projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Merge branch 'Accomodate-DSA-front-end-into-Ocelot'
[linux-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
JB		 217static const struct nla_policy ip_tun_policy[LWTUNNEL_IP_MAX + 1] = {
218 [LWTUNNEL_IP_ID] = { .type = NLA_U64 },
219 [LWTUNNEL_IP_DST] = { .type = NLA_U32 },
220 [LWTUNNEL_IP_SRC] = { .type = NLA_U32 },
221 [LWTUNNEL_IP_TTL] = { .type = NLA_U8 },
222 [LWTUNNEL_IP_TOS] = { .type = NLA_U8 },
a1c234f9 223 [LWTUNNEL_IP_FLAGS] = { .type = NLA_U16 },
4ece4778 224 [LWTUNNEL_IP_OPTS] = { .type = NLA_NESTED },
3093fbe7
TG		 225};
226
4ece4778
XL		 227static const struct nla_policy ip_opts_policy[LWTUNNEL_IP_OPTS_MAX + 1] = {
228 [LWTUNNEL_IP_OPTS_GENEVE] = { .type = NLA_NESTED },
edf31cbb 229 [LWTUNNEL_IP_OPTS_VXLAN] = { .type = NLA_NESTED },
b0a21810 230 [LWTUNNEL_IP_OPTS_ERSPAN] = { .type = NLA_NESTED },
4ece4778
XL		 231};
232
233static const struct nla_policy
234geneve_opt_policy[LWTUNNEL_IP_OPT_GENEVE_MAX + 1] = {
235 [LWTUNNEL_IP_OPT_GENEVE_CLASS] = { .type = NLA_U16 },
236 [LWTUNNEL_IP_OPT_GENEVE_TYPE] = { .type = NLA_U8 },
237 [LWTUNNEL_IP_OPT_GENEVE_DATA] = { .type = NLA_BINARY, .len = 128 },
238};
239
edf31cbb
XL		 240static const struct nla_policy
241vxlan_opt_policy[LWTUNNEL_IP_OPT_VXLAN_MAX + 1] = {
242 [LWTUNNEL_IP_OPT_VXLAN_GBP] = { .type = NLA_U32 },
243};
244
b0a21810
XL		 245static const struct nla_policy
246erspan_opt_policy[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1] = {
247 [LWTUNNEL_IP_OPT_ERSPAN_VER] = { .type = NLA_U8 },
248 [LWTUNNEL_IP_OPT_ERSPAN_INDEX] = { .type = NLA_U32 },
249 [LWTUNNEL_IP_OPT_ERSPAN_DIR] = { .type = NLA_U8 },
250 [LWTUNNEL_IP_OPT_ERSPAN_HWID] = { .type = NLA_U8 },
251};
252
4ece4778
XL		 253static int ip_tun_parse_opts_geneve(struct nlattr *attr,
254 struct ip_tunnel_info *info,
255 struct netlink_ext_ack *extack)
256{
257 struct nlattr *tb[LWTUNNEL_IP_OPT_GENEVE_MAX + 1];
258 int data_len, err;
259
260 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_OPT_GENEVE_MAX,
261 attr, geneve_opt_policy, extack);
262 if (err)
263 return err;
264
265 if (!tb[LWTUNNEL_IP_OPT_GENEVE_CLASS] ||
266 !tb[LWTUNNEL_IP_OPT_GENEVE_TYPE] ||
267 !tb[LWTUNNEL_IP_OPT_GENEVE_DATA])
268 return -EINVAL;
269
270 attr = tb[LWTUNNEL_IP_OPT_GENEVE_DATA];
271 data_len = nla_len(attr);
272 if (data_len % 4)
273 return -EINVAL;
274
275 if (info) {
276 struct geneve_opt *opt = ip_tunnel_info_opts(info);
277
278 memcpy(opt->opt_data, nla_data(attr), data_len);
279 opt->length = data_len / 4;
280 attr = tb[LWTUNNEL_IP_OPT_GENEVE_CLASS];
281 opt->opt_class = nla_get_be16(attr);
282 attr = tb[LWTUNNEL_IP_OPT_GENEVE_TYPE];
283 opt->type = nla_get_u8(attr);
284 info->key.tun_flags |= TUNNEL_GENEVE_OPT;
285 }
286
287 return sizeof(struct geneve_opt) + data_len;
288}
289
edf31cbb
XL		 290static int ip_tun_parse_opts_vxlan(struct nlattr *attr,
291 struct ip_tunnel_info *info,
292 struct netlink_ext_ack *extack)
293{
294 struct nlattr *tb[LWTUNNEL_IP_OPT_VXLAN_MAX + 1];
295 int err;
296
297 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_OPT_VXLAN_MAX,
298 attr, vxlan_opt_policy, extack);
299 if (err)
300 return err;
301
302 if (!tb[LWTUNNEL_IP_OPT_VXLAN_GBP])
303 return -EINVAL;
304
305 if (info) {
306 struct vxlan_metadata *md = ip_tunnel_info_opts(info);
307
308 attr = tb[LWTUNNEL_IP_OPT_VXLAN_GBP];
309 md->gbp = nla_get_u32(attr);
310 info->key.tun_flags |= TUNNEL_VXLAN_OPT;
311 }
312
313 return sizeof(struct vxlan_metadata);
314}
315
b0a21810
XL		 316static int ip_tun_parse_opts_erspan(struct nlattr *attr,
317 struct ip_tunnel_info *info,
318 struct netlink_ext_ack *extack)
319{
320 struct nlattr *tb[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1];
321 int err;
322
323 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_OPT_ERSPAN_MAX,
324 attr, erspan_opt_policy, extack);
325 if (err)
326 return err;
327
328 if (!tb[LWTUNNEL_IP_OPT_ERSPAN_VER])
329 return -EINVAL;
330
331 if (info) {
332 struct erspan_metadata *md = ip_tunnel_info_opts(info);
333
334 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_VER];
335 md->version = nla_get_u8(attr);
336
337 if (md->version == 1 && tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX]) {
338 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX];
339 md->u.index = nla_get_be32(attr);
340 } else if (md->version == 2 && tb[LWTUNNEL_IP_OPT_ERSPAN_DIR] &&
341 tb[LWTUNNEL_IP_OPT_ERSPAN_HWID]) {
342 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_DIR];
343 md->u.md2.dir = nla_get_u8(attr);
344 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_HWID];
345 set_hwid(&md->u.md2, nla_get_u8(attr));
346 } else {
347 return -EINVAL;
348 }
349
350 info->key.tun_flags |= TUNNEL_ERSPAN_OPT;
351 }
352
353 return sizeof(struct erspan_metadata);
354}
355
4ece4778
XL		 356static int ip_tun_parse_opts(struct nlattr *attr, struct ip_tunnel_info *info,
357 struct netlink_ext_ack *extack)
358{
359 struct nlattr *tb[LWTUNNEL_IP_OPTS_MAX + 1];
360 int err;
361
362 if (!attr)
363 return 0;
364
365 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_OPTS_MAX, attr,
366 ip_opts_policy, extack);
367 if (err)
368 return err;
369
370 if (tb[LWTUNNEL_IP_OPTS_GENEVE])
371 err = ip_tun_parse_opts_geneve(tb[LWTUNNEL_IP_OPTS_GENEVE],
372 info, extack);
edf31cbb
XL		 373 else if (tb[LWTUNNEL_IP_OPTS_VXLAN])
374 err = ip_tun_parse_opts_vxlan(tb[LWTUNNEL_IP_OPTS_VXLAN],
375 info, extack);
b0a21810
XL		 376 else if (tb[LWTUNNEL_IP_OPTS_ERSPAN])
377 err = ip_tun_parse_opts_erspan(tb[LWTUNNEL_IP_OPTS_ERSPAN],
378 info, extack);
4ece4778
XL		 379 else
380 err = -EINVAL;
381
382 return err;
383}
384
385static int ip_tun_get_optlen(struct nlattr *attr,
386 struct netlink_ext_ack *extack)
387{
388 return ip_tun_parse_opts(attr, NULL, extack);
389}
390
391static int ip_tun_set_opts(struct nlattr *attr, struct ip_tunnel_info *info,
392 struct netlink_ext_ack *extack)
393{
394 return ip_tun_parse_opts(attr, info, extack);
395}
396
30357d7d 397static int ip_tun_build_state(struct nlattr *attr,
127eb7cd 398 unsigned int family, const void *cfg,
9ae28727
DA		 399 struct lwtunnel_state **ts,
400 struct netlink_ext_ack *extack)
3093fbe7 401{
a1c234f9 402 struct nlattr *tb[LWTUNNEL_IP_MAX + 1];
4ece4778
XL		 403 struct lwtunnel_state *new_state;
404 struct ip_tunnel_info *tun_info;
405 int err, opt_len;
3093fbe7 406
8cb08174
JB		 407 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_MAX, attr,
408 ip_tun_policy, extack);
3093fbe7
TG		 409 if (err < 0)
410 return err;
411
4ece4778
XL		 412 opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP_OPTS], extack);
413 if (opt_len < 0)
414 return opt_len;
415
416 new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len);
3093fbe7
TG		 417 if (!new_state)
418 return -ENOMEM;
419
420 new_state->type = LWTUNNEL_ENCAP_IP;
421
422 tun_info = lwt_tun_info(new_state);
423
4ece4778
XL		 424 err = ip_tun_set_opts(tb[LWTUNNEL_IP_OPTS], tun_info, extack);
425 if (err < 0) {
426 lwtstate_free(new_state);
427 return err;
428 }
429
3d25eabb 430#ifdef CONFIG_DST_CACHE
431 err = dst_cache_init(&tun_info->dst_cache, GFP_KERNEL);
432 if (err) {
433 lwtstate_free(new_state);
434 return err;
435 }
436#endif
437
a1c234f9 438 if (tb[LWTUNNEL_IP_ID])
30d3d83a 439 tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP_ID]);
3093fbe7 440
a1c234f9 441 if (tb[LWTUNNEL_IP_DST])
7822ce73 442 tun_info->key.u.ipv4.dst = nla_get_in_addr(tb[LWTUNNEL_IP_DST]);
3093fbe7 443
a1c234f9 444 if (tb[LWTUNNEL_IP_SRC])
7822ce73 445 tun_info->key.u.ipv4.src = nla_get_in_addr(tb[LWTUNNEL_IP_SRC]);
3093fbe7 446
a1c234f9 447 if (tb[LWTUNNEL_IP_TTL])
7c383fb2 448 tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP_TTL]);
3093fbe7 449
a1c234f9 450 if (tb[LWTUNNEL_IP_TOS])
7c383fb2 451 tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP_TOS]);
3093fbe7 452
a1c234f9 453 if (tb[LWTUNNEL_IP_FLAGS])
4ece4778 454 tun_info->key.tun_flags |= nla_get_be16(tb[LWTUNNEL_IP_FLAGS]);
3093fbe7
TG		 455
456 tun_info->mode = IP_TUNNEL_INFO_TX;
4ece4778 457 tun_info->options_len = opt_len;
3093fbe7
TG		 458
459 *ts = new_state;
460
461 return 0;
462}
463
3d25eabb 464static void ip_tun_destroy_state(struct lwtunnel_state *lwtstate)
465{
466#ifdef CONFIG_DST_CACHE
467 struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
468
469 dst_cache_destroy(&tun_info->dst_cache);
470#endif
471}
472
4ece4778
XL		 473static int ip_tun_fill_encap_opts_geneve(struct sk_buff *skb,
474 struct ip_tunnel_info *tun_info)
475{
476 struct geneve_opt *opt;
477 struct nlattr *nest;
478
479 nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_GENEVE);
480 if (!nest)
481 return -ENOMEM;
482
483 opt = ip_tunnel_info_opts(tun_info);
484 if (nla_put_be16(skb, LWTUNNEL_IP_OPT_GENEVE_CLASS, opt->opt_class) ||
485 nla_put_u8(skb, LWTUNNEL_IP_OPT_GENEVE_TYPE, opt->type) ||
486 nla_put(skb, LWTUNNEL_IP_OPT_GENEVE_DATA, opt->length * 4,
487 opt->opt_data)) {
488 nla_nest_cancel(skb, nest);
489 return -ENOMEM;
490 }
491
492 nla_nest_end(skb, nest);
493 return 0;
494}
495
edf31cbb
XL		 496static int ip_tun_fill_encap_opts_vxlan(struct sk_buff *skb,
497 struct ip_tunnel_info *tun_info)
498{
499 struct vxlan_metadata *md;
500 struct nlattr *nest;
501
502 nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_VXLAN);
503 if (!nest)
504 return -ENOMEM;
505
506 md = ip_tunnel_info_opts(tun_info);
507 if (nla_put_u32(skb, LWTUNNEL_IP_OPT_VXLAN_GBP, md->gbp)) {
508 nla_nest_cancel(skb, nest);
509 return -ENOMEM;
510 }
511
512 nla_nest_end(skb, nest);
513 return 0;
514}
515
b0a21810
XL		 516static int ip_tun_fill_encap_opts_erspan(struct sk_buff *skb,
517 struct ip_tunnel_info *tun_info)
518{
519 struct erspan_metadata *md;
520 struct nlattr *nest;
521
522 nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_ERSPAN);
523 if (!nest)
524 return -ENOMEM;
525
526 md = ip_tunnel_info_opts(tun_info);
527 if (nla_put_u32(skb, LWTUNNEL_IP_OPT_ERSPAN_VER, md->version))
528 goto err;
529
530 if (md->version == 1 &&
531 nla_put_be32(skb, LWTUNNEL_IP_OPT_ERSPAN_INDEX, md->u.index))
532 goto err;
533
534 if (md->version == 2 &&
535 (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_DIR, md->u.md2.dir) ||
536 nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_HWID,
537 get_hwid(&md->u.md2))))
538 goto err;
539
540 nla_nest_end(skb, nest);
541 return 0;
542err:
543 nla_nest_cancel(skb, nest);
544 return -ENOMEM;
545}
546
4ece4778
XL		 547static int ip_tun_fill_encap_opts(struct sk_buff *skb, int type,
548 struct ip_tunnel_info *tun_info)
549{
550 struct nlattr *nest;
551 int err = 0;
552
edf31cbb 553 if (!(tun_info->key.tun_flags &
b0a21810 554 (TUNNEL_GENEVE_OPT | TUNNEL_VXLAN_OPT | TUNNEL_ERSPAN_OPT)))
4ece4778
XL		 555 return 0;
556
557 nest = nla_nest_start_noflag(skb, type);
558 if (!nest)
559 return -ENOMEM;
560
561 if (tun_info->key.tun_flags & TUNNEL_GENEVE_OPT)
562 err = ip_tun_fill_encap_opts_geneve(skb, tun_info);
edf31cbb
XL		 563 else if (tun_info->key.tun_flags & TUNNEL_VXLAN_OPT)
564 err = ip_tun_fill_encap_opts_vxlan(skb, tun_info);
b0a21810
XL		 565 else if (tun_info->key.tun_flags & TUNNEL_ERSPAN_OPT)
566 err = ip_tun_fill_encap_opts_erspan(skb, tun_info);
4ece4778
XL		 567
568 if (err) {
569 nla_nest_cancel(skb, nest);
570 return err;
571 }
572
573 nla_nest_end(skb, nest);
574 return 0;
575}
576
3093fbe7
TG		 577static int ip_tun_fill_encap_info(struct sk_buff *skb,
578 struct lwtunnel_state *lwtstate)
579{
580 struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
581
b46f6ded
ND		 582 if (nla_put_be64(skb, LWTUNNEL_IP_ID, tun_info->key.tun_id,
583 LWTUNNEL_IP_PAD) ||
7822ce73
HY		 584 nla_put_in_addr(skb, LWTUNNEL_IP_DST, tun_info->key.u.ipv4.dst) ||
585 nla_put_in_addr(skb, LWTUNNEL_IP_SRC, tun_info->key.u.ipv4.src) ||
7c383fb2
JB		 586 nla_put_u8(skb, LWTUNNEL_IP_TOS, tun_info->key.tos) ||
587 nla_put_u8(skb, LWTUNNEL_IP_TTL, tun_info->key.ttl) ||
4ece4778
XL		 588 nla_put_be16(skb, LWTUNNEL_IP_FLAGS, tun_info->key.tun_flags) ||
589 ip_tun_fill_encap_opts(skb, LWTUNNEL_IP_OPTS, tun_info))
3093fbe7
TG		 590 return -ENOMEM;
591
592 return 0;
593}
594
4ece4778
XL		 595static int ip_tun_opts_nlsize(struct ip_tunnel_info *info)
596{
597 int opt_len;
598
edf31cbb 599 if (!(info->key.tun_flags &
b0a21810 600 (TUNNEL_GENEVE_OPT | TUNNEL_VXLAN_OPT | TUNNEL_ERSPAN_OPT)))
4ece4778
XL		 601 return 0;
602
603 opt_len = nla_total_size(0); /* LWTUNNEL_IP_OPTS */
604 if (info->key.tun_flags & TUNNEL_GENEVE_OPT) {
605 struct geneve_opt *opt = ip_tunnel_info_opts(info);
606
607 opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_GENEVE */
608 + nla_total_size(2) /* OPT_GENEVE_CLASS */
609 + nla_total_size(1) /* OPT_GENEVE_TYPE */
610 + nla_total_size(opt->length * 4);
611 /* OPT_GENEVE_DATA */
edf31cbb
XL		 612 } else if (info->key.tun_flags & TUNNEL_VXLAN_OPT) {
613 opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_VXLAN */
614 + nla_total_size(4); /* OPT_VXLAN_GBP */
b0a21810
XL		 615 } else if (info->key.tun_flags & TUNNEL_ERSPAN_OPT) {
616 opt_len += nla_total_size(0) /* LWTUNNEL_IP_OPTS_ERSPAN */
617 + nla_total_size(1) /* OPT_ERSPAN_VER */
618 + nla_total_size(4); /* OPT_ERSPAN_INDEX/DIR/HWID */
4ece4778
XL		 619 }
620
621 return opt_len;
622}
623
3093fbe7
TG		 624static int ip_tun_encap_nlsize(struct lwtunnel_state *lwtstate)
625{
b46f6ded 626 return nla_total_size_64bit(8) /* LWTUNNEL_IP_ID */
a1c234f9
JB		 627 + nla_total_size(4) /* LWTUNNEL_IP_DST */
628 + nla_total_size(4) /* LWTUNNEL_IP_SRC */
629 + nla_total_size(1) /* LWTUNNEL_IP_TOS */
630 + nla_total_size(1) /* LWTUNNEL_IP_TTL */
4ece4778
XL		 631 + nla_total_size(2) /* LWTUNNEL_IP_FLAGS */
632 + ip_tun_opts_nlsize(lwt_tun_info(lwtstate));
633 /* LWTUNNEL_IP_OPTS */
3093fbe7
TG		 634}
635
2d798499
JB		 636static int ip_tun_cmp_encap(struct lwtunnel_state *a, struct lwtunnel_state *b)
637{
0eb8eb2f
XL		 638 struct ip_tunnel_info *info_a = lwt_tun_info(a);
639 struct ip_tunnel_info *info_b = lwt_tun_info(b);
640
641 return memcmp(info_a, info_b, sizeof(info_a->key)) ||
642 info_a->mode != info_b->mode ||
643 info_a->options_len != info_b->options_len ||
644 memcmp(ip_tunnel_info_opts(info_a),
645 ip_tunnel_info_opts(info_b), info_a->options_len);
2d798499
JB		 646}
647
3093fbe7
TG		 648static const struct lwtunnel_encap_ops ip_tun_lwt_ops = {
649 .build_state = ip_tun_build_state,
3d25eabb 650 .destroy_state = ip_tun_destroy_state,
3093fbe7
TG		 651 .fill_encap = ip_tun_fill_encap_info,
652 .get_encap_size = ip_tun_encap_nlsize,
2d798499 653 .cmp_encap = ip_tun_cmp_encap,
88ff7334 654 .owner = THIS_MODULE,
3093fbe7
TG		 655};
656
32a2b002
JB		 657static const struct nla_policy ip6_tun_policy[LWTUNNEL_IP6_MAX + 1] = {
658 [LWTUNNEL_IP6_ID] = { .type = NLA_U64 },
659 [LWTUNNEL_IP6_DST] = { .len = sizeof(struct in6_addr) },
660 [LWTUNNEL_IP6_SRC] = { .len = sizeof(struct in6_addr) },
661 [LWTUNNEL_IP6_HOPLIMIT] = { .type = NLA_U8 },
662 [LWTUNNEL_IP6_TC] = { .type = NLA_U8 },
32a2b002
JB		 663 [LWTUNNEL_IP6_FLAGS] = { .type = NLA_U16 },
664};
665
30357d7d 666static int ip6_tun_build_state(struct nlattr *attr,
127eb7cd 667 unsigned int family, const void *cfg,
9ae28727
DA		 668 struct lwtunnel_state **ts,
669 struct netlink_ext_ack *extack)
32a2b002 670{
32a2b002 671 struct nlattr *tb[LWTUNNEL_IP6_MAX + 1];
4ece4778
XL		 672 struct lwtunnel_state *new_state;
673 struct ip_tunnel_info *tun_info;
674 int err, opt_len;
32a2b002 675
8cb08174
JB		 676 err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP6_MAX, attr,
677 ip6_tun_policy, extack);
32a2b002
JB		 678 if (err < 0)
679 return err;
680
4ece4778
XL		 681 opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP6_OPTS], extack);
682 if (opt_len < 0)
683 return opt_len;
684
685 new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len);
32a2b002
JB		 686 if (!new_state)
687 return -ENOMEM;
688
689 new_state->type = LWTUNNEL_ENCAP_IP6;
690
691 tun_info = lwt_tun_info(new_state);
692
4ece4778
XL		 693 err = ip_tun_set_opts(tb[LWTUNNEL_IP6_OPTS], tun_info, extack);
694 if (err < 0) {
695 lwtstate_free(new_state);
696 return err;
697 }
698
32a2b002 699 if (tb[LWTUNNEL_IP6_ID])
30d3d83a 700 tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP6_ID]);
32a2b002
JB		 701
702 if (tb[LWTUNNEL_IP6_DST])
703 tun_info->key.u.ipv6.dst = nla_get_in6_addr(tb[LWTUNNEL_IP6_DST]);
704
705 if (tb[LWTUNNEL_IP6_SRC])
706 tun_info->key.u.ipv6.src = nla_get_in6_addr(tb[LWTUNNEL_IP6_SRC]);
707
708 if (tb[LWTUNNEL_IP6_HOPLIMIT])
709 tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP6_HOPLIMIT]);
710
711 if (tb[LWTUNNEL_IP6_TC])
712 tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP6_TC]);
713
32a2b002 714 if (tb[LWTUNNEL_IP6_FLAGS])
4ece4778 715 tun_info->key.tun_flags |= nla_get_be16(tb[LWTUNNEL_IP6_FLAGS]);
32a2b002 716
7f9562a1 717 tun_info->mode = IP_TUNNEL_INFO_TX | IP_TUNNEL_INFO_IPV6;
4ece4778 718 tun_info->options_len = opt_len;
32a2b002
JB		 719
720 *ts = new_state;
721
722 return 0;
723}
724
725static int ip6_tun_fill_encap_info(struct sk_buff *skb,
726 struct lwtunnel_state *lwtstate)
727{
728 struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
729
b46f6ded
ND		 730 if (nla_put_be64(skb, LWTUNNEL_IP6_ID, tun_info->key.tun_id,
731 LWTUNNEL_IP6_PAD) ||
32a2b002
JB		 732 nla_put_in6_addr(skb, LWTUNNEL_IP6_DST, &tun_info->key.u.ipv6.dst) ||
733 nla_put_in6_addr(skb, LWTUNNEL_IP6_SRC, &tun_info->key.u.ipv6.src) ||
995096a0
QA		 734 nla_put_u8(skb, LWTUNNEL_IP6_TC, tun_info->key.tos) ||
735 nla_put_u8(skb, LWTUNNEL_IP6_HOPLIMIT, tun_info->key.ttl) ||
4ece4778
XL		 736 nla_put_be16(skb, LWTUNNEL_IP6_FLAGS, tun_info->key.tun_flags) ||
737 ip_tun_fill_encap_opts(skb, LWTUNNEL_IP6_OPTS, tun_info))
32a2b002
JB		 738 return -ENOMEM;
739
740 return 0;
741}
742
743static int ip6_tun_encap_nlsize(struct lwtunnel_state *lwtstate)
744{
b46f6ded 745 return nla_total_size_64bit(8) /* LWTUNNEL_IP6_ID */
32a2b002
JB		 746 + nla_total_size(16) /* LWTUNNEL_IP6_DST */
747 + nla_total_size(16) /* LWTUNNEL_IP6_SRC */
748 + nla_total_size(1) /* LWTUNNEL_IP6_HOPLIMIT */
749 + nla_total_size(1) /* LWTUNNEL_IP6_TC */
4ece4778
XL		 750 + nla_total_size(2) /* LWTUNNEL_IP6_FLAGS */
751 + ip_tun_opts_nlsize(lwt_tun_info(lwtstate));
752 /* LWTUNNEL_IP6_OPTS */
32a2b002
JB		 753}
754
755static const struct lwtunnel_encap_ops ip6_tun_lwt_ops = {
756 .build_state = ip6_tun_build_state,
757 .fill_encap = ip6_tun_fill_encap_info,
758 .get_encap_size = ip6_tun_encap_nlsize,
759 .cmp_encap = ip_tun_cmp_encap,
88ff7334 760 .owner = THIS_MODULE,
32a2b002
JB		 761};
762
045a0fa0 763void __init ip_tunnel_core_init(void)
3093fbe7 764{
fca5fdf6
DB		 765 /* If you land here, make sure whether increasing ip_tunnel_info's
766 * options_len is a reasonable choice with its usage in front ends
767 * (f.e., it's part of flow keys, etc).
768 */
769 BUILD_BUG_ON(IP_TUNNEL_OPTS_MAX != 255);
770
3093fbe7 771 lwtunnel_encap_add_ops(&ip_tun_lwt_ops, LWTUNNEL_ENCAP_IP);
32a2b002 772 lwtunnel_encap_add_ops(&ip6_tun_lwt_ops, LWTUNNEL_ENCAP_IP6);
3093fbe7 773}
e7030878 774
5263a98f 775DEFINE_STATIC_KEY_FALSE(ip_tunnel_metadata_cnt);
e7030878
TG		 776EXPORT_SYMBOL(ip_tunnel_metadata_cnt);
777
778void ip_tunnel_need_metadata(void)
779{
5263a98f 780 static_branch_inc(&ip_tunnel_metadata_cnt);
e7030878
TG		 781}
782EXPORT_SYMBOL_GPL(ip_tunnel_need_metadata);
783
784void ip_tunnel_unneed_metadata(void)
785{
5263a98f 786 static_branch_dec(&ip_tunnel_metadata_cnt);
e7030878
TG		 787}
788EXPORT_SYMBOL_GPL(ip_tunnel_unneed_metadata);
Linux 6.x block layer and io_uring tree(s)