Loading...
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 <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>
12
13#define IANA_VXLAN_UDP_PORT 4789
14#define IANA_VXLAN_GPE_UDP_PORT 4790
15
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 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
22 *
23 * I = VXLAN Network Identifier (VNI) present.
24 */
25struct vxlanhdr {
26 __be32 vx_flags;
27 __be32 vx_vni;
28};
29
30/* VXLAN header flags. */
31#define VXLAN_HF_VNI cpu_to_be32(BIT(27))
32
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))
37
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)
42
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 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
49 *
50 * C = Remote checksum offload bit. When set indicates that the
51 * remote checksum offload data is present.
52 *
53 * O = Offset bit. Indicates the checksum offset relative to
54 * checksum start.
55 *
56 * Csum start = Checksum start divided by two.
57 *
58 * http://tools.ietf.org/html/draft-herbert-vxlan-rco
59 */
60
61/* VXLAN-RCO header flags. */
62#define VXLAN_HF_RCO cpu_to_be32(BIT(21))
63
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)
70
71/*
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 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
78 *
79 * G = Group Policy ID present.
80 *
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.
83 *
84 * A = Indicates that the group policy has already been applied to
85 * this packet. Policies MUST NOT be applied by devices when the
86 * A bit is set.
87 *
88 * https://tools.ietf.org/html/draft-smith-vxlan-group-policy
89 */
90struct vxlanhdr_gbp {
91 u8 vx_flags;
92#ifdef __LITTLE_ENDIAN_BITFIELD
93 u8 reserved_flags1:3,
94 policy_applied:1,
95 reserved_flags2:2,
96 dont_learn:1,
97 reserved_flags3:1;
98#elif defined(__BIG_ENDIAN_BITFIELD)
99 u8 reserved_flags1:1,
100 dont_learn:1,
101 reserved_flags2:2,
102 policy_applied:1,
103 reserved_flags3:3;
104#else
105#error "Please fix <asm/byteorder.h>"
106#endif
107 __be16 policy_id;
108 __be32 vx_vni;
109};
110
111/* VXLAN-GBP header flags. */
112#define VXLAN_HF_GBP cpu_to_be32(BIT(31))
113
114#define VXLAN_GBP_USED_BITS (VXLAN_HF_GBP | cpu_to_be32(0xFFFFFF))
115
116/* skb->mark mapping
117 *
118 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
119 * |R|R|R|R|R|R|R|R|R|D|R|R|A|R|R|R| Group Policy ID |
120 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
121 */
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)
125
126#define VXLAN_GBP_MASK (VXLAN_GBP_DONT_LEARN | VXLAN_GBP_POLICY_APPLIED | \
127 VXLAN_GBP_ID_MASK)
128
129/*
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 * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
136 *
137 * Ver = Version. Indicates VXLAN GPE protocol version.
138 *
139 * P = Next Protocol Bit. The P bit is set to indicate that the
140 * Next Protocol field is present.
141 *
142 * O = OAM Flag Bit. The O bit is set to indicate that the packet
143 * is an OAM packet.
144 *
145 * Next Protocol = This 8 bit field indicates the protocol header
146 * immediately following the VXLAN GPE header.
147 *
148 * https://tools.ietf.org/html/draft-ietf-nvo3-vxlan-gpe-01
149 */
150
151struct vxlanhdr_gpe {
152#if defined(__LITTLE_ENDIAN_BITFIELD)
153 u8 oam_flag:1,
154 reserved_flags1:1,
155 np_applied:1,
156 instance_applied:1,
157 version:2,
158 reserved_flags2:2;
159#elif defined(__BIG_ENDIAN_BITFIELD)
160 u8 reserved_flags2:2,
161 version:2,
162 instance_applied:1,
163 np_applied:1,
164 reserved_flags1:1,
165 oam_flag:1;
166#endif
167 u8 reserved_flags3;
168 u8 reserved_flags4;
169 u8 next_protocol;
170 __be32 vx_vni;
171};
172
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))
177
178#define VXLAN_GPE_USED_BITS (VXLAN_HF_VER | VXLAN_HF_NP | VXLAN_HF_OAM | \
179 cpu_to_be32(0xff))
180
181struct vxlan_metadata {
182 u32 gbp;
183};
184
185/* per UDP socket information */
186struct vxlan_sock {
187 struct hlist_node hlist;
188 struct socket *sock;
189 struct hlist_head vni_list[VNI_HASH_SIZE];
190 refcount_t refcnt;
191 u32 flags;
192};
193
194union vxlan_addr {
195 struct sockaddr_in sin;
196 struct sockaddr_in6 sin6;
197 struct sockaddr sa;
198};
199
200struct vxlan_rdst {
201 union vxlan_addr remote_ip;
202 __be16 remote_port;
203 u8 offloaded:1;
204 __be32 remote_vni;
205 u32 remote_ifindex;
206 struct net_device *remote_dev;
207 struct list_head list;
208 struct rcu_head rcu;
209 struct dst_cache dst_cache;
210};
211
212struct vxlan_config {
213 union vxlan_addr remote_ip;
214 union vxlan_addr saddr;
215 __be32 vni;
216 int remote_ifindex;
217 int mtu;
218 __be16 dst_port;
219 u16 port_min;
220 u16 port_max;
221 u8 tos;
222 u8 ttl;
223 __be32 label;
224 enum ifla_vxlan_label_policy label_policy;
225 u32 flags;
226 unsigned long age_interval;
227 unsigned int addrmax;
228 bool no_share;
229 enum ifla_vxlan_df df;
230};
231
232enum {
233 VXLAN_VNI_STATS_RX,
234 VXLAN_VNI_STATS_RX_DROPS,
235 VXLAN_VNI_STATS_RX_ERRORS,
236 VXLAN_VNI_STATS_TX,
237 VXLAN_VNI_STATS_TX_DROPS,
238 VXLAN_VNI_STATS_TX_ERRORS,
239};
240
241struct vxlan_vni_stats {
242 u64 rx_packets;
243 u64 rx_bytes;
244 u64 rx_drops;
245 u64 rx_errors;
246 u64 tx_packets;
247 u64 tx_bytes;
248 u64 tx_drops;
249 u64 tx_errors;
250};
251
252struct vxlan_vni_stats_pcpu {
253 struct vxlan_vni_stats stats;
254 struct u64_stats_sync syncp;
255};
256
257struct vxlan_dev_node {
258 struct hlist_node hlist;
259 struct vxlan_dev *vxlan;
260};
261
262struct 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 */
267#endif
268 struct list_head vlist;
269 __be32 vni;
270 union vxlan_addr remote_ip; /* default remote ip for this vni */
271 struct vxlan_vni_stats_pcpu __percpu *stats;
272
273 struct rcu_head rcu;
274};
275
276struct vxlan_vni_group {
277 struct rhashtable vni_hash;
278 struct list_head vni_list;
279 u32 num_vnis;
280};
281
282/* Pseudo network device */
283struct vxlan_dev {
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 */
287#endif
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 */
292#endif
293 struct net_device *dev;
294 struct net *net; /* netns for packet i/o */
295 struct vxlan_rdst default_dst; /* default destination */
296
297 struct timer_list age_timer;
298 spinlock_t hash_lock[FDB_HASH_SIZE];
299 unsigned int addrcnt;
300 struct gro_cells gro_cells;
301
302 struct vxlan_config cfg;
303
304 struct vxlan_vni_group __rcu *vnigrp;
305
306 struct hlist_head fdb_head[FDB_HASH_SIZE];
307
308 struct rhashtable mdb_tbl;
309 struct hlist_head mdb_list;
310 unsigned int mdb_seq;
311};
312
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
333
334/* Flags that are used in the receive path. These flags must match in
335 * order for a socket to be shareable
336 */
337#define VXLAN_F_RCV_FLAGS (VXLAN_F_GBP | \
338 VXLAN_F_GPE | \
339 VXLAN_F_UDP_ZERO_CSUM6_RX | \
340 VXLAN_F_REMCSUM_RX | \
341 VXLAN_F_REMCSUM_NOPARTIAL | \
342 VXLAN_F_COLLECT_METADATA | \
343 VXLAN_F_VNIFILTER)
344
345/* Flags that can be set together with VXLAN_F_GPE. */
346#define VXLAN_F_ALLOWED_GPE (VXLAN_F_GPE | \
347 VXLAN_F_IPV6 | \
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 | \
354 VXLAN_F_LOCALBYPASS)
355
356struct net_device *vxlan_dev_create(struct net *net, const char *name,
357 u8 name_assign_type, struct vxlan_config *conf);
358
359static inline netdev_features_t vxlan_features_check(struct sk_buff *skb,
360 netdev_features_t features)
361{
362 u8 l4_hdr = 0;
363
364 if (!skb->encapsulation)
365 return features;
366
367 switch (vlan_get_protocol(skb)) {
368 case htons(ETH_P_IP):
369 l4_hdr = ip_hdr(skb)->protocol;
370 break;
371 case htons(ETH_P_IPV6):
372 l4_hdr = ipv6_hdr(skb)->nexthdr;
373 break;
374 default:
375 return features;
376 }
377
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);
386
387 return features;
388}
389
390static inline int vxlan_headroom(u32 flags)
391{
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);
398}
399
400static inline struct vxlanhdr *vxlan_hdr(struct sk_buff *skb)
401{
402 return (struct vxlanhdr *)(udp_hdr(skb) + 1);
403}
404
405static inline __be32 vxlan_vni(__be32 vni_field)
406{
407#if defined(__BIG_ENDIAN)
408 return (__force __be32)((__force u32)vni_field >> 8);
409#else
410 return (__force __be32)((__force u32)(vni_field & VXLAN_VNI_MASK) << 8);
411#endif
412}
413
414static inline __be32 vxlan_vni_field(__be32 vni)
415{
416#if defined(__BIG_ENDIAN)
417 return (__force __be32)((__force u32)vni << 8);
418#else
419 return (__force __be32)((__force u32)vni >> 8);
420#endif
421}
422
423static inline size_t vxlan_rco_start(__be32 vni_field)
424{
425 return be32_to_cpu(vni_field & VXLAN_RCO_MASK) << VXLAN_RCO_SHIFT;
426}
427
428static inline size_t vxlan_rco_offset(__be32 vni_field)
429{
430 return (vni_field & VXLAN_RCO_UDP) ?
431 offsetof(struct udphdr, check) :
432 offsetof(struct tcphdr, check);
433}
434
435static inline __be32 vxlan_compute_rco(unsigned int start, unsigned int offset)
436{
437 __be32 vni_field = cpu_to_be32(start >> VXLAN_RCO_SHIFT);
438
439 if (offset == offsetof(struct udphdr, check))
440 vni_field |= VXLAN_RCO_UDP;
441 return vni_field;
442}
443
444static inline unsigned short vxlan_get_sk_family(struct vxlan_sock *vs)
445{
446 return vs->sock->sk->sk_family;
447}
448
449#if IS_ENABLED(CONFIG_IPV6)
450
451static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
452{
453 if (ipa->sa.sa_family == AF_INET6)
454 return ipv6_addr_any(&ipa->sin6.sin6_addr);
455 else
456 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
457}
458
459static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
460{
461 if (ipa->sa.sa_family == AF_INET6)
462 return ipv6_addr_is_multicast(&ipa->sin6.sin6_addr);
463 else
464 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr);
465}
466
467#else /* !IS_ENABLED(CONFIG_IPV6) */
468
469static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
470{
471 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
472}
473
474static inline bool vxlan_addr_multicast(const union vxlan_addr *ipa)
475{
476 return ipv4_is_multicast(ipa->sin.sin_addr.s_addr);
477}
478
479#endif /* IS_ENABLED(CONFIG_IPV6) */
480
481static inline bool netif_is_vxlan(const struct net_device *dev)
482{
483 return dev->rtnl_link_ops &&
484 !strcmp(dev->rtnl_link_ops->kind, "vxlan");
485}
486
487struct switchdev_notifier_vxlan_fdb_info {
488 struct switchdev_notifier_info info; /* must be first */
489 union vxlan_addr remote_ip;
490 __be16 remote_port;
491 __be32 remote_vni;
492 u32 remote_ifindex;
493 u8 eth_addr[ETH_ALEN];
494 __be32 vni;
495 bool offloaded;
496 bool added_by_user;
497};
498
499#if IS_ENABLED(CONFIG_VXLAN)
500int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
501 struct switchdev_notifier_vxlan_fdb_info *fdb_info);
502int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
503 struct notifier_block *nb,
504 struct netlink_ext_ack *extack);
505void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni);
506
507#else
508static inline int
509vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
510 struct switchdev_notifier_vxlan_fdb_info *fdb_info)
511{
512 return -ENOENT;
513}
514
515static inline int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
516 struct notifier_block *nb,
517 struct netlink_ext_ack *extack)
518{
519 return -EOPNOTSUPP;
520}
521
522static inline void
523vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni)
524{
525}
526#endif
527
528static inline void vxlan_flag_attr_error(int attrtype,
529 struct netlink_ext_ack *extack)
530{
531#define VXLAN_FLAG(flg) \
532 case IFLA_VXLAN_##flg: \
533 NL_SET_ERR_MSG_MOD(extack, \
534 "cannot change " #flg " flag"); \
535 break
536 switch (attrtype) {
537 VXLAN_FLAG(TTL_INHERIT);
538 VXLAN_FLAG(LEARNING);
539 VXLAN_FLAG(PROXY);
540 VXLAN_FLAG(RSC);
541 VXLAN_FLAG(L2MISS);
542 VXLAN_FLAG(L3MISS);
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);
548 VXLAN_FLAG(GBP);
549 VXLAN_FLAG(GPE);
550 VXLAN_FLAG(REMCSUM_NOPARTIAL);
551 default:
552 NL_SET_ERR_MSG_MOD(extack, \
553 "cannot change flag");
554 break;
555 }
556#undef VXLAN_FLAG
557}
558
559static inline bool vxlan_fdb_nh_path_select(struct nexthop *nh,
560 u32 hash,
561 struct vxlan_rdst *rdst)
562{
563 struct fib_nh_common *nhc;
564
565 nhc = nexthop_path_fdb_result(nh, hash >> 1);
566 if (unlikely(!nhc))
567 return false;
568
569 switch (nhc->nhc_gw_family) {
570 case AF_INET:
571 rdst->remote_ip.sin.sin_addr.s_addr = nhc->nhc_gw.ipv4;
572 rdst->remote_ip.sa.sa_family = AF_INET;
573 break;
574 case AF_INET6:
575 rdst->remote_ip.sin6.sin6_addr = nhc->nhc_gw.ipv6;
576 rdst->remote_ip.sa.sa_family = AF_INET6;
577 break;
578 }
579
580 return true;
581}
582
583static inline void vxlan_build_gbp_hdr(struct vxlanhdr *vxh, const struct vxlan_metadata *md)
584{
585 struct vxlanhdr_gbp *gbp;
586
587 if (!md->gbp)
588 return;
589
590 gbp = (struct vxlanhdr_gbp *)vxh;
591 vxh->vx_flags |= VXLAN_HF_GBP;
592
593 if (md->gbp & VXLAN_GBP_DONT_LEARN)
594 gbp->dont_learn = 1;
595
596 if (md->gbp & VXLAN_GBP_POLICY_APPLIED)
597 gbp->policy_applied = 1;
598
599 gbp->policy_id = htons(md->gbp & VXLAN_GBP_ID_MASK);
600}
601
602#endif
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
230enum {
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
239struct 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
250struct vxlan_vni_stats_pcpu {
251 struct vxlan_vni_stats stats;
252 struct u64_stats_sync syncp;
253};
254
255struct vxlan_dev_node {
256 struct hlist_node hlist;
257 struct vxlan_dev *vxlan;
258};
259
260struct 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
274struct vxlan_vni_group {
275 struct rhashtable vni_hash;
276 struct list_head vni_list;
277 u32 num_vnis;
278};
279
280/* Pseudo network device */
281struct 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
347struct net_device *vxlan_dev_create(struct net *net, const char *name,
348 u8 name_assign_type, struct vxlan_config *conf);
349
350static 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
386static inline struct vxlanhdr *vxlan_hdr(struct sk_buff *skb)
387{
388 return (struct vxlanhdr *)(udp_hdr(skb) + 1);
389}
390
391static 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
400static 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
409static 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
414static 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
421static 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
430static 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
437static 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
445static 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
455static inline bool vxlan_addr_any(const union vxlan_addr *ipa)
456{
457 return ipa->sin.sin_addr.s_addr == htonl(INADDR_ANY);
458}
459
460static 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
467static 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
473struct 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)
486int vxlan_fdb_find_uc(struct net_device *dev, const u8 *mac, __be32 vni,
487 struct switchdev_notifier_vxlan_fdb_info *fdb_info);
488int vxlan_fdb_replay(const struct net_device *dev, __be32 vni,
489 struct notifier_block *nb,
490 struct netlink_ext_ack *extack);
491void vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni);
492
493#else
494static inline int
495vxlan_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
501static 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
508static inline void
509vxlan_fdb_clear_offload(const struct net_device *dev, __be32 vni)
510{
511}
512#endif
513
514static 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
545static 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