1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef __NET_VXLAN_H 3 #define __NET_VXLAN_H 1 4 5 #include <linux/if_vlan.h> 6 #include <net/udp_tunnel.h> 7 #include <net/dst_metadata.h> 8 #include <net/rtnetlink.h> 9 #include <net/switchdev.h> 10 #include <net/nexthop.h> 11 12 #define IANA_VXLAN_UDP_PORT 4789 13 #define IANA_VXLAN_GPE_UDP_PORT 4790 14 15 /* VXLAN protocol (RFC 7348) header: 16 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 17 * |R|R|R|R|I|R|R|R| Reserved | 18 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 19 * | VXLAN Network Identifier (VNI) | Reserved | 20 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 21 * 22 * I = VXLAN Network Identifier (VNI) present. 23 */ 24 struct vxlanhdr { 25 __be32 vx_flags; 26 __be32 vx_vni; 27 }; 28 29 /* VXLAN header flags. */ 30 #define VXLAN_HF_VNI cpu_to_be32(BIT(27)) 31 32 #define VXLAN_N_VID (1u << 24) 33 #define VXLAN_VID_MASK (VXLAN_N_VID - 1) 34 #define VXLAN_VNI_MASK cpu_to_be32(VXLAN_VID_MASK << 8) 35 #define VXLAN_HLEN (sizeof(struct udphdr) + sizeof(struct vxlanhdr)) 36 37 #define VNI_HASH_BITS 10 38 #define VNI_HASH_SIZE (1<<VNI_HASH_BITS) 39 #define FDB_HASH_BITS 8 40 #define FDB_HASH_SIZE (1<<FDB_HASH_BITS) 41 42 /* Remote checksum offload for VXLAN (VXLAN_F_REMCSUM_[RT]X): 43 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 44 * |R|R|R|R|I|R|R|R|R|R|C| Reserved | 45 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 46 * | VXLAN Network Identifier (VNI) |O| Csum start | 47 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 48 * 49 * C = Remote checksum offload bit. When set indicates that the 50 * remote checksum offload data is present. 51 * 52 * O = Offset bit. Indicates the checksum offset relative to 53 * checksum start. 54 * 55 * Csum start = Checksum start divided by two. 56 * 57 * http://tools.ietf.org/html/draft-herbert-vxlan-rco 58 */ 59 60 /* VXLAN-RCO header flags. */ 61 #define VXLAN_HF_RCO cpu_to_be32(BIT(21)) 62 63 /* Remote checksum offload header option */ 64 #define VXLAN_RCO_MASK cpu_to_be32(0x7f) /* Last byte of vni field */ 65 #define VXLAN_RCO_UDP cpu_to_be32(0x80) /* Indicate UDP RCO (TCP when not set *) */ 66 #define VXLAN_RCO_SHIFT 1 /* Left shift of start */ 67 #define VXLAN_RCO_SHIFT_MASK ((1 << VXLAN_RCO_SHIFT) - 1) 68 #define VXLAN_MAX_REMCSUM_START (0x7f << VXLAN_RCO_SHIFT) 69 70 /* 71 * VXLAN Group Based Policy Extension (VXLAN_F_GBP): 72 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 73 * |G|R|R|R|I|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID | 74 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 75 * | VXLAN Network Identifier (VNI) | Reserved | 76 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 77 * 78 * G = Group Policy ID present. 79 * 80 * D = Don't Learn bit. When set, this bit indicates that the egress 81 * VTEP MUST NOT learn the source address of the encapsulated frame. 82 * 83 * A = Indicates that the group policy has already been applied to 84 * this packet. Policies MUST NOT be applied by devices when the 85 * A bit is set. 86 * 87 * https://tools.ietf.org/html/draft-smith-vxlan-group-policy 88 */ 89 struct vxlanhdr_gbp { 90 u8 vx_flags; 91 #ifdef __LITTLE_ENDIAN_BITFIELD 92 u8 reserved_flags1:3, 93 policy_applied:1, 94 reserved_flags2:2, 95 dont_learn:1, 96 reserved_flags3:1; 97 #elif defined(__BIG_ENDIAN_BITFIELD) 98 u8 reserved_flags1:1, 99 dont_learn:1, 100 reserved_flags2:2, 101 policy_applied:1, 102 reserved_flags3:3; 103 #else 104 #error "Please fix <asm/byteorder.h>" 105 #endif 106 __be16 policy_id; 107 __be32 vx_vni; 108 }; 109 110 /* VXLAN-GBP header flags. */ 111 #define VXLAN_HF_GBP cpu_to_be32(BIT(31)) 112 113 #define VXLAN_GBP_USED_BITS (VXLAN_HF_GBP | cpu_to_be32(0xFFFFFF)) 114 115 /* skb->mark mapping 116 * 117 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 118 * |R|R|R|R|R|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID | 119 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 120 */ 121 #define VXLAN_GBP_DONT_LEARN (BIT(6) << 16) 122 #define VXLAN_GBP_POLICY_APPLIED (BIT(3) << 16) 123 #define VXLAN_GBP_ID_MASK (0xFFFF) 124 125 #define VXLAN_GBP_MASK (VXLAN_GBP_DONT_LEARN | VXLAN_GBP_POLICY_APPLIED | \ 126 VXLAN_GBP_ID_MASK) 127 128 /* 129 * VXLAN Generic Protocol Extension (VXLAN_F_GPE): 130 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 131 * |R|R|Ver|I|P|R|O| Reserved |Next Protocol | 132 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 133 * | VXLAN Network Identifier (VNI) | Reserved | 134 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ 135 * 136 * Ver = Version. Indicates VXLAN GPE protocol version. 137 * 138 * P = Next Protocol Bit. The P bit is set to indicate that the 139 * Next Protocol field is present. 140 * 141 * O = OAM Flag Bit. The O bit is set to indicate that the packet 142 * is an OAM packet. 143 * 144 * Next Protocol = This 8 bit field indicates the protocol header 145 * immediately following the VXLAN GPE header. 146 * 147 * https://tools.ietf.org/html/draft-ietf-nvo3-vxlan-gpe-01 148 */ 149 150 struct vxlanhdr_gpe { 151 #if defined(__LITTLE_ENDIAN_BITFIELD) 152 u8 oam_flag:1, 153 reserved_flags1:1, 154 np_applied:1, 155 instance_applied:1, 156 version:2, 157 reserved_flags2:2; 158 #elif defined(__BIG_ENDIAN_BITFIELD) 159 u8 reserved_flags2:2, 160 version:2, 161 instance_applied:1, 162 np_applied:1, 163 reserved_flags1:1, 164 oam_flag:1; 165 #endif 166 u8 reserved_flags3; 167 u8 reserved_flags4; 168 u8 next_protocol; 169 __be32 vx_vni; 170 }; 171 172 /* VXLAN-GPE header flags. */ 173 #define VXLAN_HF_VER cpu_to_be32(BIT(29) | BIT(28)) 174 #define VXLAN_HF_NP cpu_to_be32(BIT(26)) 175 #define VXLAN_HF_OAM cpu_to_be32(BIT(24)) 176 177 #define VXLAN_GPE_USED_BITS (VXLAN_HF_VER | VXLAN_HF_NP | VXLAN_HF_OAM | \ 178 cpu_to_be32(0xff)) 179 180 struct vxlan_metadata { 181 u32 gbp; 182 }; 183 184 /* per UDP socket information */ 185 struct vxlan_sock { 186 struct hlist_node hlist; 187 struct socket *sock; 188 struct hlist_head vni_list[VNI_HASH_SIZE]; 189 refcount_t refcnt; 190 u32 flags; 191 }; 192 193 union vxlan_addr { 194 struct sockaddr_in sin; 195 struct sockaddr_in6 sin6; 196 struct sockaddr sa; 197 }; 198 199 struct vxlan_rdst { 200 union vxlan_addr remote_ip; 201 __be16 remote_port; 202 u8 offloaded:1; 203 __be32 remote_vni; 204 u32 remote_ifindex; 205 struct net_device *remote_dev; 206 struct list_head list; 207 struct rcu_head rcu; 208 struct dst_cache dst_cache; 209 }; 210 211 struct vxlan_config { 212 union vxlan_addr remote_ip; 213 union vxlan_addr saddr; 214 __be32 vni; 215 int remote_ifindex; 216 int mtu; 217 __be16 dst_port; 218 u16 port_min; 219 u16 port_max; 220 u8 tos; 221 u8 ttl; 222 __be32 label; 223 u32 flags; 224 unsigned long age_interval; 225 unsigned int addrmax; 226 bool no_share; 227 enum ifla_vxlan_df df; 228 }; 229 230 enum { 231 VXLAN_VNI_STATS_RX, 232 VXLAN_VNI_STATS_RX_DROPS, 233 VXLAN_VNI_STATS_RX_ERRORS, 234 VXLAN_VNI_STATS_TX, 235 VXLAN_VNI_STATS_TX_DROPS, 236 VXLAN_VNI_STATS_TX_ERRORS, 237 }; 238 239 struct vxlan_vni_stats { 240 u64 rx_packets; 241 u64 rx_bytes; 242 u64 rx_drops; 243 u64 rx_errors; 244 u64 tx_packets; 245 u64 tx_bytes; 246 u64 tx_drops; 247 u64 tx_errors; 248 }; 249 250 struct vxlan_vni_stats_pcpu { 251 struct vxlan_vni_stats stats; 252 struct u64_stats_sync syncp; 253 }; 254 255 struct vxlan_dev_node { 256 struct hlist_node hlist; 257 struct vxlan_dev *vxlan; 258 }; 259 260 struct vxlan_vni_node { 261 struct rhash_head vnode; 262 struct vxlan_dev_node hlist4; /* vni hash table for IPv4 socket */ 263 #if IS_ENABLED(CONFIG_IPV6) 264 struct vxlan_dev_node hlist6; /* vni hash table for IPv6 socket */ 265 #endif 266 struct list_head vlist; 267 __be32 vni; 268 union vxlan_addr remote_ip; /* default remote ip for this vni */ 269 struct vxlan_vni_stats_pcpu __percpu *stats; 270 271 struct rcu_head rcu; 272 }; 273 274 struct vxlan_vni_group { 275 struct rhashtable vni_hash; 276 struct list_head vni_list; 277 u32 num_vnis; 278 }; 279 280 /* Pseudo network device */ 281 struct vxlan_dev { 282 struct vxlan_dev_node hlist4; /* vni hash table for IPv4 socket */ 283 #if IS_ENABLED(CONFIG_IPV6) 284 struct vxlan_dev_node hlist6; /* vni hash table for IPv6 socket */ 285 #endif 286 struct list_head next; /* vxlan's per namespace list */ 287 struct vxlan_sock __rcu *vn4_sock; /* listening socket for IPv4 */ 288 #if IS_ENABLED(CONFIG_IPV6) 289 struct vxlan_sock __rcu *vn6_sock; /* listening socket for IPv6 */ 290 #endif 291 struct net_device *dev; 292 struct net *net; /* netns for packet i/o */ 293 struct vxlan_rdst default_dst; /* default destination */ 294 295 struct timer_list age_timer; 296 spinlock_t hash_lock[FDB_HASH_SIZE]; 297 unsigned int addrcnt; 298 struct gro_cells gro_cells; 299 300 struct vxlan_config cfg; 301 302 struct vxlan_vni_group __rcu *vnigrp; 303 304 struct hlist_head fdb_head[FDB_HASH_SIZE]; 305 }; 306 307 #define VXLAN_F_LEARN 0x01 308 #define VXLAN_F_PROXY 0x02 309 #define VXLAN_F_RSC 0x04 310 #define VXLAN_F_L2MISS 0x08 311 #define VXLAN_F_L3MISS 0x10 312 #define VXLAN_F_IPV6 0x20 313 #define VXLAN_F_UDP_ZERO_CSUM_TX 0x40 314 #define VXLAN_F_UDP_ZERO_CSUM6_TX 0x80 315 #define VXLAN_F_UDP_ZERO_CSUM6_RX 0x100 316 #define VXLAN_F_REMCSUM_TX 0x200 317 #define VXLAN_F_REMCSUM_RX 0x400 318 #define VXLAN_F_GBP 0x800 319 #define VXLAN_F_REMCSUM_NOPARTIAL 0x1000 320 #define VXLAN_F_COLLECT_METADATA 0x2000 321 #define VXLAN_F_GPE 0x4000 322 #define VXLAN_F_IPV6_LINKLOCAL 0x8000 323 #define VXLAN_F_TTL_INHERIT 0x10000 324 #define VXLAN_F_VNIFILTER 0x20000 325 326 /* Flags that are used in the receive path. These flags must match in 327 * order for a socket to be shareable 328 */ 329 #define VXLAN_F_RCV_FLAGS (VXLAN_F_GBP | \ 330 VXLAN_F_GPE | \ 331 VXLAN_F_UDP_ZERO_CSUM6_RX | \ 332 VXLAN_F_REMCSUM_RX | \ 333 VXLAN_F_REMCSUM_NOPARTIAL | \ 334 VXLAN_F_COLLECT_METADATA | \ 335 VXLAN_F_VNIFILTER) 336 337 /* Flags that can be set together with VXLAN_F_GPE. */ 338 #define VXLAN_F_ALLOWED_GPE (VXLAN_F_GPE | \ 339 VXLAN_F_IPV6 | \ 340 VXLAN_F_IPV6_LINKLOCAL | \ 341 VXLAN_F_UDP_ZERO_CSUM_TX | \ 342 VXLAN_F_UDP_ZERO_CSUM6_TX | \ 343 VXLAN_F_UDP_ZERO_CSUM6_RX | \ 344 VXLAN_F_COLLECT_METADATA | \ 345 VXLAN_F_VNIFILTER) 346 347 struct net_device *vxlan_dev_create(struct net *net, const char *name, 348 u8 name_assign_type, struct vxlan_config *conf); 349 350 static inline netdev_features_t vxlan_features_check(struct sk_buff *skb, 351 netdev_features_t features) 352 { 353 u8 l4_hdr = 0; 354 355 if (!skb->encapsulation) 356 return features; 357 358 switch (vlan_get_protocol(skb)) { 359 case htons(ETH_P_IP): 360 l4_hdr = ip_hdr(skb)->protocol; 361 break; 362 case htons(ETH_P_IPV6): 363 l4_hdr = ipv6_hdr(skb)->nexthdr; 364 break; 365 default: 366 return features; 367 } 368 369 if ((l4_hdr == IPPROTO_UDP) && 370 (skb->inner_protocol_type != ENCAP_TYPE_ETHER || 371 skb->inner_protocol != htons(ETH_P_TEB) || 372 (skb_inner_mac_header(skb) - skb_transport_header(skb) != 373 sizeof(struct udphdr) + sizeof(struct vxlanhdr)) || 374 (skb->ip_summed != CHECKSUM_NONE && 375 !can_checksum_protocol(features, inner_eth_hdr(skb)->h_proto)))) 376 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 377 378 return features; 379 } 380 381 /* IP header + UDP + VXLAN + Ethernet header */ 382 #define VXLAN_HEADROOM (20 + 8 + 8 + 14) 383 /* IPv6 header + UDP + VXLAN + Ethernet header */ 384 #define VXLAN6_HEADROOM (40 + 8 + 8 + 14) 385 386 static inline struct vxlanhdr *vxlan_hdr(struct sk_buff *skb) 387 { 388 return (struct vxlanhdr *)(udp_hdr(skb) + 1); 389 } 390 391 static inline __be32 vxlan_vni(__be32 vni_field) 392 { 393 #if defined(__BIG_ENDIAN) 394 return (__force __be32)((__force u32)vni_field >> 8); 395 #else 396 return (__force __be32)((__force u32)(vni_field & VXLAN_VNI_MASK) << 8); 397 #endif 398 } 399 400 static inline __be32 vxlan_vni_field(__be32 vni) 401 { 402 #if defined(__BIG_ENDIAN) 403 return (__force __be32)((__force u32)vni << 8); 404 #else 405 return (__force __be32)((__force u32)vni >> 8); 406 #endif 407 } 408 409 static inline size_t vxlan_rco_start(__be32 vni_field) 410 { 411 return be32_to_cpu(vni_field & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT; 412 } 413 414 static inline size_t vxlan_rco_offset(__be32 vni_field) 415 { 416 return (vni_field & VXLAN_RCO_UDP) ? 417 offsetof(struct udphdr, check) : 418 offsetof(struct tcphdr, check); 419 } 420 421 static inline __be32 vxlan_compute_rco(unsigned int start, unsigned int offset) 422 { 423 __be32 vni_field = cpu_to_be32(start >> VXLAN_RCO_SHIFT); 424 425 if (offset == offsetof(struct udphdr, check)) 426 vni_field |= VXLAN_RCO_UDP; 427 return vni_field; 428 } 429 430 static inline unsigned short vxlan_get_sk_family(struct vxlan_sock *vs) 431 { 432 return vs->sock->sk->sk_family; 433 } 434 435 #if IS_ENABLED(CONFIG_IPV6) 436 437 static inline bool vxlan_addr_any(const union vxlan_addr *ipa) 438 { 439 if (ipa->sa.sa_family == AF_INET6) 440 return ipv6_addr_any(&ipa->sin6.sin6_addr); 441 else 442 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY); 443 } 444 445 static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa) 446 { 447 if (ipa->sa.sa_family == AF_INET6) 448 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr); 449 else 450 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr); 451 } 452 453 #else /* !IS_ENABLED(CONFIG_IPV6) */ 454 455 static inline bool vxlan_addr_any(const union vxlan_addr *ipa) 456 { 457 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY); 458 } 459 460 static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa) 461 { 462 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr); 463 } 464 465 #endif /* IS_ENABLED(CONFIG_IPV6) */ 466 467 static inline bool netif_is_vxlan(const struct net_device *dev) 468 { 469 return dev->rtnl_link_ops && 470 !strcmp(dev->rtnl_link_ops->kind, "vxlan"); 471 } 472 473 struct switchdev_notifier_vxlan_fdb_info { 474 struct switchdev_notifier_info info; /* must be first */ 475 union vxlan_addr remote_ip; 476 __be16 remote_port; 477 __be32 remote_vni; 478 u32 remote_ifindex; 479 u8 eth_addr[ETH_ALEN]; 480 __be32 vni; 481 bool offloaded; 482 bool added_by_user; 483 }; 484 485 #if IS_ENABLED(CONFIG_VXLAN) 486 int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni, 487 struct switchdev_notifier_vxlan_fdb_info *fdb_info); 488 int vxlan_fdb_replay(const struct net_device *dev, __be32 vni, 489 struct notifier_block *nb, 490 struct netlink_ext_ack *extack); 491 void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni); 492 493 #else 494 static inline int 495 vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni, 496 struct switchdev_notifier_vxlan_fdb_info *fdb_info) 497 { 498 return -ENOENT; 499 } 500 501 static inline int vxlan_fdb_replay(const struct net_device *dev, __be32 vni, 502 struct notifier_block *nb, 503 struct netlink_ext_ack *extack) 504 { 505 return -EOPNOTSUPP; 506 } 507 508 static inline void 509 vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni) 510 { 511 } 512 #endif 513 514 static inline void vxlan_flag_attr_error(int attrtype, 515 struct netlink_ext_ack *extack) 516 { 517 #define VXLAN_FLAG(flg) \ 518 case IFLA_VXLAN_##flg: \ 519 NL_SET_ERR_MSG_MOD(extack, \ 520 "cannot change " #flg " flag"); \ 521 break 522 switch (attrtype) { 523 VXLAN_FLAG(TTL_INHERIT); 524 VXLAN_FLAG(LEARNING); 525 VXLAN_FLAG(PROXY); 526 VXLAN_FLAG(RSC); 527 VXLAN_FLAG(L2MISS); 528 VXLAN_FLAG(L3MISS); 529 VXLAN_FLAG(COLLECT_METADATA); 530 VXLAN_FLAG(UDP_ZERO_CSUM6_TX); 531 VXLAN_FLAG(UDP_ZERO_CSUM6_RX); 532 VXLAN_FLAG(REMCSUM_TX); 533 VXLAN_FLAG(REMCSUM_RX); 534 VXLAN_FLAG(GBP); 535 VXLAN_FLAG(GPE); 536 VXLAN_FLAG(REMCSUM_NOPARTIAL); 537 default: 538 NL_SET_ERR_MSG_MOD(extack, \ 539 "cannot change flag"); 540 break; 541 } 542 #undef VXLAN_FLAG 543 } 544 545 static inline bool vxlan_fdb_nh_path_select(struct nexthop *nh, 546 int hash, 547 struct vxlan_rdst *rdst) 548 { 549 struct fib_nh_common *nhc; 550 551 nhc = nexthop_path_fdb_result(nh, hash); 552 if (unlikely(!nhc)) 553 return false; 554 555 switch (nhc->nhc_gw_family) { 556 case AF_INET: 557 rdst->remote_ip.sin.sin_addr.s_addr = nhc->nhc_gw.ipv4; 558 rdst->remote_ip.sa.sa_family = AF_INET; 559 break; 560 case AF_INET6: 561 rdst->remote_ip.sin6.sin6_addr = nhc->nhc_gw.ipv6; 562 rdst->remote_ip.sa.sa_family = AF_INET6; 563 break; 564 } 565 566 return true; 567 } 568 569 #endif 570