tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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 */
24struct 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 */
89struct 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
150struct 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
180struct vxlan_metadata {
181 u32 gbp;
182};
183
184/* per UDP socket information */
185struct 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
193union vxlan_addr {
194 struct sockaddr_in sin;
195 struct sockaddr_in6 sin6;
196 struct sockaddr sa;
197};
198
199struct 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
211struct 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
230struct vxlan_dev_node {
231 struct hlist_node hlist;
232 struct vxlan_dev *vxlan;
233};
234
235/* Pseudo network device */
236struct vxlan_dev {
237 struct vxlan_dev_node hlist4; /* vni hash table for IPv4 socket */
238#if IS_ENABLED(CONFIG_IPV6)
239 struct vxlan_dev_node hlist6; /* vni hash table for IPv6 socket */
240#endif
241 struct list_head next; /* vxlan's per namespace list */
242 struct vxlan_sock __rcu *vn4_sock; /* listening socket for IPv4 */
243#if IS_ENABLED(CONFIG_IPV6)
244 struct vxlan_sock __rcu *vn6_sock; /* listening socket for IPv6 */
245#endif
246 struct net_device *dev;
247 struct net *net; /* netns for packet i/o */
248 struct vxlan_rdst default_dst; /* default destination */
249
250 struct timer_list age_timer;
251 spinlock_t hash_lock[FDB_HASH_SIZE];
252 unsigned int addrcnt;
253 struct gro_cells gro_cells;
254
255 struct vxlan_config cfg;
256
257 struct hlist_head fdb_head[FDB_HASH_SIZE];
258};
259
260#define VXLAN_F_LEARN 0x01
261#define VXLAN_F_PROXY 0x02
262#define VXLAN_F_RSC 0x04
263#define VXLAN_F_L2MISS 0x08
264#define VXLAN_F_L3MISS 0x10
265#define VXLAN_F_IPV6 0x20
266#define VXLAN_F_UDP_ZERO_CSUM_TX 0x40
267#define VXLAN_F_UDP_ZERO_CSUM6_TX 0x80
268#define VXLAN_F_UDP_ZERO_CSUM6_RX 0x100
269#define VXLAN_F_REMCSUM_TX 0x200
270#define VXLAN_F_REMCSUM_RX 0x400
271#define VXLAN_F_GBP 0x800
272#define VXLAN_F_REMCSUM_NOPARTIAL 0x1000
273#define VXLAN_F_COLLECT_METADATA 0x2000
274#define VXLAN_F_GPE 0x4000
275#define VXLAN_F_IPV6_LINKLOCAL 0x8000
276#define VXLAN_F_TTL_INHERIT 0x10000
277
278/* Flags that are used in the receive path. These flags must match in
279 * order for a socket to be shareable
280 */
281#define VXLAN_F_RCV_FLAGS (VXLAN_F_GBP | \
282 VXLAN_F_GPE | \
283 VXLAN_F_UDP_ZERO_CSUM6_RX | \
284 VXLAN_F_REMCSUM_RX | \
285 VXLAN_F_REMCSUM_NOPARTIAL | \
286 VXLAN_F_COLLECT_METADATA)
287
288/* Flags that can be set together with VXLAN_F_GPE. */
289#define VXLAN_F_ALLOWED_GPE (VXLAN_F_GPE | \
290 VXLAN_F_IPV6 | \
291 VXLAN_F_IPV6_LINKLOCAL | \
292 VXLAN_F_UDP_ZERO_CSUM_TX | \
293 VXLAN_F_UDP_ZERO_CSUM6_TX | \
294 VXLAN_F_UDP_ZERO_CSUM6_RX | \
295 VXLAN_F_COLLECT_METADATA)
296
297struct net_device *vxlan_dev_create(struct net *net, const char *name,
298 u8 name_assign_type, struct vxlan_config *conf);
299
300static inline netdev_features_t vxlan_features_check(struct sk_buff *skb,
301 netdev_features_t features)
302{
303 u8 l4_hdr = 0;
304
305 if (!skb->encapsulation)
306 return features;
307
308 switch (vlan_get_protocol(skb)) {
309 case htons(ETH_P_IP):
310 l4_hdr = ip_hdr(skb)->protocol;
311 break;
312 case htons(ETH_P_IPV6):
313 l4_hdr = ipv6_hdr(skb)->nexthdr;
314 break;
315 default:
316 return features;
317 }
318
319 if ((l4_hdr == IPPROTO_UDP) &&
320 (skb->inner_protocol_type != ENCAP_TYPE_ETHER ||
321 skb->inner_protocol != htons(ETH_P_TEB) ||
322 (skb_inner_mac_header(skb) - skb_transport_header(skb) !=
323 sizeof(struct udphdr) + sizeof(struct vxlanhdr)) ||
324 (skb->ip_summed != CHECKSUM_NONE &&
325 !can_checksum_protocol(features, inner_eth_hdr(skb)->h_proto))))
326 return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
327
328 return features;
329}
330
331/* IP header + UDP + VXLAN + Ethernet header */
332#define VXLAN_HEADROOM (20 + 8 + 8 + 14)
333/* IPv6 header + UDP + VXLAN + Ethernet header */
334#define VXLAN6_HEADROOM (40 + 8 + 8 + 14)
335
336static inline struct vxlanhdr *vxlan_hdr(struct sk_buff *skb)
337{
338 return (struct vxlanhdr *)(udp_hdr(skb) + 1);
339}
340
341static inline __be32 vxlan_vni(__be32 vni_field)
342{
343#if defined(__BIG_ENDIAN)
344 return (__force __be32)((__force u32)vni_field >> 8);
345#else
346 return (__force __be32)((__force u32)(vni_field & VXLAN_VNI_MASK) << 8);
347#endif
348}
349
350static inline __be32 vxlan_vni_field(__be32 vni)
351{
352#if defined(__BIG_ENDIAN)
353 return (__force __be32)((__force u32)vni << 8);
354#else
355 return (__force __be32)((__force u32)vni >> 8);
356#endif
357}
358
359static inline size_t vxlan_rco_start(__be32 vni_field)
360{
361 return be32_to_cpu(vni_field & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT;
362}
363
364static inline size_t vxlan_rco_offset(__be32 vni_field)
365{
366 return (vni_field & VXLAN_RCO_UDP) ?
367 offsetof(struct udphdr, check) :
368 offsetof(struct tcphdr, check);
369}
370
371static inline __be32 vxlan_compute_rco(unsigned int start, unsigned int offset)
372{
373 __be32 vni_field = cpu_to_be32(start >> VXLAN_RCO_SHIFT);
374
375 if (offset == offsetof(struct udphdr, check))
376 vni_field |= VXLAN_RCO_UDP;
377 return vni_field;
378}
379
380static inline unsigned short vxlan_get_sk_family(struct vxlan_sock *vs)
381{
382 return vs->sock->sk->sk_family;
383}
384
385#if IS_ENABLED(CONFIG_IPV6)
386
387static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
388{
389 if (ipa->sa.sa_family == AF_INET6)
390 return ipv6_addr_any(&ipa->sin6.sin6_addr);
391 else
392 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
393}
394
395static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
396{
397 if (ipa->sa.sa_family == AF_INET6)
398 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
399 else
400 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr);
401}
402
403#else /* !IS_ENABLED(CONFIG_IPV6) */
404
405static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
406{
407 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
408}
409
410static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
411{
412 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr);
413}
414
415#endif /* IS_ENABLED(CONFIG_IPV6) */
416
417static inline bool netif_is_vxlan(const struct net_device *dev)
418{
419 return dev->rtnl_link_ops &&
420 !strcmp(dev->rtnl_link_ops->kind, "vxlan");
421}
422
423struct switchdev_notifier_vxlan_fdb_info {
424 struct switchdev_notifier_info info; /* must be first */
425 union vxlan_addr remote_ip;
426 __be16 remote_port;
427 __be32 remote_vni;
428 u32 remote_ifindex;
429 u8 eth_addr[ETH_ALEN];
430 __be32 vni;
431 bool offloaded;
432 bool added_by_user;
433};
434
435#if IS_ENABLED(CONFIG_VXLAN)
436int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
437 struct switchdev_notifier_vxlan_fdb_info *fdb_info);
438int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
439 struct notifier_block *nb,
440 struct netlink_ext_ack *extack);
441void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni);
442
443#else
444static inline int
445vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
446 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
447{
448 return -ENOENT;
449}
450
451static inline int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
452 struct notifier_block *nb,
453 struct netlink_ext_ack *extack)
454{
455 return -EOPNOTSUPP;
456}
457
458static inline void
459vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni)
460{
461}
462#endif
463
464static inline void vxlan_flag_attr_error(int attrtype,
465 struct netlink_ext_ack *extack)
466{
467#define VXLAN_FLAG(flg) \
468 case IFLA_VXLAN_##flg: \
469 NL_SET_ERR_MSG_MOD(extack, \
470 "cannot change " #flg " flag"); \
471 break
472 switch (attrtype) {
473 VXLAN_FLAG(TTL_INHERIT);
474 VXLAN_FLAG(LEARNING);
475 VXLAN_FLAG(PROXY);
476 VXLAN_FLAG(RSC);
477 VXLAN_FLAG(L2MISS);
478 VXLAN_FLAG(L3MISS);
479 VXLAN_FLAG(COLLECT_METADATA);
480 VXLAN_FLAG(UDP_ZERO_CSUM6_TX);
481 VXLAN_FLAG(UDP_ZERO_CSUM6_RX);
482 VXLAN_FLAG(REMCSUM_TX);
483 VXLAN_FLAG(REMCSUM_RX);
484 VXLAN_FLAG(GBP);
485 VXLAN_FLAG(GPE);
486 VXLAN_FLAG(REMCSUM_NOPARTIAL);
487 default:
488 NL_SET_ERR_MSG_MOD(extack, \
489 "cannot change flag");
490 break;
491 }
492#undef VXLAN_FLAG
493}
494
495static inline bool vxlan_fdb_nh_path_select(struct nexthop *nh,
496 int hash,
497 struct vxlan_rdst *rdst)
498{
499 struct fib_nh_common *nhc;
500
501 nhc = nexthop_path_fdb_result(nh, hash);
502 if (unlikely(!nhc))
503 return false;
504
505 switch (nhc->nhc_gw_family) {
506 case AF_INET:
507 rdst->remote_ip.sin.sin_addr.s_addr = nhc->nhc_gw.ipv4;
508 rdst->remote_ip.sa.sa_family = AF_INET;
509 break;
510 case AF_INET6:
511 rdst->remote_ip.sin6.sin6_addr = nhc->nhc_gw.ipv6;
512 rdst->remote_ip.sa.sa_family = AF_INET6;
513 break;
514 }
515
516 return true;
517}
518
519#endif