1 /* SPDX-License-Identifier: GPL-2.0 */
3 #define __NET_VXLAN_H 1
5 #include <linux/if_vlan.h>
6 #include <linux/rhashtable-types.h>
7 #include <net/udp_tunnel.h>
8 #include <net/dst_metadata.h>
9 #include <net/rtnetlink.h>
10 #include <net/switchdev.h>
11 #include <net/nexthop.h>
13 #define IANA_VXLAN_UDP_PORT 4789
14 #define IANA_VXLAN_GPE_UDP_PORT 4790
16 /* VXLAN protocol (RFC 7348) header:
17 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
18 * |R|R|R|R|I|R|R|R| Reserved |
19 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
20 * | VXLAN Network Identifier (VNI) | Reserved |
21 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
23 * I = VXLAN Network Identifier (VNI) present.
30 /* VXLAN header flags. */
31 #define VXLAN_HF_VNI cpu_to_be32(BIT(27))
33 #define VXLAN_N_VID (1u << 24)
34 #define VXLAN_VID_MASK (VXLAN_N_VID - 1)
35 #define VXLAN_VNI_MASK cpu_to_be32(VXLAN_VID_MASK << 8)
36 #define VXLAN_HLEN (sizeof(struct udphdr) + sizeof(struct vxlanhdr))
38 #define VNI_HASH_BITS 10
39 #define VNI_HASH_SIZE (1<<VNI_HASH_BITS)
40 #define FDB_HASH_BITS 8
41 #define FDB_HASH_SIZE (1<<FDB_HASH_BITS)
43 /* Remote checksum offload for VXLAN (VXLAN_F_REMCSUM_[RT]X):
44 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
45 * |R|R|R|R|I|R|R|R|R|R|C| Reserved |
46 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
47 * | VXLAN Network Identifier (VNI) |O| Csum start |
48 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
50 * C = Remote checksum offload bit. When set indicates that the
51 * remote checksum offload data is present.
53 * O = Offset bit. Indicates the checksum offset relative to
56 * Csum start = Checksum start divided by two.
58 * http://tools.ietf.org/html/draft-herbert-vxlan-rco
61 /* VXLAN-RCO header flags. */
62 #define VXLAN_HF_RCO cpu_to_be32(BIT(21))
64 /* Remote checksum offload header option */
65 #define VXLAN_RCO_MASK cpu_to_be32(0x7f) /* Last byte of vni field */
66 #define VXLAN_RCO_UDP cpu_to_be32(0x80) /* Indicate UDP RCO (TCP when not set *) */
67 #define VXLAN_RCO_SHIFT 1 /* Left shift of start */
68 #define VXLAN_RCO_SHIFT_MASK ((1 << VXLAN_RCO_SHIFT) - 1)
69 #define VXLAN_MAX_REMCSUM_START (0x7f << VXLAN_RCO_SHIFT)
72 * VXLAN Group Based Policy Extension (VXLAN_F_GBP):
73 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
74 * |G|R|R|R|I|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID |
75 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
76 * | VXLAN Network Identifier (VNI) | Reserved |
77 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
79 * G = Group Policy ID present.
81 * D = Don't Learn bit. When set, this bit indicates that the egress
82 * VTEP MUST NOT learn the source address of the encapsulated frame.
84 * A = Indicates that the group policy has already been applied to
85 * this packet. Policies MUST NOT be applied by devices when the
88 * https://tools.ietf.org/html/draft-smith-vxlan-group-policy
92 #ifdef __LITTLE_ENDIAN_BITFIELD
98 #elif defined(__BIG_ENDIAN_BITFIELD)
105 #error "Please fix <asm/byteorder.h>"
111 /* VXLAN-GBP header flags. */
112 #define VXLAN_HF_GBP cpu_to_be32(BIT(31))
114 #define VXLAN_GBP_USED_BITS (VXLAN_HF_GBP | cpu_to_be32(0xFFFFFF))
118 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
119 * |R|R|R|R|R|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID |
120 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
122 #define VXLAN_GBP_DONT_LEARN (BIT(6) << 16)
123 #define VXLAN_GBP_POLICY_APPLIED (BIT(3) << 16)
124 #define VXLAN_GBP_ID_MASK (0xFFFF)
126 #define VXLAN_GBP_MASK (VXLAN_GBP_DONT_LEARN | VXLAN_GBP_POLICY_APPLIED | \
130 * VXLAN Generic Protocol Extension (VXLAN_F_GPE):
131 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
132 * |R|R|Ver|I|P|R|O| Reserved |Next Protocol |
133 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
134 * | VXLAN Network Identifier (VNI) | Reserved |
135 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
137 * Ver = Version. Indicates VXLAN GPE protocol version.
139 * P = Next Protocol Bit. The P bit is set to indicate that the
140 * Next Protocol field is present.
142 * O = OAM Flag Bit. The O bit is set to indicate that the packet
145 * Next Protocol = This 8 bit field indicates the protocol header
146 * immediately following the VXLAN GPE header.
148 * https://tools.ietf.org/html/draft-ietf-nvo3-vxlan-gpe-01
151 struct vxlanhdr_gpe {
152 #if defined(__LITTLE_ENDIAN_BITFIELD)
159 #elif defined(__BIG_ENDIAN_BITFIELD)
160 u8 reserved_flags2:2,
173 /* VXLAN-GPE header flags. */
174 #define VXLAN_HF_VER cpu_to_be32(BIT(29) | BIT(28))
175 #define VXLAN_HF_NP cpu_to_be32(BIT(26))
176 #define VXLAN_HF_OAM cpu_to_be32(BIT(24))
178 #define VXLAN_GPE_USED_BITS (VXLAN_HF_VER | VXLAN_HF_NP | VXLAN_HF_OAM | \
181 struct vxlan_metadata {
185 /* per UDP socket information */
187 struct hlist_node hlist;
189 struct hlist_head vni_list[VNI_HASH_SIZE];
195 struct sockaddr_in sin;
196 struct sockaddr_in6 sin6;
201 union vxlan_addr remote_ip;
206 struct net_device *remote_dev;
207 struct list_head list;
209 struct dst_cache dst_cache;
212 struct vxlan_config {
213 union vxlan_addr remote_ip;
214 union vxlan_addr saddr;
225 unsigned long age_interval;
226 unsigned int addrmax;
228 enum ifla_vxlan_df df;
233 VXLAN_VNI_STATS_RX_DROPS,
234 VXLAN_VNI_STATS_RX_ERRORS,
236 VXLAN_VNI_STATS_TX_DROPS,
237 VXLAN_VNI_STATS_TX_ERRORS,
240 struct vxlan_vni_stats {
251 struct vxlan_vni_stats_pcpu {
252 struct vxlan_vni_stats stats;
253 struct u64_stats_sync syncp;
256 struct vxlan_dev_node {
257 struct hlist_node hlist;
258 struct vxlan_dev *vxlan;
261 struct vxlan_vni_node {
262 struct rhash_head vnode;
263 struct vxlan_dev_node hlist4; /* vni hash table for IPv4 socket */
264 #if IS_ENABLED(CONFIG_IPV6)
265 struct vxlan_dev_node hlist6; /* vni hash table for IPv6 socket */
267 struct list_head vlist;
269 union vxlan_addr remote_ip; /* default remote ip for this vni */
270 struct vxlan_vni_stats_pcpu __percpu *stats;
275 struct vxlan_vni_group {
276 struct rhashtable vni_hash;
277 struct list_head vni_list;
281 /* Pseudo network device */
283 struct vxlan_dev_node hlist4; /* vni hash table for IPv4 socket */
284 #if IS_ENABLED(CONFIG_IPV6)
285 struct vxlan_dev_node hlist6; /* vni hash table for IPv6 socket */
287 struct list_head next; /* vxlan's per namespace list */
288 struct vxlan_sock __rcu *vn4_sock; /* listening socket for IPv4 */
289 #if IS_ENABLED(CONFIG_IPV6)
290 struct vxlan_sock __rcu *vn6_sock; /* listening socket for IPv6 */
292 struct net_device *dev;
293 struct net *net; /* netns for packet i/o */
294 struct vxlan_rdst default_dst; /* default destination */
296 struct timer_list age_timer;
297 spinlock_t hash_lock[FDB_HASH_SIZE];
298 unsigned int addrcnt;
299 struct gro_cells gro_cells;
301 struct vxlan_config cfg;
303 struct vxlan_vni_group __rcu *vnigrp;
305 struct hlist_head fdb_head[FDB_HASH_SIZE];
307 struct rhashtable mdb_tbl;
308 struct hlist_head mdb_list;
309 unsigned int mdb_seq;
312 #define VXLAN_F_LEARN 0x01
313 #define VXLAN_F_PROXY 0x02
314 #define VXLAN_F_RSC 0x04
315 #define VXLAN_F_L2MISS 0x08
316 #define VXLAN_F_L3MISS 0x10
317 #define VXLAN_F_IPV6 0x20
318 #define VXLAN_F_UDP_ZERO_CSUM_TX 0x40
319 #define VXLAN_F_UDP_ZERO_CSUM6_TX 0x80
320 #define VXLAN_F_UDP_ZERO_CSUM6_RX 0x100
321 #define VXLAN_F_REMCSUM_TX 0x200
322 #define VXLAN_F_REMCSUM_RX 0x400
323 #define VXLAN_F_GBP 0x800
324 #define VXLAN_F_REMCSUM_NOPARTIAL 0x1000
325 #define VXLAN_F_COLLECT_METADATA 0x2000
326 #define VXLAN_F_GPE 0x4000
327 #define VXLAN_F_IPV6_LINKLOCAL 0x8000
328 #define VXLAN_F_TTL_INHERIT 0x10000
329 #define VXLAN_F_VNIFILTER 0x20000
330 #define VXLAN_F_MDB 0x40000
331 #define VXLAN_F_LOCALBYPASS 0x80000
333 /* Flags that are used in the receive path. These flags must match in
334 * order for a socket to be shareable
336 #define VXLAN_F_RCV_FLAGS (VXLAN_F_GBP | \
338 VXLAN_F_UDP_ZERO_CSUM6_RX | \
339 VXLAN_F_REMCSUM_RX | \
340 VXLAN_F_REMCSUM_NOPARTIAL | \
341 VXLAN_F_COLLECT_METADATA | \
344 /* Flags that can be set together with VXLAN_F_GPE. */
345 #define VXLAN_F_ALLOWED_GPE (VXLAN_F_GPE | \
347 VXLAN_F_IPV6_LINKLOCAL | \
348 VXLAN_F_UDP_ZERO_CSUM_TX | \
349 VXLAN_F_UDP_ZERO_CSUM6_TX | \
350 VXLAN_F_UDP_ZERO_CSUM6_RX | \
351 VXLAN_F_COLLECT_METADATA | \
352 VXLAN_F_VNIFILTER | \
355 struct net_device *vxlan_dev_create(struct net *net, const char *name,
356 u8 name_assign_type, struct vxlan_config *conf);
358 static inline netdev_features_t vxlan_features_check(struct sk_buff *skb,
359 netdev_features_t features)
363 if (!skb->encapsulation)
366 switch (vlan_get_protocol(skb)) {
367 case htons(ETH_P_IP):
368 l4_hdr = ip_hdr(skb)->protocol;
370 case htons(ETH_P_IPV6):
371 l4_hdr = ipv6_hdr(skb)->nexthdr;
377 if ((l4_hdr == IPPROTO_UDP) &&
378 (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
379 skb->inner_protocol != htons(ETH_P_TEB) ||
380 (skb_inner_mac_header(skb) - skb_transport_header(skb) !=
381 sizeof(struct udphdr) + sizeof(struct vxlanhdr)) ||
382 (skb->ip_summed != CHECKSUM_NONE &&
383 !can_checksum_protocol(features, inner_eth_hdr(skb)->h_proto))))
384 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
389 static inline int vxlan_headroom(u32 flags)
391 /* VXLAN: IP4/6 header + UDP + VXLAN + Ethernet header */
392 /* VXLAN-GPE: IP4/6 header + UDP + VXLAN */
393 return (flags & VXLAN_F_IPV6 ? sizeof(struct ipv6hdr) :
394 sizeof(struct iphdr)) +
395 sizeof(struct udphdr) + sizeof(struct vxlanhdr) +
396 (flags & VXLAN_F_GPE ? 0 : ETH_HLEN);
399 static inline struct vxlanhdr *vxlan_hdr(struct sk_buff *skb)
401 return (struct vxlanhdr *)(udp_hdr(skb) + 1);
404 static inline __be32 vxlan_vni(__be32 vni_field)
406 #if defined(__BIG_ENDIAN)
407 return (__force __be32)((__force u32)vni_field >> 8);
409 return (__force __be32)((__force u32)(vni_field & VXLAN_VNI_MASK) << 8);
413 static inline __be32 vxlan_vni_field(__be32 vni)
415 #if defined(__BIG_ENDIAN)
416 return (__force __be32)((__force u32)vni << 8);
418 return (__force __be32)((__force u32)vni >> 8);
422 static inline size_t vxlan_rco_start(__be32 vni_field)
424 return be32_to_cpu(vni_field & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT;
427 static inline size_t vxlan_rco_offset(__be32 vni_field)
429 return (vni_field & VXLAN_RCO_UDP) ?
430 offsetof(struct udphdr, check) :
431 offsetof(struct tcphdr, check);
434 static inline __be32 vxlan_compute_rco(unsigned int start, unsigned int offset)
436 __be32 vni_field = cpu_to_be32(start >> VXLAN_RCO_SHIFT);
438 if (offset == offsetof(struct udphdr, check))
439 vni_field |= VXLAN_RCO_UDP;
443 static inline unsigned short vxlan_get_sk_family(struct vxlan_sock *vs)
445 return vs->sock->sk->sk_family;
448 #if IS_ENABLED(CONFIG_IPV6)
450 static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
452 if (ipa->sa.sa_family == AF_INET6)
453 return ipv6_addr_any(&ipa->sin6.sin6_addr);
455 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
458 static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
460 if (ipa->sa.sa_family == AF_INET6)
461 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
463 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr);
466 #else /* !IS_ENABLED(CONFIG_IPV6) */
468 static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
470 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
473 static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
475 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr);
478 #endif /* IS_ENABLED(CONFIG_IPV6) */
480 static inline bool netif_is_vxlan(const struct net_device *dev)
482 return dev->rtnl_link_ops &&
483 !strcmp(dev->rtnl_link_ops->kind, "vxlan");
486 struct switchdev_notifier_vxlan_fdb_info {
487 struct switchdev_notifier_info info; /* must be first */
488 union vxlan_addr remote_ip;
492 u8 eth_addr[ETH_ALEN];
498 #if IS_ENABLED(CONFIG_VXLAN)
499 int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
500 struct switchdev_notifier_vxlan_fdb_info *fdb_info);
501 int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
502 struct notifier_block *nb,
503 struct netlink_ext_ack *extack);
504 void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni);
508 vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
509 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
514 static inline int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
515 struct notifier_block *nb,
516 struct netlink_ext_ack *extack)
522 vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni)
527 static inline void vxlan_flag_attr_error(int attrtype,
528 struct netlink_ext_ack *extack)
530 #define VXLAN_FLAG(flg) \
531 case IFLA_VXLAN_##flg: \
532 NL_SET_ERR_MSG_MOD(extack, \
533 "cannot change " #flg " flag"); \
536 VXLAN_FLAG(TTL_INHERIT);
537 VXLAN_FLAG(LEARNING);
542 VXLAN_FLAG(COLLECT_METADATA);
543 VXLAN_FLAG(UDP_ZERO_CSUM6_TX);
544 VXLAN_FLAG(UDP_ZERO_CSUM6_RX);
545 VXLAN_FLAG(REMCSUM_TX);
546 VXLAN_FLAG(REMCSUM_RX);
549 VXLAN_FLAG(REMCSUM_NOPARTIAL);
551 NL_SET_ERR_MSG_MOD(extack, \
552 "cannot change flag");
558 static inline bool vxlan_fdb_nh_path_select(struct nexthop *nh,
560 struct vxlan_rdst *rdst)
562 struct fib_nh_common *nhc;
564 nhc = nexthop_path_fdb_result(nh, hash >> 1);
568 switch (nhc->nhc_gw_family) {
570 rdst->remote_ip.sin.sin_addr.s_addr = nhc->nhc_gw.ipv4;
571 rdst->remote_ip.sa.sa_family = AF_INET;
574 rdst->remote_ip.sin6.sin6_addr = nhc->nhc_gw.ipv6;
575 rdst->remote_ip.sa.sa_family = AF_INET6;
582 static inline void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, const struct vxlan_metadata *md)
584 struct vxlanhdr_gbp *gbp;
589 gbp = (struct vxlanhdr_gbp *)vxh;
590 vxh->vx_flags |= VXLAN_HF_GBP;
592 if (md->gbp & VXLAN_GBP_DONT_LEARN)
595 if (md->gbp & VXLAN_GBP_POLICY_APPLIED)
596 gbp->policy_applied = 1;
598 gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK);