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
;
224 enum ifla_vxlan_label_policy label_policy
;
226 unsigned long age_interval
;
227 unsigned int addrmax
;
229 enum ifla_vxlan_df df
;
234 VXLAN_VNI_STATS_RX_DROPS
,
235 VXLAN_VNI_STATS_RX_ERRORS
,
237 VXLAN_VNI_STATS_TX_DROPS
,
238 VXLAN_VNI_STATS_TX_ERRORS
,
241 struct vxlan_vni_stats
{
252 struct vxlan_vni_stats_pcpu
{
253 struct vxlan_vni_stats stats
;
254 struct u64_stats_sync syncp
;
257 struct vxlan_dev_node
{
258 struct hlist_node hlist
;
259 struct vxlan_dev
*vxlan
;
262 struct vxlan_vni_node
{
263 struct rhash_head vnode
;
264 struct vxlan_dev_node hlist4
; /* vni hash table for IPv4 socket */
265 #if IS_ENABLED(CONFIG_IPV6)
266 struct vxlan_dev_node hlist6
; /* vni hash table for IPv6 socket */
268 struct list_head vlist
;
270 union vxlan_addr remote_ip
; /* default remote ip for this vni */
271 struct vxlan_vni_stats_pcpu __percpu
*stats
;
276 struct vxlan_vni_group
{
277 struct rhashtable vni_hash
;
278 struct list_head vni_list
;
282 /* Pseudo network device */
284 struct vxlan_dev_node hlist4
; /* vni hash table for IPv4 socket */
285 #if IS_ENABLED(CONFIG_IPV6)
286 struct vxlan_dev_node hlist6
; /* vni hash table for IPv6 socket */
288 struct list_head next
; /* vxlan's per namespace list */
289 struct vxlan_sock __rcu
*vn4_sock
; /* listening socket for IPv4 */
290 #if IS_ENABLED(CONFIG_IPV6)
291 struct vxlan_sock __rcu
*vn6_sock
; /* listening socket for IPv6 */
293 struct net_device
*dev
;
294 struct net
*net
; /* netns for packet i/o */
295 struct vxlan_rdst default_dst
; /* default destination */
297 struct timer_list age_timer
;
298 spinlock_t hash_lock
[FDB_HASH_SIZE
];
299 unsigned int addrcnt
;
300 struct gro_cells gro_cells
;
302 struct vxlan_config cfg
;
304 struct vxlan_vni_group __rcu
*vnigrp
;
306 struct hlist_head fdb_head
[FDB_HASH_SIZE
];
308 struct rhashtable mdb_tbl
;
309 struct hlist_head mdb_list
;
310 unsigned int mdb_seq
;
313 #define VXLAN_F_LEARN 0x01
314 #define VXLAN_F_PROXY 0x02
315 #define VXLAN_F_RSC 0x04
316 #define VXLAN_F_L2MISS 0x08
317 #define VXLAN_F_L3MISS 0x10
318 #define VXLAN_F_IPV6 0x20
319 #define VXLAN_F_UDP_ZERO_CSUM_TX 0x40
320 #define VXLAN_F_UDP_ZERO_CSUM6_TX 0x80
321 #define VXLAN_F_UDP_ZERO_CSUM6_RX 0x100
322 #define VXLAN_F_REMCSUM_TX 0x200
323 #define VXLAN_F_REMCSUM_RX 0x400
324 #define VXLAN_F_GBP 0x800
325 #define VXLAN_F_REMCSUM_NOPARTIAL 0x1000
326 #define VXLAN_F_COLLECT_METADATA 0x2000
327 #define VXLAN_F_GPE 0x4000
328 #define VXLAN_F_IPV6_LINKLOCAL 0x8000
329 #define VXLAN_F_TTL_INHERIT 0x10000
330 #define VXLAN_F_VNIFILTER 0x20000
331 #define VXLAN_F_MDB 0x40000
332 #define VXLAN_F_LOCALBYPASS 0x80000
334 /* Flags that are used in the receive path. These flags must match in
335 * order for a socket to be shareable
337 #define VXLAN_F_RCV_FLAGS (VXLAN_F_GBP | \
339 VXLAN_F_UDP_ZERO_CSUM6_RX | \
340 VXLAN_F_REMCSUM_RX | \
341 VXLAN_F_REMCSUM_NOPARTIAL | \
342 VXLAN_F_COLLECT_METADATA | \
345 /* Flags that can be set together with VXLAN_F_GPE. */
346 #define VXLAN_F_ALLOWED_GPE (VXLAN_F_GPE | \
348 VXLAN_F_IPV6_LINKLOCAL | \
349 VXLAN_F_UDP_ZERO_CSUM_TX | \
350 VXLAN_F_UDP_ZERO_CSUM6_TX | \
351 VXLAN_F_UDP_ZERO_CSUM6_RX | \
352 VXLAN_F_COLLECT_METADATA | \
353 VXLAN_F_VNIFILTER | \
356 struct net_device
*vxlan_dev_create(struct net
*net
, const char *name
,
357 u8 name_assign_type
, struct vxlan_config
*conf
);
359 static inline netdev_features_t
vxlan_features_check(struct sk_buff
*skb
,
360 netdev_features_t features
)
364 if (!skb
->encapsulation
)
367 switch (vlan_get_protocol(skb
)) {
368 case htons(ETH_P_IP
):
369 l4_hdr
= ip_hdr(skb
)->protocol
;
371 case htons(ETH_P_IPV6
):
372 l4_hdr
= ipv6_hdr(skb
)->nexthdr
;
378 if ((l4_hdr
== IPPROTO_UDP
) &&
379 (skb
->inner_protocol_type
!= ENCAP_TYPE_ETHER
||
380 skb
->inner_protocol
!= htons(ETH_P_TEB
) ||
381 (skb_inner_mac_header(skb
) - skb_transport_header(skb
) !=
382 sizeof(struct udphdr
) + sizeof(struct vxlanhdr
)) ||
383 (skb
->ip_summed
!= CHECKSUM_NONE
&&
384 !can_checksum_protocol(features
, inner_eth_hdr(skb
)->h_proto
))))
385 return features
& ~(NETIF_F_CSUM_MASK
| NETIF_F_GSO_MASK
);
390 static inline int vxlan_headroom(u32 flags
)
392 /* VXLAN: IP4/6 header + UDP + VXLAN + Ethernet header */
393 /* VXLAN-GPE: IP4/6 header + UDP + VXLAN */
394 return (flags
& VXLAN_F_IPV6
? sizeof(struct ipv6hdr
) :
395 sizeof(struct iphdr
)) +
396 sizeof(struct udphdr
) + sizeof(struct vxlanhdr
) +
397 (flags
& VXLAN_F_GPE
? 0 : ETH_HLEN
);
400 static inline struct vxlanhdr
*vxlan_hdr(struct sk_buff
*skb
)
402 return (struct vxlanhdr
*)(udp_hdr(skb
) + 1);
405 static inline __be32
vxlan_vni(__be32 vni_field
)
407 #if defined(__BIG_ENDIAN)
408 return (__force __be32
)((__force u32
)vni_field
>> 8);
410 return (__force __be32
)((__force u32
)(vni_field
& VXLAN_VNI_MASK
) << 8);
414 static inline __be32
vxlan_vni_field(__be32 vni
)
416 #if defined(__BIG_ENDIAN)
417 return (__force __be32
)((__force u32
)vni
<< 8);
419 return (__force __be32
)((__force u32
)vni
>> 8);
423 static inline size_t vxlan_rco_start(__be32 vni_field
)
425 return be32_to_cpu(vni_field
& VXLAN_RCO_MASK
) << VXLAN_RCO_SHIFT
;
428 static inline size_t vxlan_rco_offset(__be32 vni_field
)
430 return (vni_field
& VXLAN_RCO_UDP
) ?
431 offsetof(struct udphdr
, check
) :
432 offsetof(struct tcphdr
, check
);
435 static inline __be32
vxlan_compute_rco(unsigned int start
, unsigned int offset
)
437 __be32 vni_field
= cpu_to_be32(start
>> VXLAN_RCO_SHIFT
);
439 if (offset
== offsetof(struct udphdr
, check
))
440 vni_field
|= VXLAN_RCO_UDP
;
444 static inline unsigned short vxlan_get_sk_family(struct vxlan_sock
*vs
)
446 return vs
->sock
->sk
->sk_family
;
449 #if IS_ENABLED(CONFIG_IPV6)
451 static inline bool vxlan_addr_any(const union vxlan_addr
*ipa
)
453 if (ipa
->sa
.sa_family
== AF_INET6
)
454 return ipv6_addr_any(&ipa
->sin6
.sin6_addr
);
456 return ipa
->sin
.sin_addr
.s_addr
== htonl(INADDR_ANY
);
459 static inline bool vxlan_addr_multicast(const union vxlan_addr
*ipa
)
461 if (ipa
->sa
.sa_family
== AF_INET6
)
462 return ipv6_addr_is_multicast(&ipa
->sin6
.sin6_addr
);
464 return ipv4_is_multicast(ipa
->sin
.sin_addr
.s_addr
);
467 #else /* !IS_ENABLED(CONFIG_IPV6) */
469 static inline bool vxlan_addr_any(const union vxlan_addr
*ipa
)
471 return ipa
->sin
.sin_addr
.s_addr
== htonl(INADDR_ANY
);
474 static inline bool vxlan_addr_multicast(const union vxlan_addr
*ipa
)
476 return ipv4_is_multicast(ipa
->sin
.sin_addr
.s_addr
);
479 #endif /* IS_ENABLED(CONFIG_IPV6) */
481 static inline bool netif_is_vxlan(const struct net_device
*dev
)
483 return dev
->rtnl_link_ops
&&
484 !strcmp(dev
->rtnl_link_ops
->kind
, "vxlan");
487 struct switchdev_notifier_vxlan_fdb_info
{
488 struct switchdev_notifier_info info
; /* must be first */
489 union vxlan_addr remote_ip
;
493 u8 eth_addr
[ETH_ALEN
];
499 #if IS_ENABLED(CONFIG_VXLAN)
500 int vxlan_fdb_find_uc(struct net_device
*dev
, const u8
*mac
, __be32 vni
,
501 struct switchdev_notifier_vxlan_fdb_info
*fdb_info
);
502 int vxlan_fdb_replay(const struct net_device
*dev
, __be32 vni
,
503 struct notifier_block
*nb
,
504 struct netlink_ext_ack
*extack
);
505 void vxlan_fdb_clear_offload(const struct net_device
*dev
, __be32 vni
);
509 vxlan_fdb_find_uc(struct net_device
*dev
, const u8
*mac
, __be32 vni
,
510 struct switchdev_notifier_vxlan_fdb_info
*fdb_info
)
515 static inline int vxlan_fdb_replay(const struct net_device
*dev
, __be32 vni
,
516 struct notifier_block
*nb
,
517 struct netlink_ext_ack
*extack
)
523 vxlan_fdb_clear_offload(const struct net_device
*dev
, __be32 vni
)
528 static inline void vxlan_flag_attr_error(int attrtype
,
529 struct netlink_ext_ack
*extack
)
531 #define VXLAN_FLAG(flg) \
532 case IFLA_VXLAN_##flg: \
533 NL_SET_ERR_MSG_MOD(extack, \
534 "cannot change " #flg " flag"); \
537 VXLAN_FLAG(TTL_INHERIT
);
538 VXLAN_FLAG(LEARNING
);
543 VXLAN_FLAG(COLLECT_METADATA
);
544 VXLAN_FLAG(UDP_ZERO_CSUM6_TX
);
545 VXLAN_FLAG(UDP_ZERO_CSUM6_RX
);
546 VXLAN_FLAG(REMCSUM_TX
);
547 VXLAN_FLAG(REMCSUM_RX
);
550 VXLAN_FLAG(REMCSUM_NOPARTIAL
);
552 NL_SET_ERR_MSG_MOD(extack
, \
553 "cannot change flag");
559 static inline bool vxlan_fdb_nh_path_select(struct nexthop
*nh
,
561 struct vxlan_rdst
*rdst
)
563 struct fib_nh_common
*nhc
;
565 nhc
= nexthop_path_fdb_result(nh
, hash
>> 1);
569 switch (nhc
->nhc_gw_family
) {
571 rdst
->remote_ip
.sin
.sin_addr
.s_addr
= nhc
->nhc_gw
.ipv4
;
572 rdst
->remote_ip
.sa
.sa_family
= AF_INET
;
575 rdst
->remote_ip
.sin6
.sin6_addr
= nhc
->nhc_gw
.ipv6
;
576 rdst
->remote_ip
.sa
.sa_family
= AF_INET6
;
583 static inline void vxlan_build_gbp_hdr(struct vxlanhdr
*vxh
, const struct vxlan_metadata
*md
)
585 struct vxlanhdr_gbp
*gbp
;
590 gbp
= (struct vxlanhdr_gbp
*)vxh
;
591 vxh
->vx_flags
|= VXLAN_HF_GBP
;
593 if (md
->gbp
& VXLAN_GBP_DONT_LEARN
)
596 if (md
->gbp
& VXLAN_GBP_POLICY_APPLIED
)
597 gbp
->policy_applied
= 1;
599 gbp
->policy_id
= htons(md
->gbp
& VXLAN_GBP_ID_MASK
);