
Alexander Duyck6ceb31c2016-02-19 11:26:31 -08003157 !(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM_TX)) ||
Tom Herbert359a0ea2014-06-04 17:20:29 -07003158 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_TX,
3159 !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_TX)) ||
3160 nla_put_u8(skb, IFLA_VXLAN_UDP_ZERO_CSUM6_RX,
Tom Herbertdfd86452015-01-12 17:00:38 -08003161 !!(vxlan->flags & VXLAN_F_UDP_ZERO_CSUM6_RX)) ||
3162 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_TX,
3163 !!(vxlan->flags & VXLAN_F_REMCSUM_TX)) ||
3164 nla_put_u8(skb, IFLA_VXLAN_REMCSUM_RX,
3165 !!(vxlan->flags & VXLAN_F_REMCSUM_RX)))
stephen hemmingerd3428942012-10-01 12:32:35 +00003166 goto nla_put_failure;
3167
stephen hemminger05f47d62012-10-09 20:35:50 +00003168 if (nla_put(skb, IFLA_VXLAN_PORT_RANGE, sizeof(ports), &ports))
3169 goto nla_put_failure;
3170
Thomas Graf35114942015-01-15 03:53:55 +01003171 if (vxlan->flags & VXLAN_F_GBP &&
3172 nla_put_flag(skb, IFLA_VXLAN_GBP))
3173 goto nla_put_failure;
3174
Jiri Bence1e53142016-04-05 14:47:13 +02003175 if (vxlan->flags & VXLAN_F_GPE &&
3176 nla_put_flag(skb, IFLA_VXLAN_GPE))
3177 goto nla_put_failure;
3178
Tom Herbert0ace2ca2015-02-10 16:30:32 -08003179 if (vxlan->flags & VXLAN_F_REMCSUM_NOPARTIAL &&
3180 nla_put_flag(skb, IFLA_VXLAN_REMCSUM_NOPARTIAL))
3181 goto nla_put_failure;
3182
stephen hemmingerd3428942012-10-01 12:32:35 +00003183 return 0;
3184
3185nla_put_failure:
3186 return -EMSGSIZE;
3187}
3188
Nicolas Dichtel1728d4f2015-01-15 15:11:17 +01003189static struct net *vxlan_get_link_net(const struct net_device *dev)
3190{
3191 struct vxlan_dev *vxlan = netdev_priv(dev);
3192
3193 return vxlan->net;
3194}
3195
stephen hemmingerd3428942012-10-01 12:32:35 +00003196static struct rtnl_link_ops vxlan_link_ops __read_mostly = {
3197 .kind = "vxlan",
3198 .maxtype = IFLA_VXLAN_MAX,
3199 .policy = vxlan_policy,
3200 .priv_size = sizeof(struct vxlan_dev),
3201 .setup = vxlan_setup,
3202 .validate = vxlan_validate,
3203 .newlink = vxlan_newlink,
3204 .dellink = vxlan_dellink,
3205 .get_size = vxlan_get_size,
3206 .fill_info = vxlan_fill_info,
Nicolas Dichtel1728d4f2015-01-15 15:11:17 +01003207 .get_link_net = vxlan_get_link_net,
stephen hemmingerd3428942012-10-01 12:32:35 +00003208};
3209
Nicolas Dichtelcf5da332016-06-13 10:31:05 +02003210struct net_device *vxlan_dev_create(struct net *net, const char *name,
3211 u8 name_assign_type,
3212 struct vxlan_config *conf)
3213{
3214 struct nlattr *tb[IFLA_MAX + 1];
3215 struct net_device *dev;
3216 int err;
3217
3218 memset(&tb, 0, sizeof(tb));
3219
3220 dev = rtnl_create_link(net, name, name_assign_type,
3221 &vxlan_link_ops, tb);
3222 if (IS_ERR(dev))
3223 return dev;
3224
3225 err = vxlan_dev_configure(net, dev, conf);
3226 if (err < 0) {
3227 free_netdev(dev);
3228 return ERR_PTR(err);
3229 }
3230
3231 err = rtnl_configure_link(dev, NULL);
3232 if (err < 0) {
3233 LIST_HEAD(list_kill);
3234
3235 vxlan_dellink(dev, &list_kill);
3236 unregister_netdevice_many(&list_kill);
3237 return ERR_PTR(err);
3238 }
3239
3240 return dev;
3241}
3242EXPORT_SYMBOL_GPL(vxlan_dev_create);
3243
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003244static void vxlan_handle_lowerdev_unregister(struct vxlan_net *vn,
3245 struct net_device *dev)
3246{
3247 struct vxlan_dev *vxlan, *next;
3248 LIST_HEAD(list_kill);
3249
3250 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
3251 struct vxlan_rdst *dst = &vxlan->default_dst;
3252
3253 /* In case we created vxlan device with carrier
3254 * and we loose the carrier due to module unload
3255 * we also need to remove vxlan device. In other
3256 * cases, it's not necessary and remote_ifindex
3257 * is 0 here, so no matches.
3258 */
3259 if (dst->remote_ifindex == dev->ifindex)
3260 vxlan_dellink(vxlan->dev, &list_kill);
3261 }
3262
3263 unregister_netdevice_many(&list_kill);
3264}
3265
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02003266static int vxlan_netdevice_event(struct notifier_block *unused,
3267 unsigned long event, void *ptr)
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003268{
3269 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
Daniel Borkmann783c1462014-01-22 21:07:53 +01003270 struct vxlan_net *vn = net_generic(dev_net(dev), vxlan_net_id);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003271
Daniel Borkmann783c1462014-01-22 21:07:53 +01003272 if (event == NETDEV_UNREGISTER)
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003273 vxlan_handle_lowerdev_unregister(vn, dev);
Alexander Duyck7c46a642016-06-16 12:21:00 -07003274 else if (event == NETDEV_UDP_TUNNEL_PUSH_INFO)
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02003275 vxlan_push_rx_ports(dev);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003276
3277 return NOTIFY_DONE;
3278}
3279
3280static struct notifier_block vxlan_notifier_block __read_mostly = {
Hannes Frederic Sowab7aade12016-04-18 21:19:47 +02003281 .notifier_call = vxlan_netdevice_event,
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003282};
3283
stephen hemmingerd3428942012-10-01 12:32:35 +00003284static __net_init int vxlan_init_net(struct net *net)
3285{
3286 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
Cong Wang31fec5a2013-05-27 22:35:52 +00003287 unsigned int h;
stephen hemmingerd3428942012-10-01 12:32:35 +00003288
stephen hemminger553675f2013-05-16 11:35:20 +00003289 INIT_LIST_HEAD(&vn->vxlan_list);
Stephen Hemminger1c51a912013-06-17 14:16:11 -07003290 spin_lock_init(&vn->sock_lock);
stephen hemmingerd3428942012-10-01 12:32:35 +00003291
stephen hemminger553675f2013-05-16 11:35:20 +00003292 for (h = 0; h < PORT_HASH_SIZE; ++h)
3293 INIT_HLIST_HEAD(&vn->sock_list[h]);
stephen hemmingerd3428942012-10-01 12:32:35 +00003294
3295 return 0;
3296}
3297
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003298static void __net_exit vxlan_exit_net(struct net *net)
3299{
3300 struct vxlan_net *vn = net_generic(net, vxlan_net_id);
3301 struct vxlan_dev *vxlan, *next;
3302 struct net_device *dev, *aux;
3303 LIST_HEAD(list);
3304
3305 rtnl_lock();
3306 for_each_netdev_safe(net, dev, aux)
3307 if (dev->rtnl_link_ops == &vxlan_link_ops)
3308 unregister_netdevice_queue(dev, &list);
3309
3310 list_for_each_entry_safe(vxlan, next, &vn->vxlan_list, next) {
3311 /* If vxlan->dev is in the same netns, it has already been added
3312 * to the list by the previous loop.
3313 */
Tom Herbert58ce31c2015-08-19 17:07:33 -07003314 if (!net_eq(dev_net(vxlan->dev), net)) {
3315 gro_cells_destroy(&vxlan->gro_cells);
John W. Linville13c3ed62015-05-18 13:51:24 -04003316 unregister_netdevice_queue(vxlan->dev, &list);
Tom Herbert58ce31c2015-08-19 17:07:33 -07003317 }
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003318 }
3319
3320 unregister_netdevice_many(&list);
3321 rtnl_unlock();
3322}
3323
stephen hemmingerd3428942012-10-01 12:32:35 +00003324static struct pernet_operations vxlan_net_ops = {
3325 .init = vxlan_init_net,
Nicolas Dichtelf01ec1c2014-04-24 10:02:49 +02003326 .exit = vxlan_exit_net,
stephen hemmingerd3428942012-10-01 12:32:35 +00003327 .id = &vxlan_net_id,
3328 .size = sizeof(struct vxlan_net),
3329};
3330
3331static int __init vxlan_init_module(void)
3332{
3333 int rc;
3334
3335 get_random_bytes(&vxlan_salt, sizeof(vxlan_salt));
3336
Daniel Borkmann783c1462014-01-22 21:07:53 +01003337 rc = register_pernet_subsys(&vxlan_net_ops);
stephen hemmingerd3428942012-10-01 12:32:35 +00003338 if (rc)
3339 goto out1;
3340
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003341 rc = register_netdevice_notifier(&vxlan_notifier_block);
stephen hemmingerd3428942012-10-01 12:32:35 +00003342 if (rc)
3343 goto out2;
3344
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003345 rc = rtnl_link_register(&vxlan_link_ops);
3346 if (rc)
3347 goto out3;
stephen hemmingerd3428942012-10-01 12:32:35 +00003348
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003349 return 0;
3350out3:
3351 unregister_netdevice_notifier(&vxlan_notifier_block);
stephen hemmingerd3428942012-10-01 12:32:35 +00003352out2:
Daniel Borkmann783c1462014-01-22 21:07:53 +01003353 unregister_pernet_subsys(&vxlan_net_ops);
stephen hemmingerd3428942012-10-01 12:32:35 +00003354out1:
3355 return rc;
3356}
Cong Wang7332a132013-05-27 22:35:53 +00003357late_initcall(vxlan_init_module);
stephen hemmingerd3428942012-10-01 12:32:35 +00003358
3359static void __exit vxlan_cleanup_module(void)
3360{
Stephen Hemmingerb7153982013-06-17 14:16:09 -07003361 rtnl_link_unregister(&vxlan_link_ops);
Daniel Borkmannacaf4e7092014-01-13 18:41:19 +01003362 unregister_netdevice_notifier(&vxlan_notifier_block);
Daniel Borkmann783c1462014-01-22 21:07:53 +01003363 unregister_pernet_subsys(&vxlan_net_ops);
3364 /* rcu_barrier() is called by netns */
stephen hemmingerd3428942012-10-01 12:32:35 +00003365}
3366module_exit(vxlan_cleanup_module);
3367
3368MODULE_LICENSE("GPL");
3369MODULE_VERSION(VXLAN_VERSION);
stephen hemminger3b8df3c2013-04-27 11:31:52 +00003370MODULE_AUTHOR("Stephen Hemminger <stephen@networkplumber.org>");
Jesse Brandeburgead51392014-01-17 11:00:33 -08003371MODULE_DESCRIPTION("Driver for VXLAN encapsulated traffic");
stephen hemmingerd3428942012-10-01 12:32:35 +00003372MODULE_ALIAS_RTNL_LINK("vxlan");