projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
ipv4: route: Ignore output interface in FIB lookup for PMTU route
[linux-2.6-block.git] / net / ipv4 / ip_tunnel_core.c
CommitLineData
c9422999 1// SPDX-License-Identifier: GPL-2.0-only
0e6fbc5b
PS		 2/*
3 * Copyright (c) 2013 Nicira, Inc.
0e6fbc5b
PS		 4 */
5
6#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7
8#include <linux/types.h>
9#include <linux/kernel.h>
10#include <linux/skbuff.h>
11#include <linux/netdevice.h>
12#include <linux/in.h>
13#include <linux/if_arp.h>
0e6fbc5b
PS		 14#include <linux/init.h>
15#include <linux/in6.h>
16#include <linux/inetdevice.h>
17#include <linux/netfilter_ipv4.h>
18#include <linux/etherdevice.h>
19#include <linux/if_ether.h>
20#include <linux/if_vlan.h>
e7030878 21#include <linux/static_key.h>
0e6fbc5b
PS		 22
23#include <net/ip.h>
24#include <net/icmp.h>
25#include <net/protocol.h>
26#include <net/ip_tunnels.h>
058214a4 27#include <net/ip6_tunnel.h>
0e6fbc5b
PS		 28#include <net/arp.h>
29#include <net/checksum.h>
30#include <net/dsfield.h>
31#include <net/inet_ecn.h>
32#include <net/xfrm.h>
33#include <net/net_namespace.h>
34#include <net/netns/generic.h>
35#include <net/rtnetlink.h>
63d008a4 36#include <net/dst_metadata.h>
4ece4778 37#include <net/geneve.h>
edf31cbb 38#include <net/vxlan.h>
b0a21810 39#include <net/erspan.h>
0e6fbc5b 40
55c2bc14
TH		 41const struct ip_tunnel_encap_ops __rcu *
42 iptun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly;
43EXPORT_SYMBOL(iptun_encaps);
44
058214a4
TH		 45const struct ip6_tnl_encap_ops __rcu *
46 ip6tun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly;
47EXPORT_SYMBOL(ip6tun_encaps);
48
039f5062
PS		 49void iptunnel_xmit(struct sock *sk, struct rtable *rt, struct sk_buff *skb,
50 __be32 src, __be32 dst, __u8 proto,
51 __u8 tos, __u8 ttl, __be16 df, bool xnet)
0e6fbc5b 52{
bc22a0e2 53 int pkt_len = skb->len - skb_inner_network_offset(skb);
f859b0f6 54 struct net *net = dev_net(rt->dst.dev);
039f5062 55 struct net_device *dev = skb->dev;
0e6fbc5b
PS		 56 struct iphdr *iph;
57 int err;
58
963a88b3
ND		 59 skb_scrub_packet(skb, xnet);
60
bf8d85d4 61 skb_clear_hash_if_not_l4(skb);
0e6fbc5b
PS		 62 skb_dst_set(skb, &rt->dst);
63 memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
64
65 /* Push down and install the IP header. */
78a3694d 66 skb_push(skb, sizeof(struct iphdr));
0e6fbc5b
PS		 67 skb_reset_network_header(skb);
68
69 iph = ip_hdr(skb);
70
71 iph->version = 4;
72 iph->ihl = sizeof(struct iphdr) >> 2;
16f7eb2b 73 iph->frag_off = ip_mtu_locked(&rt->dst) ? 0 : df;
0e6fbc5b
PS		 74 iph->protocol = proto;
75 iph->tos = tos;
76 iph->daddr = dst;
77 iph->saddr = src;
78 iph->ttl = ttl;
f859b0f6 79 __ip_select_ident(net, iph, skb_shinfo(skb)->gso_segs ?: 1);
0e6fbc5b 80
33224b16 81 err = ip_local_out(net, sk, skb);
5684abf7
XL		 82
83 if (dev) {
84 if (unlikely(net_xmit_eval(err)))
85 pkt_len = 0;
86 iptunnel_xmit_stats(dev, pkt_len);
87 }
0e6fbc5b
PS		 88}
89EXPORT_SYMBOL_GPL(iptunnel_xmit);
3d7b46cd 90
a6d5bbf3
JB		 91int __iptunnel_pull_header(struct sk_buff *skb, int hdr_len,
92 __be16 inner_proto, bool raw_proto, bool xnet)
3d7b46cd
PS		 93{
94 if (unlikely(!pskb_may_pull(skb, hdr_len)))
95 return -ENOMEM;
96
97 skb_pull_rcsum(skb, hdr_len);
98
a6d5bbf3 99 if (!raw_proto && inner_proto == htons(ETH_P_TEB)) {
1245dfc8 100 struct ethhdr *eh;
3d7b46cd
PS		 101
102 if (unlikely(!pskb_may_pull(skb, ETH_HLEN)))
103 return -ENOMEM;
104
1245dfc8 105 eh = (struct ethhdr *)skb->data;
d181ddca 106 if (likely(eth_proto_is_802_3(eh->h_proto)))
3d7b46cd
PS		 107 skb->protocol = eh->h_proto;
108 else
109 skb->protocol = htons(ETH_P_802_2);
110
111 } else {
112 skb->protocol = inner_proto;
113 }
114
7539fadc 115 skb_clear_hash_if_not_l4(skb);
3e2ed0c2 116 __vlan_hwaccel_clear_tag(skb);
3d7b46cd 117 skb_set_queue_mapping(skb, 0);
7f290c94 118 skb_scrub_packet(skb, xnet);
a09a4c8d
JG		 119
120 return iptunnel_pull_offloads(skb);
3d7b46cd 121}
a6d5bbf3 122EXPORT_SYMBOL_GPL(__iptunnel_pull_header);
2d26f0a3 123
63d008a4
JB		 124struct metadata_dst *iptunnel_metadata_reply(struct metadata_dst *md,
125 gfp_t flags)
126{
127 struct metadata_dst *res;
128 struct ip_tunnel_info *dst, *src;
129
3fcece12
JK		 130 if (!md || md->type != METADATA_IP_TUNNEL ||
131 md->u.tun_info.mode & IP_TUNNEL_INFO_TX)
63d008a4
JB		 132 return NULL;
133
f52f11ec
XL		 134 src = &md->u.tun_info;
135 res = metadata_dst_alloc(src->options_len, METADATA_IP_TUNNEL, flags);
63d008a4
JB		 136 if (!res)
137 return NULL;
138
139 dst = &res->u.tun_info;
63d008a4
JB		 140 dst->key.tun_id = src->key.tun_id;
141 if (src->mode & IP_TUNNEL_INFO_IPV6)
142 memcpy(&dst->key.u.ipv6.dst, &src->key.u.ipv6.src,
143 sizeof(struct in6_addr));
144 else
145 dst->key.u.ipv4.dst = src->key.u.ipv4.src;
7bdca378 146 dst->key.tun_flags = src->key.tun_flags;
63d008a4 147 dst->mode = src->mode | IP_TUNNEL_INFO_TX;
f52f11ec
XL		 148 ip_tunnel_info_opts_set(dst, ip_tunnel_info_opts(src),
149 src->options_len, 0);
63d008a4
JB		 150
151 return res;
152}
153EXPORT_SYMBOL_GPL(iptunnel_metadata_reply);
154
aed069df
AD		 155int iptunnel_handle_offloads(struct sk_buff *skb,
156 int gso_type_mask)
2d26f0a3
ED		 157{
158 int err;
159
160 if (likely(!skb->encapsulation)) {
161 skb_reset_inner_headers(skb);
162 skb->encapsulation = 1;
163 }
164
165 if (skb_is_gso(skb)) {
9580bf2e 166 err = skb_header_unclone(skb, GFP_ATOMIC);
2d26f0a3 167 if (unlikely(err))
aed069df 168 return err;
2d26f0a3 169 skb_shinfo(skb)->gso_type |= gso_type_mask;
aed069df 170 return 0;
2d26f0a3
ED		 171 }
172
6fa79666 173 if (skb->ip_summed != CHECKSUM_PARTIAL) {
2d26f0a3 174 skb->ip_summed = CHECKSUM_NONE;
6fa79666
EC		 175 /* We clear encapsulation here to prevent badly-written
176 * drivers potentially deciding to offload an inner checksum
177 * if we set CHECKSUM_PARTIAL on the outer header.
178 * This should go away when the drivers are all fixed.
179 */
179bc67f
EC		 180 skb->encapsulation = 0;
181 }
2d26f0a3 182
aed069df 183 return 0;
2d26f0a3
ED		 184}
185EXPORT_SYMBOL_GPL(iptunnel_handle_offloads);
ebe44f35
DM		 186
187/* Often modified stats are per cpu, other are shared (netdev->stats) */
bc1f4470 188void ip_tunnel_get_stats64(struct net_device *dev,
189 struct rtnl_link_stats64 *tot)
ebe44f35
DM		 190{
191 int i;
192
c24a5964
AD		 193 netdev_stats_to_stats64(tot, &dev->stats);
194
ebe44f35
DM		 195 for_each_possible_cpu(i) {
196 const struct pcpu_sw_netstats *tstats =
197 per_cpu_ptr(dev->tstats, i);
198 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
199 unsigned int start;
200
201 do {
57a7744e 202 start = u64_stats_fetch_begin_irq(&tstats->syncp);
ebe44f35
DM		 203 rx_packets = tstats->rx_packets;
204 tx_packets = tstats->tx_packets;
205 rx_bytes = tstats->rx_bytes;
206 tx_bytes = tstats->tx_bytes;
57a7744e 207 } while (u64_stats_fetch_retry_irq(&tstats->syncp, start));
ebe44f35
DM		 208
209 tot->rx_packets += rx_packets;
210 tot->tx_packets += tx_packets;
211 tot->rx_bytes += rx_bytes;
212 tot->tx_bytes += tx_bytes;
213 }
ebe44f35
DM		 214}
215EXPORT_SYMBOL_GPL(ip_tunnel_get_stats64);
3093fbe7 216
a1c234f9 217static const struct nla_policy ip_tun_policy[LWTUNNEL_IP_MAX + 1] = {
7b6a70f7 218 [LWTUNNEL_IP_UNSPEC] = { .strict_start_type = LWTUNNEL_IP_OPTS },
a1c234f9
JB		 219 [LWTUNNEL_IP_ID] = { .type = NLA_U64 },
220 [LWTUNNEL_IP_DST] = { .type = NLA_U32 },
221 [LWTUNNEL_IP_SRC] = { .type = NLA_U32 },
222 [LWTUNNEL_IP_TTL] = { .type = NLA_U8 },
223 [LWTUNNEL_IP_TOS] = { .type = NLA_U8 },
a1c234f9 224 [LWTUNNEL_IP_FLAGS] = { .type = NLA_U16 },
4ece4778 225 [LWTUNNEL_IP_OPTS] = { .type = NLA_NESTED },
3093fbe7
TG		 226};
227
4ece4778
XL		 228static const struct nla_policy ip_opts_policy[LWTUNNEL_IP_OPTS_MAX + 1] = {
229 [LWTUNNEL_IP_OPTS_GENEVE] = { .type = NLA_NESTED },
edf31cbb 230 [LWTUNNEL_IP_OPTS_VXLAN] = { .type = NLA_NESTED },
b0a21810 231 [LWTUNNEL_IP_OPTS_ERSPAN] = { .type = NLA_NESTED },
4ece4778
XL		 232};
233
234static const struct nla_policy
235geneve_opt_policy[LWTUNNEL_IP_OPT_GENEVE_MAX + 1] = {
236 [LWTUNNEL_IP_OPT_GENEVE_CLASS] = { .type = NLA_U16 },
237 [LWTUNNEL_IP_OPT_GENEVE_TYPE] = { .type = NLA_U8 },
238 [LWTUNNEL_IP_OPT_GENEVE_DATA] = { .type = NLA_BINARY, .len = 128 },
239};
240
edf31cbb
XL		 241static const struct nla_policy
242vxlan_opt_policy[LWTUNNEL_IP_OPT_VXLAN_MAX + 1] = {
243 [LWTUNNEL_IP_OPT_VXLAN_GBP] = { .type = NLA_U32 },
244};
245
b0a21810
XL		 246static const struct nla_policy
247erspan_opt_policy[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1] = {
248 [LWTUNNEL_IP_OPT_ERSPAN_VER] = { .type = NLA_U8 },
249 [LWTUNNEL_IP_OPT_ERSPAN_INDEX] = { .type = NLA_U32 },
250 [LWTUNNEL_IP_OPT_ERSPAN_DIR] = { .type = NLA_U8 },
251 [LWTUNNEL_IP_OPT_ERSPAN_HWID] = { .type = NLA_U8 },
252};
253
4ece4778 254static int ip_tun_parse_opts_geneve(struct nlattr *attr,
2f1d370b 255 struct ip_tunnel_info *info, int opts_len,
4ece4778
XL		 256 struct netlink_ext_ack *extack)
257{
258 struct nlattr *tb[LWTUNNEL_IP_OPT_GENEVE_MAX + 1];
259 int data_len, err;
260
ed02551f
XL		 261 err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_GENEVE_MAX, attr,
262 geneve_opt_policy, extack);
4ece4778
XL		 263 if (err)
264 return err;
265
266 if (!tb[LWTUNNEL_IP_OPT_GENEVE_CLASS] ||
267 !tb[LWTUNNEL_IP_OPT_GENEVE_TYPE] ||
268 !tb[LWTUNNEL_IP_OPT_GENEVE_DATA])
269 return -EINVAL;
270
271 attr = tb[LWTUNNEL_IP_OPT_GENEVE_DATA];
272 data_len = nla_len(attr);
273 if (data_len % 4)
274 return -EINVAL;
275
276 if (info) {
2f1d370b 277 struct geneve_opt *opt = ip_tunnel_info_opts(info) + opts_len;
4ece4778
XL		 278
279 memcpy(opt->opt_data, nla_data(attr), data_len);
280 opt->length = data_len / 4;
281 attr = tb[LWTUNNEL_IP_OPT_GENEVE_CLASS];
282 opt->opt_class = nla_get_be16(attr);
283 attr = tb[LWTUNNEL_IP_OPT_GENEVE_TYPE];
284 opt->type = nla_get_u8(attr);
285 info->key.tun_flags |= TUNNEL_GENEVE_OPT;
286 }
287
288 return sizeof(struct geneve_opt) + data_len;
289}
290
edf31cbb 291static int ip_tun_parse_opts_vxlan(struct nlattr *attr,
2f1d370b 292 struct ip_tunnel_info *info, int opts_len,
edf31cbb
XL		 293 struct netlink_ext_ack *extack)
294{
295 struct nlattr *tb[LWTUNNEL_IP_OPT_VXLAN_MAX + 1];
296 int err;
297
ed02551f
XL		 298 err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_VXLAN_MAX, attr,
299 vxlan_opt_policy, extack);
edf31cbb
XL		 300 if (err)
301 return err;
302
303 if (!tb[LWTUNNEL_IP_OPT_VXLAN_GBP])
304 return -EINVAL;
305
306 if (info) {
2f1d370b
XL		 307 struct vxlan_metadata *md =
308 ip_tunnel_info_opts(info) + opts_len;
edf31cbb
XL		 309
310 attr = tb[LWTUNNEL_IP_OPT_VXLAN_GBP];
311 md->gbp = nla_get_u32(attr);
312 info->key.tun_flags |= TUNNEL_VXLAN_OPT;
313 }
314
315 return sizeof(struct vxlan_metadata);
316}
317
b0a21810 318static int ip_tun_parse_opts_erspan(struct nlattr *attr,
2f1d370b 319 struct ip_tunnel_info *info, int opts_len,
b0a21810
XL		 320 struct netlink_ext_ack *extack)
321{
322 struct nlattr *tb[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1];
323 int err;
1841b982 324 u8 ver;
b0a21810 325
ed02551f
XL		 326 err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_ERSPAN_MAX, attr,
327 erspan_opt_policy, extack);
b0a21810
XL		 328 if (err)
329 return err;
330
331 if (!tb[LWTUNNEL_IP_OPT_ERSPAN_VER])
332 return -EINVAL;
333
1841b982
XL		 334 ver = nla_get_u8(tb[LWTUNNEL_IP_OPT_ERSPAN_VER]);
335 if (ver == 1) {
336 if (!tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX])
337 return -EINVAL;
338 } else if (ver == 2) {
339 if (!tb[LWTUNNEL_IP_OPT_ERSPAN_DIR] ||
340 !tb[LWTUNNEL_IP_OPT_ERSPAN_HWID])
341 return -EINVAL;
342 } else {
343 return -EINVAL;
344 }
345
b0a21810 346 if (info) {
2f1d370b
XL		 347 struct erspan_metadata *md =
348 ip_tunnel_info_opts(info) + opts_len;
b0a21810 349
1841b982
XL		 350 md->version = ver;
351 if (ver == 1) {
b0a21810
XL		 352 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX];
353 md->u.index = nla_get_be32(attr);
1841b982 354 } else {
b0a21810
XL		 355 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_DIR];
356 md->u.md2.dir = nla_get_u8(attr);
357 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_HWID];
358 set_hwid(&md->u.md2, nla_get_u8(attr));
b0a21810
XL		 359 }
360
361 info->key.tun_flags |= TUNNEL_ERSPAN_OPT;
362 }
363
364 return sizeof(struct erspan_metadata);
365}
366
4ece4778
XL		 367static int ip_tun_parse_opts(struct nlattr *attr, struct ip_tunnel_info *info,
368 struct netlink_ext_ack *extack)
369{
2f1d370b
XL		 370 int err, rem, opt_len, opts_len = 0, type = 0;
371 struct nlattr *nla;
4ece4778
XL		 372
373 if (!attr)
374 return 0;
375
2f1d370b
XL		 376 err = nla_validate(nla_data(attr), nla_len(attr), LWTUNNEL_IP_OPTS_MAX,
377 ip_opts_policy, extack);
4ece4778
XL		 378 if (err)
379 return err;
380
2f1d370b
XL		 381 nla_for_each_attr(nla, nla_data(attr), nla_len(attr), rem) {
382 switch (nla_type(nla)) {
383 case LWTUNNEL_IP_OPTS_GENEVE:
384 if (type && type != TUNNEL_GENEVE_OPT)
385 return -EINVAL;
386 opt_len = ip_tun_parse_opts_geneve(nla, info, opts_len,
387 extack);
388 if (opt_len < 0)
389 return opt_len;
390 opts_len += opt_len;
391 if (opts_len > IP_TUNNEL_OPTS_MAX)
392 return -EINVAL;
393 type = TUNNEL_GENEVE_OPT;
394 break;
395 case LWTUNNEL_IP_OPTS_VXLAN:
396 if (type)
397 return -EINVAL;
398 opt_len = ip_tun_parse_opts_vxlan(nla, info, opts_len,
399 extack);
400 if (opt_len < 0)
401 return opt_len;
402 opts_len += opt_len;
403 type = TUNNEL_VXLAN_OPT;
404 break;
405 case LWTUNNEL_IP_OPTS_ERSPAN:
406 if (type)
407 return -EINVAL;
408 opt_len = ip_tun_parse_opts_erspan(nla, info, opts_len,
409 extack);
410 if (opt_len < 0)
411 return opt_len;
412 opts_len += opt_len;
413 type = TUNNEL_ERSPAN_OPT;
414 break;
415 default:
416 return -EINVAL;
417 }
418 }
4ece4778 419
2f1d370b 420 return opts_len;
4ece4778
XL		 421}
422
423static int ip_tun_get_optlen(struct nlattr *attr,
424 struct netlink_ext_ack *extack)
425{
426 return ip_tun_parse_opts(attr, NULL, extack);
427}
428
429static int ip_tun_set_opts(struct nlattr *attr, struct ip_tunnel_info *info,
430 struct netlink_ext_ack *extack)
431{
432 return ip_tun_parse_opts(attr, info, extack);
433}
434
faee6769 435static int ip_tun_build_state(struct net *net, struct nlattr *attr,
127eb7cd 436 unsigned int family, const void *cfg,
9ae28727
DA		 437 struct lwtunnel_state **ts,
438 struct netlink_ext_ack *extack)
3093fbe7 439{
a1c234f9 440 struct nlattr *tb[LWTUNNEL_IP_MAX + 1];
4ece4778
XL		 441 struct lwtunnel_state *new_state;
442 struct ip_tunnel_info *tun_info;
443 int err, opt_len;
3093fbe7 444
8cb08174
JB		 445 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_MAX, attr,
446 ip_tun_policy, extack);
3093fbe7
TG		 447 if (err < 0)
448 return err;
449
4ece4778
XL		 450 opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP_OPTS], extack);
451 if (opt_len < 0)
452 return opt_len;
453
454 new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len);
3093fbe7
TG		 455 if (!new_state)
456 return -ENOMEM;
457
458 new_state->type = LWTUNNEL_ENCAP_IP;
459
460 tun_info = lwt_tun_info(new_state);
461
4ece4778
XL		 462 err = ip_tun_set_opts(tb[LWTUNNEL_IP_OPTS], tun_info, extack);
463 if (err < 0) {
464 lwtstate_free(new_state);
465 return err;
466 }
467
3d25eabb 468#ifdef CONFIG_DST_CACHE
469 err = dst_cache_init(&tun_info->dst_cache, GFP_KERNEL);
470 if (err) {
471 lwtstate_free(new_state);
472 return err;
473 }
474#endif
475
a1c234f9 476 if (tb[LWTUNNEL_IP_ID])
30d3d83a 477 tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP_ID]);
3093fbe7 478
a1c234f9 479 if (tb[LWTUNNEL_IP_DST])
7822ce73 480 tun_info->key.u.ipv4.dst = nla_get_in_addr(tb[LWTUNNEL_IP_DST]);
3093fbe7 481
a1c234f9 482 if (tb[LWTUNNEL_IP_SRC])
7822ce73 483 tun_info->key.u.ipv4.src = nla_get_in_addr(tb[LWTUNNEL_IP_SRC]);
3093fbe7 484
a1c234f9 485 if (tb[LWTUNNEL_IP_TTL])
7c383fb2 486 tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP_TTL]);
3093fbe7 487
a1c234f9 488 if (tb[LWTUNNEL_IP_TOS])
7c383fb2 489 tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP_TOS]);
3093fbe7 490
a1c234f9 491 if (tb[LWTUNNEL_IP_FLAGS])
0c06d166
XL		 492 tun_info->key.tun_flags |=
493 (nla_get_be16(tb[LWTUNNEL_IP_FLAGS]) &
494 ~TUNNEL_OPTIONS_PRESENT);
3093fbe7
TG		 495
496 tun_info->mode = IP_TUNNEL_INFO_TX;
4ece4778 497 tun_info->options_len = opt_len;
3093fbe7
TG		 498
499 *ts = new_state;
500
501 return 0;
502}
503
3d25eabb 504static void ip_tun_destroy_state(struct lwtunnel_state *lwtstate)
505{
506#ifdef CONFIG_DST_CACHE
507 struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
508
509 dst_cache_destroy(&tun_info->dst_cache);
510#endif
511}
512
4ece4778
XL		 513static int ip_tun_fill_encap_opts_geneve(struct sk_buff *skb,
514 struct ip_tunnel_info *tun_info)
515{
516 struct geneve_opt *opt;
517 struct nlattr *nest;
2f1d370b 518 int offset = 0;
4ece4778
XL		 519
520 nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_GENEVE);
521 if (!nest)
522 return -ENOMEM;
523
2f1d370b
XL		 524 while (tun_info->options_len > offset) {
525 opt = ip_tunnel_info_opts(tun_info) + offset;
526 if (nla_put_be16(skb, LWTUNNEL_IP_OPT_GENEVE_CLASS,
527 opt->opt_class) ||
528 nla_put_u8(skb, LWTUNNEL_IP_OPT_GENEVE_TYPE, opt->type) ||
529 nla_put(skb, LWTUNNEL_IP_OPT_GENEVE_DATA, opt->length * 4,
530 opt->opt_data)) {
531 nla_nest_cancel(skb, nest);
532 return -ENOMEM;
533 }
534 offset += sizeof(*opt) + opt->length * 4;
4ece4778
XL		 535 }
536
537 nla_nest_end(skb, nest);
538 return 0;
539}
540
edf31cbb
XL		 541static int ip_tun_fill_encap_opts_vxlan(struct sk_buff *skb,
542 struct ip_tunnel_info *tun_info)
543{
544 struct vxlan_metadata *md;
545 struct nlattr *nest;
546
547 nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_VXLAN);
548 if (!nest)
549 return -ENOMEM;
550
551 md = ip_tunnel_info_opts(tun_info);
552 if (nla_put_u32(skb, LWTUNNEL_IP_OPT_VXLAN_GBP, md->gbp)) {
553 nla_nest_cancel(skb, nest);
554 return -ENOMEM;
555 }
556
557 nla_nest_end(skb, nest);
558 return 0;
559}
560
b0a21810
XL		 561static int ip_tun_fill_encap_opts_erspan(struct sk_buff *skb,
562 struct ip_tunnel_info *tun_info)
563{
564 struct erspan_metadata *md;
565 struct nlattr *nest;
566
567 nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_ERSPAN);
568 if (!nest)
569 return -ENOMEM;
570
571 md = ip_tunnel_info_opts(tun_info);
3132174b 572 if (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_VER, md->version))
b0a21810
XL		 573 goto err;
574
575 if (md->version == 1 &&
576 nla_put_be32(skb, LWTUNNEL_IP_OPT_ERSPAN_INDEX, md->u.index))
577 goto err;
578
579 if (md->version == 2 &&
580 (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_DIR, md->u.md2.dir) ||
581 nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_HWID,
582 get_hwid(&md->u.md2))))
583 goto err;
584
585 nla_nest_end(skb, nest);
586 return 0;
587err:
588 nla_nest_cancel(skb, nest);
589 return -ENOMEM;
590}
591
4ece4778
XL		 592static int ip_tun_fill_encap_opts(struct sk_buff *skb, int type,
593 struct ip_tunnel_info *tun_info)
594{
595 struct nlattr *nest;
596 int err = 0;
597
0c06d166 598 if (!(tun_info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))
4ece4778
XL		 599 return 0;
600
601 nest = nla_nest_start_noflag(skb, type);
602 if (!nest)
603 return -ENOMEM;
604
605 if (tun_info->key.tun_flags & TUNNEL_GENEVE_OPT)
606 err = ip_tun_fill_encap_opts_geneve(skb, tun_info);
edf31cbb
XL		 607 else if (tun_info->key.tun_flags & TUNNEL_VXLAN_OPT)
608 err = ip_tun_fill_encap_opts_vxlan(skb, tun_info);
b0a21810
XL		 609 else if (tun_info->key.tun_flags & TUNNEL_ERSPAN_OPT)
610 err = ip_tun_fill_encap_opts_erspan(skb, tun_info);
4ece4778
XL		 611
612 if (err) {
613 nla_nest_cancel(skb, nest);
614 return err;
615 }
616
617 nla_nest_end(skb, nest);
618 return 0;
619}
620
3093fbe7
TG		 621static int ip_tun_fill_encap_info(struct sk_buff *skb,
622 struct lwtunnel_state *lwtstate)
623{
624 struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
625
b46f6ded
ND		 626 if (nla_put_be64(skb, LWTUNNEL_IP_ID, tun_info->key.tun_id,
627 LWTUNNEL_IP_PAD) ||
7822ce73
HY		 628 nla_put_in_addr(skb, LWTUNNEL_IP_DST, tun_info->key.u.ipv4.dst) ||
629 nla_put_in_addr(skb, LWTUNNEL_IP_SRC, tun_info->key.u.ipv4.src) ||
7c383fb2
JB		 630 nla_put_u8(skb, LWTUNNEL_IP_TOS, tun_info->key.tos) ||
631 nla_put_u8(skb, LWTUNNEL_IP_TTL, tun_info->key.ttl) ||
4ece4778
XL		 632 nla_put_be16(skb, LWTUNNEL_IP_FLAGS, tun_info->key.tun_flags) ||
633 ip_tun_fill_encap_opts(skb, LWTUNNEL_IP_OPTS, tun_info))
3093fbe7
TG		 634 return -ENOMEM;
635
636 return 0;
637}
638
4ece4778
XL		 639static int ip_tun_opts_nlsize(struct ip_tunnel_info *info)
640{
641 int opt_len;
642
0c06d166 643 if (!(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))
4ece4778
XL		 644 return 0;
645
646 opt_len = nla_total_size(0); /* LWTUNNEL_IP_OPTS */
647 if (info->key.tun_flags & TUNNEL_GENEVE_OPT) {
2f1d370b
XL		 648 struct geneve_opt *opt;
649 int offset = 0;
650
651 opt_len += nla_total_size(0); /* LWTUNNEL_IP_OPTS_GENEVE */
652 while (info->options_len > offset) {
653 opt = ip_tunnel_info_opts(info) + offset;
654 opt_len += nla_total_size(2) /* OPT_GENEVE_CLASS */
655 + nla_total_size(1) /* OPT_GENEVE_TYPE */
656 + nla_total_size(opt->length * 4);
657 /* OPT_GENEVE_DATA */
658 offset += sizeof(*opt) + opt->length * 4;
659 }
edf31cbb
XL		 660 } else if (info->key.tun_flags & TUNNEL_VXLAN_OPT) {
661 opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_VXLAN */
662 + nla_total_size(4); /* OPT_VXLAN_GBP */
b0a21810 663 } else if (info->key.tun_flags & TUNNEL_ERSPAN_OPT) {
58e8494e
XL		 664 struct erspan_metadata *md = ip_tunnel_info_opts(info);
665
b0a21810
XL		 666 opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_ERSPAN */
667 + nla_total_size(1) /* OPT_ERSPAN_VER */
58e8494e
XL		 668 + (md->version == 1 ? nla_total_size(4)
669 /* OPT_ERSPAN_INDEX (v1) */
670 : nla_total_size(1) +
671 nla_total_size(1));
672 /* OPT_ERSPAN_DIR + HWID (v2) */
4ece4778
XL		 673 }
674
675 return opt_len;
676}
677
3093fbe7
TG		 678static int ip_tun_encap_nlsize(struct lwtunnel_state *lwtstate)
679{
b46f6ded 680 return nla_total_size_64bit(8) /* LWTUNNEL_IP_ID */
a1c234f9
JB		 681 + nla_total_size(4) /* LWTUNNEL_IP_DST */
682 + nla_total_size(4) /* LWTUNNEL_IP_SRC */
683 + nla_total_size(1) /* LWTUNNEL_IP_TOS */
684 + nla_total_size(1) /* LWTUNNEL_IP_TTL */
4ece4778
XL		 685 + nla_total_size(2) /* LWTUNNEL_IP_FLAGS */
686 + ip_tun_opts_nlsize(lwt_tun_info(lwtstate));
687 /* LWTUNNEL_IP_OPTS */
3093fbe7
TG		 688}
689
2d798499
JB		 690static int ip_tun_cmp_encap(struct lwtunnel_state *a, struct lwtunnel_state *b)
691{
0eb8eb2f
XL		 692 struct ip_tunnel_info *info_a = lwt_tun_info(a);
693 struct ip_tunnel_info *info_b = lwt_tun_info(b);
694
695 return memcmp(info_a, info_b, sizeof(info_a->key)) ||
696 info_a->mode != info_b->mode ||
697 info_a->options_len != info_b->options_len ||
698 memcmp(ip_tunnel_info_opts(info_a),
699 ip_tunnel_info_opts(info_b), info_a->options_len);
2d798499
JB		 700}
701
3093fbe7
TG		 702static const struct lwtunnel_encap_ops ip_tun_lwt_ops = {
703 .build_state = ip_tun_build_state,
3d25eabb 704 .destroy_state = ip_tun_destroy_state,
3093fbe7
TG		 705 .fill_encap = ip_tun_fill_encap_info,
706 .get_encap_size = ip_tun_encap_nlsize,
2d798499 707 .cmp_encap = ip_tun_cmp_encap,
88ff7334 708 .owner = THIS_MODULE,
3093fbe7
TG		 709};
710
32a2b002 711static const struct nla_policy ip6_tun_policy[LWTUNNEL_IP6_MAX + 1] = {
7b6a70f7 712 [LWTUNNEL_IP6_UNSPEC] = { .strict_start_type = LWTUNNEL_IP6_OPTS },
32a2b002
JB		 713 [LWTUNNEL_IP6_ID] = { .type = NLA_U64 },
714 [LWTUNNEL_IP6_DST] = { .len = sizeof(struct in6_addr) },
715 [LWTUNNEL_IP6_SRC] = { .len = sizeof(struct in6_addr) },
716 [LWTUNNEL_IP6_HOPLIMIT] = { .type = NLA_U8 },
717 [LWTUNNEL_IP6_TC] = { .type = NLA_U8 },
32a2b002 718 [LWTUNNEL_IP6_FLAGS] = { .type = NLA_U16 },
7b6a70f7 719 [LWTUNNEL_IP6_OPTS] = { .type = NLA_NESTED },
32a2b002
JB		 720};
721
faee6769 722static int ip6_tun_build_state(struct net *net, struct nlattr *attr,
127eb7cd 723 unsigned int family, const void *cfg,
9ae28727
DA		 724 struct lwtunnel_state **ts,
725 struct netlink_ext_ack *extack)
32a2b002 726{
32a2b002 727 struct nlattr *tb[LWTUNNEL_IP6_MAX + 1];
4ece4778
XL		 728 struct lwtunnel_state *new_state;
729 struct ip_tunnel_info *tun_info;
730 int err, opt_len;
32a2b002 731
8cb08174
JB		 732 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP6_MAX, attr,
733 ip6_tun_policy, extack);
32a2b002
JB		 734 if (err < 0)
735 return err;
736
4ece4778
XL		 737 opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP6_OPTS], extack);
738 if (opt_len < 0)
739 return opt_len;
740
741 new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len);
32a2b002
JB		 742 if (!new_state)
743 return -ENOMEM;
744
745 new_state->type = LWTUNNEL_ENCAP_IP6;
746
747 tun_info = lwt_tun_info(new_state);
748
4ece4778
XL		 749 err = ip_tun_set_opts(tb[LWTUNNEL_IP6_OPTS], tun_info, extack);
750 if (err < 0) {
751 lwtstate_free(new_state);
752 return err;
753 }
754
32a2b002 755 if (tb[LWTUNNEL_IP6_ID])
30d3d83a 756 tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP6_ID]);
32a2b002
JB		 757
758 if (tb[LWTUNNEL_IP6_DST])
759 tun_info->key.u.ipv6.dst = nla_get_in6_addr(tb[LWTUNNEL_IP6_DST]);
760
761 if (tb[LWTUNNEL_IP6_SRC])
762 tun_info->key.u.ipv6.src = nla_get_in6_addr(tb[LWTUNNEL_IP6_SRC]);
763
764 if (tb[LWTUNNEL_IP6_HOPLIMIT])
765 tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP6_HOPLIMIT]);
766
767 if (tb[LWTUNNEL_IP6_TC])
768 tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP6_TC]);
769
32a2b002 770 if (tb[LWTUNNEL_IP6_FLAGS])
0c06d166
XL		 771 tun_info->key.tun_flags |=
772 (nla_get_be16(tb[LWTUNNEL_IP6_FLAGS]) &
773 ~TUNNEL_OPTIONS_PRESENT);
32a2b002 774
7f9562a1 775 tun_info->mode = IP_TUNNEL_INFO_TX | IP_TUNNEL_INFO_IPV6;
4ece4778 776 tun_info->options_len = opt_len;
32a2b002
JB		 777
778 *ts = new_state;
779
780 return 0;
781}
782
783static int ip6_tun_fill_encap_info(struct sk_buff *skb,
784 struct lwtunnel_state *lwtstate)
785{
786 struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
787
b46f6ded
ND		 788 if (nla_put_be64(skb, LWTUNNEL_IP6_ID, tun_info->key.tun_id,
789 LWTUNNEL_IP6_PAD) ||
32a2b002
JB		 790 nla_put_in6_addr(skb, LWTUNNEL_IP6_DST, &tun_info->key.u.ipv6.dst) ||
791 nla_put_in6_addr(skb, LWTUNNEL_IP6_SRC, &tun_info->key.u.ipv6.src) ||
995096a0
QA		 792 nla_put_u8(skb, LWTUNNEL_IP6_TC, tun_info->key.tos) ||
793 nla_put_u8(skb, LWTUNNEL_IP6_HOPLIMIT, tun_info->key.ttl) ||
4ece4778
XL		 794 nla_put_be16(skb, LWTUNNEL_IP6_FLAGS, tun_info->key.tun_flags) ||
795 ip_tun_fill_encap_opts(skb, LWTUNNEL_IP6_OPTS, tun_info))
32a2b002
JB		 796 return -ENOMEM;
797
798 return 0;
799}
800
801static int ip6_tun_encap_nlsize(struct lwtunnel_state *lwtstate)
802{
b46f6ded 803 return nla_total_size_64bit(8) /* LWTUNNEL_IP6_ID */
32a2b002
JB		 804 + nla_total_size(16) /* LWTUNNEL_IP6_DST */
805 + nla_total_size(16) /* LWTUNNEL_IP6_SRC */
806 + nla_total_size(1) /* LWTUNNEL_IP6_HOPLIMIT */
807 + nla_total_size(1) /* LWTUNNEL_IP6_TC */
4ece4778
XL		 808 + nla_total_size(2) /* LWTUNNEL_IP6_FLAGS */
809 + ip_tun_opts_nlsize(lwt_tun_info(lwtstate));
810 /* LWTUNNEL_IP6_OPTS */
32a2b002
JB		 811}
812
813static const struct lwtunnel_encap_ops ip6_tun_lwt_ops = {
814 .build_state = ip6_tun_build_state,
815 .fill_encap = ip6_tun_fill_encap_info,
816 .get_encap_size = ip6_tun_encap_nlsize,
817 .cmp_encap = ip_tun_cmp_encap,
88ff7334 818 .owner = THIS_MODULE,
32a2b002
JB		 819};
820
045a0fa0 821void __init ip_tunnel_core_init(void)
3093fbe7 822{
fca5fdf6
DB		 823 /* If you land here, make sure whether increasing ip_tunnel_info's
824 * options_len is a reasonable choice with its usage in front ends
825 * (f.e., it's part of flow keys, etc).
826 */
827 BUILD_BUG_ON(IP_TUNNEL_OPTS_MAX != 255);
828
3093fbe7 829 lwtunnel_encap_add_ops(&ip_tun_lwt_ops, LWTUNNEL_ENCAP_IP);
32a2b002 830 lwtunnel_encap_add_ops(&ip6_tun_lwt_ops, LWTUNNEL_ENCAP_IP6);
3093fbe7 831}
e7030878 832
5263a98f 833DEFINE_STATIC_KEY_FALSE(ip_tunnel_metadata_cnt);
e7030878
TG		 834EXPORT_SYMBOL(ip_tunnel_metadata_cnt);
835
836void ip_tunnel_need_metadata(void)
837{
5263a98f 838 static_branch_inc(&ip_tunnel_metadata_cnt);
e7030878
TG		 839}
840EXPORT_SYMBOL_GPL(ip_tunnel_need_metadata);
841
842void ip_tunnel_unneed_metadata(void)
843{
5263a98f 844 static_branch_dec(&ip_tunnel_metadata_cnt);
e7030878
TG		 845}
846EXPORT_SYMBOL_GPL(ip_tunnel_unneed_metadata);
2606aff9
JD		 847
848/* Returns either the correct skb->protocol value, or 0 if invalid. */
849__be16 ip_tunnel_parse_protocol(const struct sk_buff *skb)
850{
851 if (skb_network_header(skb) >= skb->head &&
852 (skb_network_header(skb) + sizeof(struct iphdr)) <= skb_tail_pointer(skb) &&
853 ip_hdr(skb)->version == 4)
854 return htons(ETH_P_IP);
855 if (skb_network_header(skb) >= skb->head &&
856 (skb_network_header(skb) + sizeof(struct ipv6hdr)) <= skb_tail_pointer(skb) &&
857 ipv6_hdr(skb)->version == 6)
858 return htons(ETH_P_IPV6);
859 return 0;
860}
861EXPORT_SYMBOL(ip_tunnel_parse_protocol);
862
863const struct header_ops ip_tunnel_header_ops = { .parse_protocol = ip_tunnel_parse_protocol };
864EXPORT_SYMBOL(ip_tunnel_header_ops);
Linux 6.x block layer and io_uring tree(s)