]>
Commit | Line | Data |
---|---|---|
012a5729 PS |
1 | #ifndef __NET_VXLAN_H |
2 | #define __NET_VXLAN_H 1 | |
3 | ||
5f35227e JG |
4 | #include <linux/ip.h> |
5 | #include <linux/ipv6.h> | |
6 | #include <linux/if_vlan.h> | |
012a5729 PS |
7 | #include <linux/skbuff.h> |
8 | #include <linux/netdevice.h> | |
9 | #include <linux/udp.h> | |
ee122c79 | 10 | #include <net/dst_metadata.h> |
012a5729 | 11 | |
828788ac JB |
12 | /* VXLAN protocol (RFC 7348) header: |
13 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
14 | * |R|R|R|R|I|R|R|R| Reserved | | |
15 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
16 | * | VXLAN Network Identifier (VNI) | Reserved | | |
17 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
18 | * | |
19 | * I = VXLAN Network Identifier (VNI) present. | |
20 | */ | |
21 | struct vxlanhdr { | |
22 | __be32 vx_flags; | |
23 | __be32 vx_vni; | |
24 | }; | |
25 | ||
26 | /* VXLAN header flags. */ | |
54bfd872 | 27 | #define VXLAN_HF_VNI cpu_to_be32(BIT(27)) |
828788ac JB |
28 | |
29 | #define VXLAN_N_VID (1u << 24) | |
30 | #define VXLAN_VID_MASK (VXLAN_N_VID - 1) | |
54bfd872 | 31 | #define VXLAN_VNI_MASK cpu_to_be32(VXLAN_VID_MASK << 8) |
828788ac JB |
32 | #define VXLAN_HLEN (sizeof(struct udphdr) + sizeof(struct vxlanhdr)) |
33 | ||
34 | #define VNI_HASH_BITS 10 | |
35 | #define VNI_HASH_SIZE (1<<VNI_HASH_BITS) | |
36 | #define FDB_HASH_BITS 8 | |
37 | #define FDB_HASH_SIZE (1<<FDB_HASH_BITS) | |
38 | ||
39 | /* Remote checksum offload for VXLAN (VXLAN_F_REMCSUM_[RT]X): | |
40 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
41 | * |R|R|R|R|I|R|R|R|R|R|C| Reserved | | |
42 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
43 | * | VXLAN Network Identifier (VNI) |O| Csum start | | |
44 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
45 | * | |
46 | * C = Remote checksum offload bit. When set indicates that the | |
47 | * remote checksum offload data is present. | |
48 | * | |
49 | * O = Offset bit. Indicates the checksum offset relative to | |
50 | * checksum start. | |
51 | * | |
52 | * Csum start = Checksum start divided by two. | |
53 | * | |
54 | * http://tools.ietf.org/html/draft-herbert-vxlan-rco | |
55 | */ | |
56 | ||
57 | /* VXLAN-RCO header flags. */ | |
54bfd872 | 58 | #define VXLAN_HF_RCO cpu_to_be32(BIT(21)) |
828788ac JB |
59 | |
60 | /* Remote checksum offload header option */ | |
54bfd872 JB |
61 | #define VXLAN_RCO_MASK cpu_to_be32(0x7f) /* Last byte of vni field */ |
62 | #define VXLAN_RCO_UDP cpu_to_be32(0x80) /* Indicate UDP RCO (TCP when not set *) */ | |
63 | #define VXLAN_RCO_SHIFT 1 /* Left shift of start */ | |
828788ac | 64 | #define VXLAN_RCO_SHIFT_MASK ((1 << VXLAN_RCO_SHIFT) - 1) |
54bfd872 | 65 | #define VXLAN_MAX_REMCSUM_START (0x7f << VXLAN_RCO_SHIFT) |
828788ac | 66 | |
3511494c | 67 | /* |
828788ac | 68 | * VXLAN Group Based Policy Extension (VXLAN_F_GBP): |
3511494c | 69 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
828788ac | 70 | * |G|R|R|R|I|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID | |
3511494c TG |
71 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |
72 | * | VXLAN Network Identifier (VNI) | Reserved | | |
73 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
74 | * | |
828788ac JB |
75 | * G = Group Policy ID present. |
76 | * | |
3511494c TG |
77 | * D = Don't Learn bit. When set, this bit indicates that the egress |
78 | * VTEP MUST NOT learn the source address of the encapsulated frame. | |
79 | * | |
80 | * A = Indicates that the group policy has already been applied to | |
81 | * this packet. Policies MUST NOT be applied by devices when the | |
82 | * A bit is set. | |
83 | * | |
828788ac | 84 | * https://tools.ietf.org/html/draft-smith-vxlan-group-policy |
3511494c TG |
85 | */ |
86 | struct vxlanhdr_gbp { | |
0e715d6f | 87 | u8 vx_flags; |
3511494c | 88 | #ifdef __LITTLE_ENDIAN_BITFIELD |
0e715d6f | 89 | u8 reserved_flags1:3, |
3511494c TG |
90 | policy_applied:1, |
91 | reserved_flags2:2, | |
92 | dont_learn:1, | |
93 | reserved_flags3:1; | |
94 | #elif defined(__BIG_ENDIAN_BITFIELD) | |
0e715d6f | 95 | u8 reserved_flags1:1, |
3511494c TG |
96 | dont_learn:1, |
97 | reserved_flags2:2, | |
98 | policy_applied:1, | |
99 | reserved_flags3:3; | |
100 | #else | |
101 | #error "Please fix <asm/byteorder.h>" | |
102 | #endif | |
103 | __be16 policy_id; | |
104 | __be32 vx_vni; | |
105 | }; | |
106 | ||
828788ac | 107 | /* VXLAN-GBP header flags. */ |
54bfd872 | 108 | #define VXLAN_HF_GBP cpu_to_be32(BIT(31)) |
828788ac | 109 | |
54bfd872 | 110 | #define VXLAN_GBP_USED_BITS (VXLAN_HF_GBP | cpu_to_be32(0xFFFFFF)) |
3511494c TG |
111 | |
112 | /* skb->mark mapping | |
113 | * | |
114 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
115 | * |R|R|R|R|R|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID | | |
116 | * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | |
117 | */ | |
118 | #define VXLAN_GBP_DONT_LEARN (BIT(6) << 16) | |
119 | #define VXLAN_GBP_POLICY_APPLIED (BIT(3) << 16) | |
120 | #define VXLAN_GBP_ID_MASK (0xFFFF) | |
121 | ||
3511494c | 122 | struct vxlan_metadata { |
3511494c TG |
123 | u32 gbp; |
124 | }; | |
125 | ||
012a5729 PS |
126 | /* per UDP socket information */ |
127 | struct vxlan_sock { | |
128 | struct hlist_node hlist; | |
012a5729 PS |
129 | struct work_struct del_work; |
130 | struct socket *sock; | |
131 | struct rcu_head rcu; | |
132 | struct hlist_head vni_list[VNI_HASH_SIZE]; | |
133 | atomic_t refcnt; | |
dc01e7d3 | 134 | struct udp_offload udp_offloads; |
dfd8645e | 135 | u32 flags; |
012a5729 PS |
136 | }; |
137 | ||
0dfbdf41 TG |
138 | union vxlan_addr { |
139 | struct sockaddr_in sin; | |
140 | struct sockaddr_in6 sin6; | |
141 | struct sockaddr sa; | |
142 | }; | |
143 | ||
144 | struct vxlan_rdst { | |
145 | union vxlan_addr remote_ip; | |
146 | __be16 remote_port; | |
54bfd872 | 147 | __be32 remote_vni; |
0dfbdf41 TG |
148 | u32 remote_ifindex; |
149 | struct list_head list; | |
150 | struct rcu_head rcu; | |
0c1d70af | 151 | struct dst_cache dst_cache; |
0dfbdf41 TG |
152 | }; |
153 | ||
154 | struct vxlan_config { | |
155 | union vxlan_addr remote_ip; | |
156 | union vxlan_addr saddr; | |
54bfd872 | 157 | __be32 vni; |
0dfbdf41 TG |
158 | int remote_ifindex; |
159 | int mtu; | |
160 | __be16 dst_port; | |
0e715d6f JB |
161 | u16 port_min; |
162 | u16 port_max; | |
163 | u8 tos; | |
164 | u8 ttl; | |
e7f70af1 | 165 | __be32 label; |
0dfbdf41 TG |
166 | u32 flags; |
167 | unsigned long age_interval; | |
168 | unsigned int addrmax; | |
169 | bool no_share; | |
170 | }; | |
171 | ||
172 | /* Pseudo network device */ | |
173 | struct vxlan_dev { | |
174 | struct hlist_node hlist; /* vni hash table */ | |
175 | struct list_head next; /* vxlan's per namespace list */ | |
b1be00a6 JB |
176 | struct vxlan_sock *vn4_sock; /* listening socket for IPv4 */ |
177 | #if IS_ENABLED(CONFIG_IPV6) | |
178 | struct vxlan_sock *vn6_sock; /* listening socket for IPv6 */ | |
179 | #endif | |
0dfbdf41 TG |
180 | struct net_device *dev; |
181 | struct net *net; /* netns for packet i/o */ | |
182 | struct vxlan_rdst default_dst; /* default destination */ | |
183 | u32 flags; /* VXLAN_F_* in vxlan.h */ | |
184 | ||
185 | struct timer_list age_timer; | |
186 | spinlock_t hash_lock; | |
187 | unsigned int addrcnt; | |
58ce31cc | 188 | struct gro_cells gro_cells; |
0dfbdf41 TG |
189 | |
190 | struct vxlan_config cfg; | |
191 | ||
192 | struct hlist_head fdb_head[FDB_HASH_SIZE]; | |
193 | }; | |
194 | ||
359a0ea9 TH |
195 | #define VXLAN_F_LEARN 0x01 |
196 | #define VXLAN_F_PROXY 0x02 | |
197 | #define VXLAN_F_RSC 0x04 | |
198 | #define VXLAN_F_L2MISS 0x08 | |
199 | #define VXLAN_F_L3MISS 0x10 | |
200 | #define VXLAN_F_IPV6 0x20 | |
6ceb31ca | 201 | #define VXLAN_F_UDP_ZERO_CSUM_TX 0x40 |
359a0ea9 TH |
202 | #define VXLAN_F_UDP_ZERO_CSUM6_TX 0x80 |
203 | #define VXLAN_F_UDP_ZERO_CSUM6_RX 0x100 | |
dfd8645e TH |
204 | #define VXLAN_F_REMCSUM_TX 0x200 |
205 | #define VXLAN_F_REMCSUM_RX 0x400 | |
3511494c | 206 | #define VXLAN_F_GBP 0x800 |
0ace2ca8 | 207 | #define VXLAN_F_REMCSUM_NOPARTIAL 0x1000 |
ee122c79 | 208 | #define VXLAN_F_COLLECT_METADATA 0x2000 |
359a0ea9 | 209 | |
d299ce14 | 210 | /* Flags that are used in the receive path. These flags must match in |
af33c1ad TH |
211 | * order for a socket to be shareable |
212 | */ | |
213 | #define VXLAN_F_RCV_FLAGS (VXLAN_F_GBP | \ | |
214 | VXLAN_F_UDP_ZERO_CSUM6_RX | \ | |
0ace2ca8 | 215 | VXLAN_F_REMCSUM_RX | \ |
ee122c79 | 216 | VXLAN_F_REMCSUM_NOPARTIAL | \ |
da8b43c0 | 217 | VXLAN_F_COLLECT_METADATA) |
ac5132d1 | 218 | |
0dfbdf41 TG |
219 | struct net_device *vxlan_dev_create(struct net *net, const char *name, |
220 | u8 name_assign_type, struct vxlan_config *conf); | |
221 | ||
b1be00a6 JB |
222 | static inline __be16 vxlan_dev_dst_port(struct vxlan_dev *vxlan, |
223 | unsigned short family) | |
614732ea | 224 | { |
b1be00a6 JB |
225 | #if IS_ENABLED(CONFIG_IPV6) |
226 | if (family == AF_INET6) | |
227 | return inet_sk(vxlan->vn6_sock->sock->sk)->inet_sport; | |
228 | #endif | |
229 | return inet_sk(vxlan->vn4_sock->sock->sk)->inet_sport; | |
614732ea | 230 | } |
49560532 | 231 | |
5f35227e JG |
232 | static inline netdev_features_t vxlan_features_check(struct sk_buff *skb, |
233 | netdev_features_t features) | |
11bf7828 | 234 | { |
5f35227e JG |
235 | u8 l4_hdr = 0; |
236 | ||
237 | if (!skb->encapsulation) | |
238 | return features; | |
239 | ||
240 | switch (vlan_get_protocol(skb)) { | |
241 | case htons(ETH_P_IP): | |
242 | l4_hdr = ip_hdr(skb)->protocol; | |
243 | break; | |
244 | case htons(ETH_P_IPV6): | |
245 | l4_hdr = ipv6_hdr(skb)->nexthdr; | |
246 | break; | |
247 | default: | |
248 | return features;; | |
249 | } | |
250 | ||
251 | if ((l4_hdr == IPPROTO_UDP) && | |
11bf7828 JS |
252 | (skb->inner_protocol_type != ENCAP_TYPE_ETHER || |
253 | skb->inner_protocol != htons(ETH_P_TEB) || | |
254 | (skb_inner_mac_header(skb) - skb_transport_header(skb) != | |
255 | sizeof(struct udphdr) + sizeof(struct vxlanhdr)))) | |
a188222b | 256 | return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); |
11bf7828 | 257 | |
5f35227e | 258 | return features; |
11bf7828 | 259 | } |
23e62de3 | 260 | |
e6cd988c JG |
261 | /* IP header + UDP + VXLAN + Ethernet header */ |
262 | #define VXLAN_HEADROOM (20 + 8 + 8 + 14) | |
263 | /* IPv6 header + UDP + VXLAN + Ethernet header */ | |
264 | #define VXLAN6_HEADROOM (40 + 8 + 8 + 14) | |
265 | ||
d4ac05ff JB |
266 | static inline struct vxlanhdr *vxlan_hdr(struct sk_buff *skb) |
267 | { | |
268 | return (struct vxlanhdr *)(udp_hdr(skb) + 1); | |
269 | } | |
270 | ||
54bfd872 JB |
271 | static inline __be32 vxlan_vni(__be32 vni_field) |
272 | { | |
273 | #if defined(__BIG_ENDIAN) | |
5692d7ea | 274 | return (__force __be32)((__force u32)vni_field >> 8); |
54bfd872 | 275 | #else |
5692d7ea | 276 | return (__force __be32)((__force u32)(vni_field & VXLAN_VNI_MASK) << 8); |
54bfd872 JB |
277 | #endif |
278 | } | |
279 | ||
280 | static inline __be32 vxlan_vni_field(__be32 vni) | |
281 | { | |
282 | #if defined(__BIG_ENDIAN) | |
5692d7ea | 283 | return (__force __be32)((__force u32)vni << 8); |
54bfd872 | 284 | #else |
5692d7ea | 285 | return (__force __be32)((__force u32)vni >> 8); |
54bfd872 JB |
286 | #endif |
287 | } | |
288 | ||
289 | static inline __be32 vxlan_tun_id_to_vni(__be64 tun_id) | |
290 | { | |
291 | #if defined(__BIG_ENDIAN) | |
5692d7ea | 292 | return (__force __be32)tun_id; |
54bfd872 | 293 | #else |
5692d7ea | 294 | return (__force __be32)((__force u64)tun_id >> 32); |
54bfd872 JB |
295 | #endif |
296 | } | |
297 | ||
07dabf20 JB |
298 | static inline __be64 vxlan_vni_to_tun_id(__be32 vni) |
299 | { | |
300 | #if defined(__BIG_ENDIAN) | |
5692d7ea | 301 | return (__force __be64)vni; |
07dabf20 | 302 | #else |
5692d7ea | 303 | return (__force __be64)((u64)(__force u32)vni << 32); |
07dabf20 JB |
304 | #endif |
305 | } | |
306 | ||
54bfd872 JB |
307 | static inline size_t vxlan_rco_start(__be32 vni_field) |
308 | { | |
309 | return be32_to_cpu(vni_field & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT; | |
310 | } | |
311 | ||
312 | static inline size_t vxlan_rco_offset(__be32 vni_field) | |
313 | { | |
314 | return (vni_field & VXLAN_RCO_UDP) ? | |
315 | offsetof(struct udphdr, check) : | |
316 | offsetof(struct tcphdr, check); | |
317 | } | |
318 | ||
319 | static inline __be32 vxlan_compute_rco(unsigned int start, unsigned int offset) | |
320 | { | |
321 | __be32 vni_field = cpu_to_be32(start >> VXLAN_RCO_SHIFT); | |
322 | ||
323 | if (offset == offsetof(struct udphdr, check)) | |
324 | vni_field |= VXLAN_RCO_UDP; | |
325 | return vni_field; | |
326 | } | |
327 | ||
e6cd988c | 328 | #if IS_ENABLED(CONFIG_VXLAN) |
53cf5275 | 329 | void vxlan_get_rx_port(struct net_device *netdev); |
e6cd988c JG |
330 | #else |
331 | static inline void vxlan_get_rx_port(struct net_device *netdev) | |
332 | { | |
333 | } | |
334 | #endif | |
705cc62f JB |
335 | |
336 | static inline unsigned short vxlan_get_sk_family(struct vxlan_sock *vs) | |
337 | { | |
338 | return vs->sock->sk->sk_family; | |
339 | } | |
48e92c44 JB |
340 | |
341 | #endif |