net/ipv4/ip_tunnel_core.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2013 Nicira, Inc.
   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>
  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>
  21 #include <linux/static_key.h>
  22
  23 #include <net/ip.h>
  24 #include <net/icmp.h>
  25 #include <net/protocol.h>
  26 #include <net/ip_tunnels.h>
  27 #include <net/ip6_tunnel.h>
  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>
  36 #include <net/dst_metadata.h>
  37 #include <net/geneve.h>
  38 #include <net/vxlan.h>
  39 #include <net/erspan.h>
  40
  41 const struct ip_tunnel_encap_ops __rcu *
  42                 iptun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly;
  43 EXPORT_SYMBOL(iptun_encaps);
  44
  45 const struct ip6_tnl_encap_ops __rcu *
  46                 ip6tun_encaps[MAX_IPTUN_ENCAP_OPS] __read_mostly;
  47 EXPORT_SYMBOL(ip6tun_encaps);
  48
  49 void 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)
  52 {
  53         int pkt_len = skb->len - skb_inner_network_offset(skb);
  54         struct net *net = dev_net(rt->dst.dev);
  55         struct net_device *dev = skb->dev;
  56         struct iphdr *iph;
  57         int err;
  58
  59         skb_scrub_packet(skb, xnet);
  60
  61         skb_clear_hash_if_not_l4(skb);
  62         skb_dst_set(skb, &rt->dst);
  63         memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
  64
  65         /* Push down and install the IP header. */
  66         skb_push(skb, sizeof(struct iphdr));
  67         skb_reset_network_header(skb);
  68
  69         iph = ip_hdr(skb);
  70
  71         iph->version    =       4;
  72         iph->ihl        =       sizeof(struct iphdr) >> 2;
  73         iph->frag_off   =       ip_mtu_locked(&rt->dst) ? 0 : df;
  74         iph->protocol   =       proto;
  75         iph->tos        =       tos;
  76         iph->daddr      =       dst;
  77         iph->saddr      =       src;
  78         iph->ttl        =       ttl;
  79         __ip_select_ident(net, iph, skb_shinfo(skb)->gso_segs ?: 1);
  80
  81         err = ip_local_out(net, sk, skb);
  82
  83         if (dev) {
  84                 if (unlikely(net_xmit_eval(err)))
  85                         pkt_len = 0;
  86                 iptunnel_xmit_stats(dev, pkt_len);
  87         }
  88 }
  89 EXPORT_SYMBOL_GPL(iptunnel_xmit);
  90
  91 int __iptunnel_pull_header(struct sk_buff *skb, int hdr_len,
  92                            __be16 inner_proto, bool raw_proto, bool xnet)
  93 {
  94         if (unlikely(!pskb_may_pull(skb, hdr_len)))
  95                 return -ENOMEM;
  96
  97         skb_pull_rcsum(skb, hdr_len);
  98
  99         if (!raw_proto && inner_proto == htons(ETH_P_TEB)) {
 100                 struct ethhdr *eh;
 101
 102                 if (unlikely(!pskb_may_pull(skb, ETH_HLEN)))
 103                         return -ENOMEM;
 104
 105                 eh = (struct ethhdr *)skb->data;
 106                 if (likely(eth_proto_is_802_3(eh->h_proto)))
 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
 115         skb_clear_hash_if_not_l4(skb);
 116         __vlan_hwaccel_clear_tag(skb);
 117         skb_set_queue_mapping(skb, 0);
 118         skb_scrub_packet(skb, xnet);
 119
 120         return iptunnel_pull_offloads(skb);
 121 }
 122 EXPORT_SYMBOL_GPL(__iptunnel_pull_header);
 123
 124 struct 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
 130         if (!md || md->type != METADATA_IP_TUNNEL ||
 131             md->u.tun_info.mode & IP_TUNNEL_INFO_TX)
 132                 return NULL;
 133
 134         src = &md->u.tun_info;
 135         res = metadata_dst_alloc(src->options_len, METADATA_IP_TUNNEL, flags);
 136         if (!res)
 137                 return NULL;
 138
 139         dst = &res->u.tun_info;
 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;
 146         dst->key.tun_flags = src->key.tun_flags;
 147         dst->mode = src->mode | IP_TUNNEL_INFO_TX;
 148         ip_tunnel_info_opts_set(dst, ip_tunnel_info_opts(src),
 149                                 src->options_len, 0);
 150
 151         return res;
 152 }
 153 EXPORT_SYMBOL_GPL(iptunnel_metadata_reply);
 154
 155 int iptunnel_handle_offloads(struct sk_buff *skb,
 156                              int gso_type_mask)
 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)) {
 166                 err = skb_header_unclone(skb, GFP_ATOMIC);
 167                 if (unlikely(err))
 168                         return err;
 169                 skb_shinfo(skb)->gso_type |= gso_type_mask;
 170                 return 0;
 171         }
 172
 173         if (skb->ip_summed != CHECKSUM_PARTIAL) {
 174                 skb->ip_summed = CHECKSUM_NONE;
 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                  */
 180                 skb->encapsulation = 0;
 181         }
 182
 183         return 0;
 184 }
 185 EXPORT_SYMBOL_GPL(iptunnel_handle_offloads);
 186
 187 /* Often modified stats are per cpu, other are shared (netdev->stats) */
 188 void ip_tunnel_get_stats64(struct net_device *dev,
 189                            struct rtnl_link_stats64 *tot)
 190 {
 191         int i;
 192
 193         netdev_stats_to_stats64(tot, &dev->stats);
 194
 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 {
 202                         start = u64_stats_fetch_begin_irq(&tstats->syncp);
 203                         rx_packets = tstats->rx_packets;
 204                         tx_packets = tstats->tx_packets;
 205                         rx_bytes = tstats->rx_bytes;
 206                         tx_bytes = tstats->tx_bytes;
 207                 } while (u64_stats_fetch_retry_irq(&tstats->syncp, start));
 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         }
 214 }
 215 EXPORT_SYMBOL_GPL(ip_tunnel_get_stats64);
 216
 217 static const struct nla_policy ip_tun_policy[LWTUNNEL_IP_MAX + 1] = {
 218         [LWTUNNEL_IP_UNSPEC]    = { .strict_start_type = LWTUNNEL_IP_OPTS },
 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 },
 224         [LWTUNNEL_IP_FLAGS]     = { .type = NLA_U16 },
 225         [LWTUNNEL_IP_OPTS]      = { .type = NLA_NESTED },
 226 };
 227
 228 static const struct nla_policy ip_opts_policy[LWTUNNEL_IP_OPTS_MAX + 1] = {
 229         [LWTUNNEL_IP_OPTS_GENEVE]       = { .type = NLA_NESTED },
 230         [LWTUNNEL_IP_OPTS_VXLAN]        = { .type = NLA_NESTED },
 231         [LWTUNNEL_IP_OPTS_ERSPAN]       = { .type = NLA_NESTED },
 232 };
 233
 234 static const struct nla_policy
 235 geneve_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
 241 static const struct nla_policy
 242 vxlan_opt_policy[LWTUNNEL_IP_OPT_VXLAN_MAX + 1] = {
 243         [LWTUNNEL_IP_OPT_VXLAN_GBP]     = { .type = NLA_U32 },
 244 };
 245
 246 static const struct nla_policy
 247 erspan_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
 254 static int ip_tun_parse_opts_geneve(struct nlattr *attr,
 255                                     struct ip_tunnel_info *info, int opts_len,
 256                                     struct netlink_ext_ack *extack)
 257 {
 258         struct nlattr *tb[LWTUNNEL_IP_OPT_GENEVE_MAX + 1];
 259         int data_len, err;
 260
 261         err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_GENEVE_MAX, attr,
 262                                geneve_opt_policy, extack);
 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) {
 277                 struct geneve_opt *opt = ip_tunnel_info_opts(info) + opts_len;
 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
 291 static int ip_tun_parse_opts_vxlan(struct nlattr *attr,
 292                                    struct ip_tunnel_info *info, int opts_len,
 293                                    struct netlink_ext_ack *extack)
 294 {
 295         struct nlattr *tb[LWTUNNEL_IP_OPT_VXLAN_MAX + 1];
 296         int err;
 297
 298         err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_VXLAN_MAX, attr,
 299                                vxlan_opt_policy, extack);
 300         if (err)
 301                 return err;
 302
 303         if (!tb[LWTUNNEL_IP_OPT_VXLAN_GBP])
 304                 return -EINVAL;
 305
 306         if (info) {
 307                 struct vxlan_metadata *md =
 308                         ip_tunnel_info_opts(info) + opts_len;
 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
 318 static int ip_tun_parse_opts_erspan(struct nlattr *attr,
 319                                     struct ip_tunnel_info *info, int opts_len,
 320                                     struct netlink_ext_ack *extack)
 321 {
 322         struct nlattr *tb[LWTUNNEL_IP_OPT_ERSPAN_MAX + 1];
 323         int err;
 324
 325         err = nla_parse_nested(tb, LWTUNNEL_IP_OPT_ERSPAN_MAX, attr,
 326                                erspan_opt_policy, extack);
 327         if (err)
 328                 return err;
 329
 330         if (!tb[LWTUNNEL_IP_OPT_ERSPAN_VER])
 331                 return -EINVAL;
 332
 333         if (info) {
 334                 struct erspan_metadata *md =
 335                         ip_tunnel_info_opts(info) + opts_len;
 336
 337                 attr = tb[LWTUNNEL_IP_OPT_ERSPAN_VER];
 338                 md->version = nla_get_u8(attr);
 339
 340                 if (md->version == 1 && tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX]) {
 341                         attr = tb[LWTUNNEL_IP_OPT_ERSPAN_INDEX];
 342                         md->u.index = nla_get_be32(attr);
 343                 } else if (md->version == 2 && tb[LWTUNNEL_IP_OPT_ERSPAN_DIR] &&
 344                            tb[LWTUNNEL_IP_OPT_ERSPAN_HWID]) {
 345                         attr = tb[LWTUNNEL_IP_OPT_ERSPAN_DIR];
 346                         md->u.md2.dir = nla_get_u8(attr);
 347                         attr = tb[LWTUNNEL_IP_OPT_ERSPAN_HWID];
 348                         set_hwid(&md->u.md2, nla_get_u8(attr));
 349                 } else {
 350                         return -EINVAL;
 351                 }
 352
 353                 info->key.tun_flags |= TUNNEL_ERSPAN_OPT;
 354         }
 355
 356         return sizeof(struct erspan_metadata);
 357 }
 358
 359 static int ip_tun_parse_opts(struct nlattr *attr, struct ip_tunnel_info *info,
 360                              struct netlink_ext_ack *extack)
 361 {
 362         int err, rem, opt_len, opts_len = 0, type = 0;
 363         struct nlattr *nla;
 364
 365         if (!attr)
 366                 return 0;
 367
 368         err = nla_validate(nla_data(attr), nla_len(attr), LWTUNNEL_IP_OPTS_MAX,
 369                            ip_opts_policy, extack);
 370         if (err)
 371                 return err;
 372
 373         nla_for_each_attr(nla, nla_data(attr), nla_len(attr), rem) {
 374                 switch (nla_type(nla)) {
 375                 case LWTUNNEL_IP_OPTS_GENEVE:
 376                         if (type && type != TUNNEL_GENEVE_OPT)
 377                                 return -EINVAL;
 378                         opt_len = ip_tun_parse_opts_geneve(nla, info, opts_len,
 379                                                            extack);
 380                         if (opt_len < 0)
 381                                 return opt_len;
 382                         opts_len += opt_len;
 383                         if (opts_len > IP_TUNNEL_OPTS_MAX)
 384                                 return -EINVAL;
 385                         type = TUNNEL_GENEVE_OPT;
 386                         break;
 387                 case LWTUNNEL_IP_OPTS_VXLAN:
 388                         if (type)
 389                                 return -EINVAL;
 390                         opt_len = ip_tun_parse_opts_vxlan(nla, info, opts_len,
 391                                                           extack);
 392                         if (opt_len < 0)
 393                                 return opt_len;
 394                         opts_len += opt_len;
 395                         type = TUNNEL_VXLAN_OPT;
 396                         break;
 397                 case LWTUNNEL_IP_OPTS_ERSPAN:
 398                         if (type)
 399                                 return -EINVAL;
 400                         opt_len = ip_tun_parse_opts_erspan(nla, info, opts_len,
 401                                                            extack);
 402                         if (opt_len < 0)
 403                                 return opt_len;
 404                         opts_len += opt_len;
 405                         type = TUNNEL_ERSPAN_OPT;
 406                         break;
 407                 default:
 408                         return -EINVAL;
 409                 }
 410         }
 411
 412         return opts_len;
 413 }
 414
 415 static int ip_tun_get_optlen(struct nlattr *attr,
 416                              struct netlink_ext_ack *extack)
 417 {
 418         return ip_tun_parse_opts(attr, NULL, extack);
 419 }
 420
 421 static int ip_tun_set_opts(struct nlattr *attr, struct ip_tunnel_info *info,
 422                            struct netlink_ext_ack *extack)
 423 {
 424         return ip_tun_parse_opts(attr, info, extack);
 425 }
 426
 427 static int ip_tun_build_state(struct nlattr *attr,
 428                               unsigned int family, const void *cfg,
 429                               struct lwtunnel_state **ts,
 430                               struct netlink_ext_ack *extack)
 431 {
 432         struct nlattr *tb[LWTUNNEL_IP_MAX + 1];
 433         struct lwtunnel_state *new_state;
 434         struct ip_tunnel_info *tun_info;
 435         int err, opt_len;
 436
 437         err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP_MAX, attr,
 438                                           ip_tun_policy, extack);
 439         if (err < 0)
 440                 return err;
 441
 442         opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP_OPTS], extack);
 443         if (opt_len < 0)
 444                 return opt_len;
 445
 446         new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len);
 447         if (!new_state)
 448                 return -ENOMEM;
 449
 450         new_state->type = LWTUNNEL_ENCAP_IP;
 451
 452         tun_info = lwt_tun_info(new_state);
 453
 454         err = ip_tun_set_opts(tb[LWTUNNEL_IP_OPTS], tun_info, extack);
 455         if (err < 0) {
 456                 lwtstate_free(new_state);
 457                 return err;
 458         }
 459
 460 #ifdef CONFIG_DST_CACHE
 461         err = dst_cache_init(&tun_info->dst_cache, GFP_KERNEL);
 462         if (err) {
 463                 lwtstate_free(new_state);
 464                 return err;
 465         }
 466 #endif
 467
 468         if (tb[LWTUNNEL_IP_ID])
 469                 tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP_ID]);
 470
 471         if (tb[LWTUNNEL_IP_DST])
 472                 tun_info->key.u.ipv4.dst = nla_get_in_addr(tb[LWTUNNEL_IP_DST]);
 473
 474         if (tb[LWTUNNEL_IP_SRC])
 475                 tun_info->key.u.ipv4.src = nla_get_in_addr(tb[LWTUNNEL_IP_SRC]);
 476
 477         if (tb[LWTUNNEL_IP_TTL])
 478                 tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP_TTL]);
 479
 480         if (tb[LWTUNNEL_IP_TOS])
 481                 tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP_TOS]);
 482
 483         if (tb[LWTUNNEL_IP_FLAGS])
 484                 tun_info->key.tun_flags |=
 485                                 (nla_get_be16(tb[LWTUNNEL_IP_FLAGS]) &
 486                                  ~TUNNEL_OPTIONS_PRESENT);
 487
 488         tun_info->mode = IP_TUNNEL_INFO_TX;
 489         tun_info->options_len = opt_len;
 490
 491         *ts = new_state;
 492
 493         return 0;
 494 }
 495
 496 static void ip_tun_destroy_state(struct lwtunnel_state *lwtstate)
 497 {
 498 #ifdef CONFIG_DST_CACHE
 499         struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
 500
 501         dst_cache_destroy(&tun_info->dst_cache);
 502 #endif
 503 }
 504
 505 static int ip_tun_fill_encap_opts_geneve(struct sk_buff *skb,
 506                                          struct ip_tunnel_info *tun_info)
 507 {
 508         struct geneve_opt *opt;
 509         struct nlattr *nest;
 510         int offset = 0;
 511
 512         nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_GENEVE);
 513         if (!nest)
 514                 return -ENOMEM;
 515
 516         while (tun_info->options_len > offset) {
 517                 opt = ip_tunnel_info_opts(tun_info) + offset;
 518                 if (nla_put_be16(skb, LWTUNNEL_IP_OPT_GENEVE_CLASS,
 519                                  opt->opt_class) ||
 520                     nla_put_u8(skb, LWTUNNEL_IP_OPT_GENEVE_TYPE, opt->type) ||
 521                     nla_put(skb, LWTUNNEL_IP_OPT_GENEVE_DATA, opt->length * 4,
 522                             opt->opt_data)) {
 523                         nla_nest_cancel(skb, nest);
 524                         return -ENOMEM;
 525                 }
 526                 offset += sizeof(*opt) + opt->length * 4;
 527         }
 528
 529         nla_nest_end(skb, nest);
 530         return 0;
 531 }
 532
 533 static int ip_tun_fill_encap_opts_vxlan(struct sk_buff *skb,
 534                                         struct ip_tunnel_info *tun_info)
 535 {
 536         struct vxlan_metadata *md;
 537         struct nlattr *nest;
 538
 539         nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_VXLAN);
 540         if (!nest)
 541                 return -ENOMEM;
 542
 543         md = ip_tunnel_info_opts(tun_info);
 544         if (nla_put_u32(skb, LWTUNNEL_IP_OPT_VXLAN_GBP, md->gbp)) {
 545                 nla_nest_cancel(skb, nest);
 546                 return -ENOMEM;
 547         }
 548
 549         nla_nest_end(skb, nest);
 550         return 0;
 551 }
 552
 553 static int ip_tun_fill_encap_opts_erspan(struct sk_buff *skb,
 554                                          struct ip_tunnel_info *tun_info)
 555 {
 556         struct erspan_metadata *md;
 557         struct nlattr *nest;
 558
 559         nest = nla_nest_start_noflag(skb, LWTUNNEL_IP_OPTS_ERSPAN);
 560         if (!nest)
 561                 return -ENOMEM;
 562
 563         md = ip_tunnel_info_opts(tun_info);
 564         if (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_VER, md->version))
 565                 goto err;
 566
 567         if (md->version == 1 &&
 568             nla_put_be32(skb, LWTUNNEL_IP_OPT_ERSPAN_INDEX, md->u.index))
 569                 goto err;
 570
 571         if (md->version == 2 &&
 572             (nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_DIR, md->u.md2.dir) ||
 573              nla_put_u8(skb, LWTUNNEL_IP_OPT_ERSPAN_HWID,
 574                         get_hwid(&md->u.md2))))
 575                 goto err;
 576
 577         nla_nest_end(skb, nest);
 578         return 0;
 579 err:
 580         nla_nest_cancel(skb, nest);
 581         return -ENOMEM;
 582 }
 583
 584 static int ip_tun_fill_encap_opts(struct sk_buff *skb, int type,
 585                                   struct ip_tunnel_info *tun_info)
 586 {
 587         struct nlattr *nest;
 588         int err = 0;
 589
 590         if (!(tun_info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))
 591                 return 0;
 592
 593         nest = nla_nest_start_noflag(skb, type);
 594         if (!nest)
 595                 return -ENOMEM;
 596
 597         if (tun_info->key.tun_flags & TUNNEL_GENEVE_OPT)
 598                 err = ip_tun_fill_encap_opts_geneve(skb, tun_info);
 599         else if (tun_info->key.tun_flags & TUNNEL_VXLAN_OPT)
 600                 err = ip_tun_fill_encap_opts_vxlan(skb, tun_info);
 601         else if (tun_info->key.tun_flags & TUNNEL_ERSPAN_OPT)
 602                 err = ip_tun_fill_encap_opts_erspan(skb, tun_info);
 603
 604         if (err) {
 605                 nla_nest_cancel(skb, nest);
 606                 return err;
 607         }
 608
 609         nla_nest_end(skb, nest);
 610         return 0;
 611 }
 612
 613 static int ip_tun_fill_encap_info(struct sk_buff *skb,
 614                                   struct lwtunnel_state *lwtstate)
 615 {
 616         struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
 617
 618         if (nla_put_be64(skb, LWTUNNEL_IP_ID, tun_info->key.tun_id,
 619                          LWTUNNEL_IP_PAD) ||
 620             nla_put_in_addr(skb, LWTUNNEL_IP_DST, tun_info->key.u.ipv4.dst) ||
 621             nla_put_in_addr(skb, LWTUNNEL_IP_SRC, tun_info->key.u.ipv4.src) ||
 622             nla_put_u8(skb, LWTUNNEL_IP_TOS, tun_info->key.tos) ||
 623             nla_put_u8(skb, LWTUNNEL_IP_TTL, tun_info->key.ttl) ||
 624             nla_put_be16(skb, LWTUNNEL_IP_FLAGS, tun_info->key.tun_flags) ||
 625             ip_tun_fill_encap_opts(skb, LWTUNNEL_IP_OPTS, tun_info))
 626                 return -ENOMEM;
 627
 628         return 0;
 629 }
 630
 631 static int ip_tun_opts_nlsize(struct ip_tunnel_info *info)
 632 {
 633         int opt_len;
 634
 635         if (!(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))
 636                 return 0;
 637
 638         opt_len = nla_total_size(0);            /* LWTUNNEL_IP_OPTS */
 639         if (info->key.tun_flags & TUNNEL_GENEVE_OPT) {
 640                 struct geneve_opt *opt;
 641                 int offset = 0;
 642
 643                 opt_len += nla_total_size(0);   /* LWTUNNEL_IP_OPTS_GENEVE */
 644                 while (info->options_len > offset) {
 645                         opt = ip_tunnel_info_opts(info) + offset;
 646                         opt_len += nla_total_size(2)    /* OPT_GENEVE_CLASS */
 647                                    + nla_total_size(1)  /* OPT_GENEVE_TYPE */
 648                                    + nla_total_size(opt->length * 4);
 649                                                         /* OPT_GENEVE_DATA */
 650                         offset += sizeof(*opt) + opt->length * 4;
 651                 }
 652         } else if (info->key.tun_flags & TUNNEL_VXLAN_OPT) {
 653                 opt_len += nla_total_size(0)    /* LWTUNNEL_IP_OPTS_VXLAN */
 654                            + nla_total_size(4); /* OPT_VXLAN_GBP */
 655         } else if (info->key.tun_flags & TUNNEL_ERSPAN_OPT) {
 656                 struct erspan_metadata *md = ip_tunnel_info_opts(info);
 657
 658                 opt_len += nla_total_size(0)    /* LWTUNNEL_IP_OPTS_ERSPAN */
 659                            + nla_total_size(1)  /* OPT_ERSPAN_VER */
 660                            + (md->version == 1 ? nla_total_size(4)
 661                                                 /* OPT_ERSPAN_INDEX (v1) */
 662                                                : nla_total_size(1) +
 663                                                  nla_total_size(1));
 664                                                 /* OPT_ERSPAN_DIR + HWID (v2) */
 665         }
 666
 667         return opt_len;
 668 }
 669
 670 static int ip_tun_encap_nlsize(struct lwtunnel_state *lwtstate)
 671 {
 672         return nla_total_size_64bit(8)  /* LWTUNNEL_IP_ID */
 673                 + nla_total_size(4)     /* LWTUNNEL_IP_DST */
 674                 + nla_total_size(4)     /* LWTUNNEL_IP_SRC */
 675                 + nla_total_size(1)     /* LWTUNNEL_IP_TOS */
 676                 + nla_total_size(1)     /* LWTUNNEL_IP_TTL */
 677                 + nla_total_size(2)     /* LWTUNNEL_IP_FLAGS */
 678                 + ip_tun_opts_nlsize(lwt_tun_info(lwtstate));
 679                                         /* LWTUNNEL_IP_OPTS */
 680 }
 681
 682 static int ip_tun_cmp_encap(struct lwtunnel_state *a, struct lwtunnel_state *b)
 683 {
 684         struct ip_tunnel_info *info_a = lwt_tun_info(a);
 685         struct ip_tunnel_info *info_b = lwt_tun_info(b);
 686
 687         return memcmp(info_a, info_b, sizeof(info_a->key)) ||
 688                info_a->mode != info_b->mode ||
 689                info_a->options_len != info_b->options_len ||
 690                memcmp(ip_tunnel_info_opts(info_a),
 691                       ip_tunnel_info_opts(info_b), info_a->options_len);
 692 }
 693
 694 static const struct lwtunnel_encap_ops ip_tun_lwt_ops = {
 695         .build_state = ip_tun_build_state,
 696         .destroy_state = ip_tun_destroy_state,
 697         .fill_encap = ip_tun_fill_encap_info,
 698         .get_encap_size = ip_tun_encap_nlsize,
 699         .cmp_encap = ip_tun_cmp_encap,
 700         .owner = THIS_MODULE,
 701 };
 702
 703 static const struct nla_policy ip6_tun_policy[LWTUNNEL_IP6_MAX + 1] = {
 704         [LWTUNNEL_IP6_UNSPEC]   = { .strict_start_type = LWTUNNEL_IP6_OPTS },
 705         [LWTUNNEL_IP6_ID]               = { .type = NLA_U64 },
 706         [LWTUNNEL_IP6_DST]              = { .len = sizeof(struct in6_addr) },
 707         [LWTUNNEL_IP6_SRC]              = { .len = sizeof(struct in6_addr) },
 708         [LWTUNNEL_IP6_HOPLIMIT]         = { .type = NLA_U8 },
 709         [LWTUNNEL_IP6_TC]               = { .type = NLA_U8 },
 710         [LWTUNNEL_IP6_FLAGS]            = { .type = NLA_U16 },
 711         [LWTUNNEL_IP6_OPTS]             = { .type = NLA_NESTED },
 712 };
 713
 714 static int ip6_tun_build_state(struct nlattr *attr,
 715                                unsigned int family, const void *cfg,
 716                                struct lwtunnel_state **ts,
 717                                struct netlink_ext_ack *extack)
 718 {
 719         struct nlattr *tb[LWTUNNEL_IP6_MAX + 1];
 720         struct lwtunnel_state *new_state;
 721         struct ip_tunnel_info *tun_info;
 722         int err, opt_len;
 723
 724         err = nla_parse_nested_deprecated(tb, LWTUNNEL_IP6_MAX, attr,
 725                                           ip6_tun_policy, extack);
 726         if (err < 0)
 727                 return err;
 728
 729         opt_len = ip_tun_get_optlen(tb[LWTUNNEL_IP6_OPTS], extack);
 730         if (opt_len < 0)
 731                 return opt_len;
 732
 733         new_state = lwtunnel_state_alloc(sizeof(*tun_info) + opt_len);
 734         if (!new_state)
 735                 return -ENOMEM;
 736
 737         new_state->type = LWTUNNEL_ENCAP_IP6;
 738
 739         tun_info = lwt_tun_info(new_state);
 740
 741         err = ip_tun_set_opts(tb[LWTUNNEL_IP6_OPTS], tun_info, extack);
 742         if (err < 0) {
 743                 lwtstate_free(new_state);
 744                 return err;
 745         }
 746
 747         if (tb[LWTUNNEL_IP6_ID])
 748                 tun_info->key.tun_id = nla_get_be64(tb[LWTUNNEL_IP6_ID]);
 749
 750         if (tb[LWTUNNEL_IP6_DST])
 751                 tun_info->key.u.ipv6.dst = nla_get_in6_addr(tb[LWTUNNEL_IP6_DST]);
 752
 753         if (tb[LWTUNNEL_IP6_SRC])
 754                 tun_info->key.u.ipv6.src = nla_get_in6_addr(tb[LWTUNNEL_IP6_SRC]);
 755
 756         if (tb[LWTUNNEL_IP6_HOPLIMIT])
 757                 tun_info->key.ttl = nla_get_u8(tb[LWTUNNEL_IP6_HOPLIMIT]);
 758
 759         if (tb[LWTUNNEL_IP6_TC])
 760                 tun_info->key.tos = nla_get_u8(tb[LWTUNNEL_IP6_TC]);
 761
 762         if (tb[LWTUNNEL_IP6_FLAGS])
 763                 tun_info->key.tun_flags |=
 764                                 (nla_get_be16(tb[LWTUNNEL_IP6_FLAGS]) &
 765                                  ~TUNNEL_OPTIONS_PRESENT);
 766
 767         tun_info->mode = IP_TUNNEL_INFO_TX | IP_TUNNEL_INFO_IPV6;
 768         tun_info->options_len = opt_len;
 769
 770         *ts = new_state;
 771
 772         return 0;
 773 }
 774
 775 static int ip6_tun_fill_encap_info(struct sk_buff *skb,
 776                                    struct lwtunnel_state *lwtstate)
 777 {
 778         struct ip_tunnel_info *tun_info = lwt_tun_info(lwtstate);
 779
 780         if (nla_put_be64(skb, LWTUNNEL_IP6_ID, tun_info->key.tun_id,
 781                          LWTUNNEL_IP6_PAD) ||
 782             nla_put_in6_addr(skb, LWTUNNEL_IP6_DST, &tun_info->key.u.ipv6.dst) ||
 783             nla_put_in6_addr(skb, LWTUNNEL_IP6_SRC, &tun_info->key.u.ipv6.src) ||
 784             nla_put_u8(skb, LWTUNNEL_IP6_TC, tun_info->key.tos) ||
 785             nla_put_u8(skb, LWTUNNEL_IP6_HOPLIMIT, tun_info->key.ttl) ||
 786             nla_put_be16(skb, LWTUNNEL_IP6_FLAGS, tun_info->key.tun_flags) ||
 787             ip_tun_fill_encap_opts(skb, LWTUNNEL_IP6_OPTS, tun_info))
 788                 return -ENOMEM;
 789
 790         return 0;
 791 }
 792
 793 static int ip6_tun_encap_nlsize(struct lwtunnel_state *lwtstate)
 794 {
 795         return nla_total_size_64bit(8)  /* LWTUNNEL_IP6_ID */
 796                 + nla_total_size(16)    /* LWTUNNEL_IP6_DST */
 797                 + nla_total_size(16)    /* LWTUNNEL_IP6_SRC */
 798                 + nla_total_size(1)     /* LWTUNNEL_IP6_HOPLIMIT */
 799                 + nla_total_size(1)     /* LWTUNNEL_IP6_TC */
 800                 + nla_total_size(2)     /* LWTUNNEL_IP6_FLAGS */
 801                 + ip_tun_opts_nlsize(lwt_tun_info(lwtstate));
 802                                         /* LWTUNNEL_IP6_OPTS */
 803 }
 804
 805 static const struct lwtunnel_encap_ops ip6_tun_lwt_ops = {
 806         .build_state = ip6_tun_build_state,
 807         .fill_encap = ip6_tun_fill_encap_info,
 808         .get_encap_size = ip6_tun_encap_nlsize,
 809         .cmp_encap = ip_tun_cmp_encap,
 810         .owner = THIS_MODULE,
 811 };
 812
 813 void __init ip_tunnel_core_init(void)
 814 {
 815         /* If you land here, make sure whether increasing ip_tunnel_info's
 816          * options_len is a reasonable choice with its usage in front ends
 817          * (f.e., it's part of flow keys, etc).
 818          */
 819         BUILD_BUG_ON(IP_TUNNEL_OPTS_MAX != 255);
 820
 821         lwtunnel_encap_add_ops(&ip_tun_lwt_ops, LWTUNNEL_ENCAP_IP);
 822         lwtunnel_encap_add_ops(&ip6_tun_lwt_ops, LWTUNNEL_ENCAP_IP6);
 823 }
 824
 825 DEFINE_STATIC_KEY_FALSE(ip_tunnel_metadata_cnt);
 826 EXPORT_SYMBOL(ip_tunnel_metadata_cnt);
 827
 828 void ip_tunnel_need_metadata(void)
 829 {
 830         static_branch_inc(&ip_tunnel_metadata_cnt);
 831 }
 832 EXPORT_SYMBOL_GPL(ip_tunnel_need_metadata);
 833
 834 void ip_tunnel_unneed_metadata(void)
 835 {
 836         static_branch_dec(&ip_tunnel_metadata_cnt);
 837 }
 838 EXPORT_SYMBOL_GPL(ip_tunnel_unneed_metadata);