net/netfilter/nft_tunnel.c

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #include <linux/kernel.h>
   3 #include <linux/init.h>
   4 #include <linux/module.h>
   5 #include <linux/seqlock.h>
   6 #include <linux/netlink.h>
   7 #include <linux/netfilter.h>
   8 #include <linux/netfilter/nf_tables.h>
   9 #include <net/netfilter/nf_tables.h>
  10 #include <net/dst_metadata.h>
  11 #include <net/ip_tunnels.h>
  12 #include <net/vxlan.h>
  13 #include <net/erspan.h>
  14 #include <net/geneve.h>
  15
  16 struct nft_tunnel {
  17         enum nft_tunnel_keys    key:8;
  18         u8                      dreg;
  19         enum nft_tunnel_mode    mode:8;
  20         u8                      len;
  21 };
  22
  23 static void nft_tunnel_get_eval(const struct nft_expr *expr,
  24                                 struct nft_regs *regs,
  25                                 const struct nft_pktinfo *pkt)
  26 {
  27         const struct nft_tunnel *priv = nft_expr_priv(expr);
  28         u32 *dest = &regs->data[priv->dreg];
  29         struct ip_tunnel_info *tun_info;
  30
  31         tun_info = skb_tunnel_info(pkt->skb);
  32
  33         switch (priv->key) {
  34         case NFT_TUNNEL_PATH:
  35                 if (!tun_info) {
  36                         nft_reg_store8(dest, false);
  37                         return;
  38                 }
  39                 if (priv->mode == NFT_TUNNEL_MODE_NONE ||
  40                     (priv->mode == NFT_TUNNEL_MODE_RX &&
  41                      !(tun_info->mode & IP_TUNNEL_INFO_TX)) ||
  42                     (priv->mode == NFT_TUNNEL_MODE_TX &&
  43                      (tun_info->mode & IP_TUNNEL_INFO_TX)))
  44                         nft_reg_store8(dest, true);
  45                 else
  46                         nft_reg_store8(dest, false);
  47                 break;
  48         case NFT_TUNNEL_ID:
  49                 if (!tun_info) {
  50                         regs->verdict.code = NFT_BREAK;
  51                         return;
  52                 }
  53                 if (priv->mode == NFT_TUNNEL_MODE_NONE ||
  54                     (priv->mode == NFT_TUNNEL_MODE_RX &&
  55                      !(tun_info->mode & IP_TUNNEL_INFO_TX)) ||
  56                     (priv->mode == NFT_TUNNEL_MODE_TX &&
  57                      (tun_info->mode & IP_TUNNEL_INFO_TX)))
  58                         *dest = ntohl(tunnel_id_to_key32(tun_info->key.tun_id));
  59                 else
  60                         regs->verdict.code = NFT_BREAK;
  61                 break;
  62         default:
  63                 WARN_ON(1);
  64                 regs->verdict.code = NFT_BREAK;
  65         }
  66 }
  67
  68 static const struct nla_policy nft_tunnel_policy[NFTA_TUNNEL_MAX + 1] = {
  69         [NFTA_TUNNEL_KEY]       = NLA_POLICY_MAX(NLA_BE32, 255),
  70         [NFTA_TUNNEL_DREG]      = { .type = NLA_U32 },
  71         [NFTA_TUNNEL_MODE]      = NLA_POLICY_MAX(NLA_BE32, 255),
  72 };
  73
  74 static int nft_tunnel_get_init(const struct nft_ctx *ctx,
  75                                const struct nft_expr *expr,
  76                                const struct nlattr * const tb[])
  77 {
  78         struct nft_tunnel *priv = nft_expr_priv(expr);
  79         u32 len;
  80
  81         if (!tb[NFTA_TUNNEL_KEY] ||
  82             !tb[NFTA_TUNNEL_DREG])
  83                 return -EINVAL;
  84
  85         priv->key = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY]));
  86         switch (priv->key) {
  87         case NFT_TUNNEL_PATH:
  88                 len = sizeof(u8);
  89                 break;
  90         case NFT_TUNNEL_ID:
  91                 len = sizeof(u32);
  92                 break;
  93         default:
  94                 return -EOPNOTSUPP;
  95         }
  96
  97         if (tb[NFTA_TUNNEL_MODE]) {
  98                 priv->mode = ntohl(nla_get_be32(tb[NFTA_TUNNEL_MODE]));
  99                 if (priv->mode > NFT_TUNNEL_MODE_MAX)
 100                         return -EOPNOTSUPP;
 101         } else {
 102                 priv->mode = NFT_TUNNEL_MODE_NONE;
 103         }
 104
 105         priv->len = len;
 106         return nft_parse_register_store(ctx, tb[NFTA_TUNNEL_DREG], &priv->dreg,
 107                                         NULL, NFT_DATA_VALUE, len);
 108 }
 109
 110 static int nft_tunnel_get_dump(struct sk_buff *skb,
 111                                const struct nft_expr *expr, bool reset)
 112 {
 113         const struct nft_tunnel *priv = nft_expr_priv(expr);
 114
 115         if (nla_put_be32(skb, NFTA_TUNNEL_KEY, htonl(priv->key)))
 116                 goto nla_put_failure;
 117         if (nft_dump_register(skb, NFTA_TUNNEL_DREG, priv->dreg))
 118                 goto nla_put_failure;
 119         if (nla_put_be32(skb, NFTA_TUNNEL_MODE, htonl(priv->mode)))
 120                 goto nla_put_failure;
 121         return 0;
 122
 123 nla_put_failure:
 124         return -1;
 125 }
 126
 127 static bool nft_tunnel_get_reduce(struct nft_regs_track *track,
 128                                   const struct nft_expr *expr)
 129 {
 130         const struct nft_tunnel *priv = nft_expr_priv(expr);
 131         const struct nft_tunnel *tunnel;
 132
 133         if (!nft_reg_track_cmp(track, expr, priv->dreg)) {
 134                 nft_reg_track_update(track, expr, priv->dreg, priv->len);
 135                 return false;
 136         }
 137
 138         tunnel = nft_expr_priv(track->regs[priv->dreg].selector);
 139         if (priv->key != tunnel->key ||
 140             priv->dreg != tunnel->dreg ||
 141             priv->mode != tunnel->mode) {
 142                 nft_reg_track_update(track, expr, priv->dreg, priv->len);
 143                 return false;
 144         }
 145
 146         if (!track->regs[priv->dreg].bitwise)
 147                 return true;
 148
 149         return false;
 150 }
 151
 152 static struct nft_expr_type nft_tunnel_type;
 153 static const struct nft_expr_ops nft_tunnel_get_ops = {
 154         .type           = &nft_tunnel_type,
 155         .size           = NFT_EXPR_SIZE(sizeof(struct nft_tunnel)),
 156         .eval           = nft_tunnel_get_eval,
 157         .init           = nft_tunnel_get_init,
 158         .dump           = nft_tunnel_get_dump,
 159         .reduce         = nft_tunnel_get_reduce,
 160 };
 161
 162 static struct nft_expr_type nft_tunnel_type __read_mostly = {
 163         .name           = "tunnel",
 164         .family         = NFPROTO_NETDEV,
 165         .ops            = &nft_tunnel_get_ops,
 166         .policy         = nft_tunnel_policy,
 167         .maxattr        = NFTA_TUNNEL_MAX,
 168         .owner          = THIS_MODULE,
 169 };
 170
 171 struct nft_tunnel_opts {
 172         union {
 173                 struct vxlan_metadata   vxlan;
 174                 struct erspan_metadata  erspan;
 175                 u8      data[IP_TUNNEL_OPTS_MAX];
 176         } u;
 177         IP_TUNNEL_DECLARE_FLAGS(flags);
 178         u32     len;
 179 };
 180
 181 struct nft_tunnel_obj {
 182         struct metadata_dst     *md;
 183         struct nft_tunnel_opts  opts;
 184 };
 185
 186 static const struct nla_policy nft_tunnel_ip_policy[NFTA_TUNNEL_KEY_IP_MAX + 1] = {
 187         [NFTA_TUNNEL_KEY_IP_SRC]        = { .type = NLA_U32 },
 188         [NFTA_TUNNEL_KEY_IP_DST]        = { .type = NLA_U32 },
 189 };
 190
 191 static int nft_tunnel_obj_ip_init(const struct nft_ctx *ctx,
 192                                   const struct nlattr *attr,
 193                                   struct ip_tunnel_info *info)
 194 {
 195         struct nlattr *tb[NFTA_TUNNEL_KEY_IP_MAX + 1];
 196         int err;
 197
 198         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_IP_MAX, attr,
 199                                           nft_tunnel_ip_policy, NULL);
 200         if (err < 0)
 201                 return err;
 202
 203         if (!tb[NFTA_TUNNEL_KEY_IP_DST])
 204                 return -EINVAL;
 205
 206         if (tb[NFTA_TUNNEL_KEY_IP_SRC])
 207                 info->key.u.ipv4.src = nla_get_be32(tb[NFTA_TUNNEL_KEY_IP_SRC]);
 208         if (tb[NFTA_TUNNEL_KEY_IP_DST])
 209                 info->key.u.ipv4.dst = nla_get_be32(tb[NFTA_TUNNEL_KEY_IP_DST]);
 210
 211         return 0;
 212 }
 213
 214 static const struct nla_policy nft_tunnel_ip6_policy[NFTA_TUNNEL_KEY_IP6_MAX + 1] = {
 215         [NFTA_TUNNEL_KEY_IP6_SRC]       = { .len = sizeof(struct in6_addr), },
 216         [NFTA_TUNNEL_KEY_IP6_DST]       = { .len = sizeof(struct in6_addr), },
 217         [NFTA_TUNNEL_KEY_IP6_FLOWLABEL] = { .type = NLA_U32, }
 218 };
 219
 220 static int nft_tunnel_obj_ip6_init(const struct nft_ctx *ctx,
 221                                    const struct nlattr *attr,
 222                                    struct ip_tunnel_info *info)
 223 {
 224         struct nlattr *tb[NFTA_TUNNEL_KEY_IP6_MAX + 1];
 225         int err;
 226
 227         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_IP6_MAX, attr,
 228                                           nft_tunnel_ip6_policy, NULL);
 229         if (err < 0)
 230                 return err;
 231
 232         if (!tb[NFTA_TUNNEL_KEY_IP6_DST])
 233                 return -EINVAL;
 234
 235         if (tb[NFTA_TUNNEL_KEY_IP6_SRC]) {
 236                 memcpy(&info->key.u.ipv6.src,
 237                        nla_data(tb[NFTA_TUNNEL_KEY_IP6_SRC]),
 238                        sizeof(struct in6_addr));
 239         }
 240         if (tb[NFTA_TUNNEL_KEY_IP6_DST]) {
 241                 memcpy(&info->key.u.ipv6.dst,
 242                        nla_data(tb[NFTA_TUNNEL_KEY_IP6_DST]),
 243                        sizeof(struct in6_addr));
 244         }
 245         if (tb[NFTA_TUNNEL_KEY_IP6_FLOWLABEL])
 246                 info->key.label = nla_get_be32(tb[NFTA_TUNNEL_KEY_IP6_FLOWLABEL]);
 247
 248         info->mode |= IP_TUNNEL_INFO_IPV6;
 249
 250         return 0;
 251 }
 252
 253 static const struct nla_policy nft_tunnel_opts_vxlan_policy[NFTA_TUNNEL_KEY_VXLAN_MAX + 1] = {
 254         [NFTA_TUNNEL_KEY_VXLAN_GBP]     = { .type = NLA_U32 },
 255 };
 256
 257 static int nft_tunnel_obj_vxlan_init(const struct nlattr *attr,
 258                                      struct nft_tunnel_opts *opts)
 259 {
 260         struct nlattr *tb[NFTA_TUNNEL_KEY_VXLAN_MAX + 1];
 261         int err;
 262
 263         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_VXLAN_MAX, attr,
 264                                           nft_tunnel_opts_vxlan_policy, NULL);
 265         if (err < 0)
 266                 return err;
 267
 268         if (!tb[NFTA_TUNNEL_KEY_VXLAN_GBP])
 269                 return -EINVAL;
 270
 271         opts->u.vxlan.gbp = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY_VXLAN_GBP]));
 272
 273         opts->len       = sizeof(struct vxlan_metadata);
 274         ip_tunnel_flags_zero(opts->flags);
 275         __set_bit(IP_TUNNEL_VXLAN_OPT_BIT, opts->flags);
 276
 277         return 0;
 278 }
 279
 280 static const struct nla_policy nft_tunnel_opts_erspan_policy[NFTA_TUNNEL_KEY_ERSPAN_MAX + 1] = {
 281         [NFTA_TUNNEL_KEY_ERSPAN_VERSION]        = { .type = NLA_U32 },
 282         [NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX]       = { .type = NLA_U32 },
 283         [NFTA_TUNNEL_KEY_ERSPAN_V2_DIR]         = { .type = NLA_U8 },
 284         [NFTA_TUNNEL_KEY_ERSPAN_V2_HWID]        = { .type = NLA_U8 },
 285 };
 286
 287 static int nft_tunnel_obj_erspan_init(const struct nlattr *attr,
 288                                       struct nft_tunnel_opts *opts)
 289 {
 290         struct nlattr *tb[NFTA_TUNNEL_KEY_ERSPAN_MAX + 1];
 291         uint8_t hwid, dir;
 292         int err, version;
 293
 294         err = nla_parse_nested_deprecated(tb, NFTA_TUNNEL_KEY_ERSPAN_MAX,
 295                                           attr, nft_tunnel_opts_erspan_policy,
 296                                           NULL);
 297         if (err < 0)
 298                 return err;
 299
 300         if (!tb[NFTA_TUNNEL_KEY_ERSPAN_VERSION])
 301                  return -EINVAL;
 302
 303         version = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY_ERSPAN_VERSION]));
 304         switch (version) {
 305         case ERSPAN_VERSION:
 306                 if (!tb[NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX])
 307                         return -EINVAL;
 308
 309                 opts->u.erspan.u.index =
 310                         nla_get_be32(tb[NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX]);
 311                 break;
 312         case ERSPAN_VERSION2:
 313                 if (!tb[NFTA_TUNNEL_KEY_ERSPAN_V2_DIR] ||
 314                     !tb[NFTA_TUNNEL_KEY_ERSPAN_V2_HWID])
 315                         return -EINVAL;
 316
 317                 hwid = nla_get_u8(tb[NFTA_TUNNEL_KEY_ERSPAN_V2_HWID]);
 318                 dir = nla_get_u8(tb[NFTA_TUNNEL_KEY_ERSPAN_V2_DIR]);
 319
 320                 set_hwid(&opts->u.erspan.u.md2, hwid);
 321                 opts->u.erspan.u.md2.dir = dir;
 322                 break;
 323         default:
 324                 return -EOPNOTSUPP;
 325         }
 326         opts->u.erspan.version = version;
 327
 328         opts->len       = sizeof(struct erspan_metadata);
 329         ip_tunnel_flags_zero(opts->flags);
 330         __set_bit(IP_TUNNEL_ERSPAN_OPT_BIT, opts->flags);
 331
 332         return 0;
 333 }
 334
 335 static const struct nla_policy nft_tunnel_opts_geneve_policy[NFTA_TUNNEL_KEY_GENEVE_MAX + 1] = {
 336         [NFTA_TUNNEL_KEY_GENEVE_CLASS]  = { .type = NLA_U16 },
 337         [NFTA_TUNNEL_KEY_GENEVE_TYPE]   = { .type = NLA_U8 },
 338         [NFTA_TUNNEL_KEY_GENEVE_DATA]   = { .type = NLA_BINARY, .len = 128 },
 339 };
 340
 341 static int nft_tunnel_obj_geneve_init(const struct nlattr *attr,
 342                                       struct nft_tunnel_opts *opts)
 343 {
 344         struct geneve_opt *opt = (struct geneve_opt *)opts->u.data + opts->len;
 345         struct nlattr *tb[NFTA_TUNNEL_KEY_GENEVE_MAX + 1];
 346         int err, data_len;
 347
 348         err = nla_parse_nested(tb, NFTA_TUNNEL_KEY_GENEVE_MAX, attr,
 349                                nft_tunnel_opts_geneve_policy, NULL);
 350         if (err < 0)
 351                 return err;
 352
 353         if (!tb[NFTA_TUNNEL_KEY_GENEVE_CLASS] ||
 354             !tb[NFTA_TUNNEL_KEY_GENEVE_TYPE] ||
 355             !tb[NFTA_TUNNEL_KEY_GENEVE_DATA])
 356                 return -EINVAL;
 357
 358         attr = tb[NFTA_TUNNEL_KEY_GENEVE_DATA];
 359         data_len = nla_len(attr);
 360         if (data_len % 4)
 361                 return -EINVAL;
 362
 363         opts->len += sizeof(*opt) + data_len;
 364         if (opts->len > IP_TUNNEL_OPTS_MAX)
 365                 return -EINVAL;
 366
 367         memcpy(opt->opt_data, nla_data(attr), data_len);
 368         opt->length = data_len / 4;
 369         opt->opt_class = nla_get_be16(tb[NFTA_TUNNEL_KEY_GENEVE_CLASS]);
 370         opt->type = nla_get_u8(tb[NFTA_TUNNEL_KEY_GENEVE_TYPE]);
 371         ip_tunnel_flags_zero(opts->flags);
 372         __set_bit(IP_TUNNEL_GENEVE_OPT_BIT, opts->flags);
 373
 374         return 0;
 375 }
 376
 377 static const struct nla_policy nft_tunnel_opts_policy[NFTA_TUNNEL_KEY_OPTS_MAX + 1] = {
 378         [NFTA_TUNNEL_KEY_OPTS_UNSPEC]   = {
 379                 .strict_start_type = NFTA_TUNNEL_KEY_OPTS_GENEVE },
 380         [NFTA_TUNNEL_KEY_OPTS_VXLAN]    = { .type = NLA_NESTED, },
 381         [NFTA_TUNNEL_KEY_OPTS_ERSPAN]   = { .type = NLA_NESTED, },
 382         [NFTA_TUNNEL_KEY_OPTS_GENEVE]   = { .type = NLA_NESTED, },
 383 };
 384
 385 static int nft_tunnel_obj_opts_init(const struct nft_ctx *ctx,
 386                                     const struct nlattr *attr,
 387                                     struct ip_tunnel_info *info,
 388                                     struct nft_tunnel_opts *opts)
 389 {
 390         struct nlattr *nla;
 391         int err, rem;
 392         u32 type = 0;
 393
 394         err = nla_validate_nested_deprecated(attr, NFTA_TUNNEL_KEY_OPTS_MAX,
 395                                              nft_tunnel_opts_policy, NULL);
 396         if (err < 0)
 397                 return err;
 398
 399         nla_for_each_attr(nla, nla_data(attr), nla_len(attr), rem) {
 400                 switch (nla_type(nla)) {
 401                 case NFTA_TUNNEL_KEY_OPTS_VXLAN:
 402                         if (type)
 403                                 return -EINVAL;
 404                         err = nft_tunnel_obj_vxlan_init(nla, opts);
 405                         if (err)
 406                                 return err;
 407                         type = IP_TUNNEL_VXLAN_OPT_BIT;
 408                         break;
 409                 case NFTA_TUNNEL_KEY_OPTS_ERSPAN:
 410                         if (type)
 411                                 return -EINVAL;
 412                         err = nft_tunnel_obj_erspan_init(nla, opts);
 413                         if (err)
 414                                 return err;
 415                         type = IP_TUNNEL_ERSPAN_OPT_BIT;
 416                         break;
 417                 case NFTA_TUNNEL_KEY_OPTS_GENEVE:
 418                         if (type && type != IP_TUNNEL_GENEVE_OPT_BIT)
 419                                 return -EINVAL;
 420                         err = nft_tunnel_obj_geneve_init(nla, opts);
 421                         if (err)
 422                                 return err;
 423                         type = IP_TUNNEL_GENEVE_OPT_BIT;
 424                         break;
 425                 default:
 426                         return -EOPNOTSUPP;
 427                 }
 428         }
 429
 430         return err;
 431 }
 432
 433 static const struct nla_policy nft_tunnel_key_policy[NFTA_TUNNEL_KEY_MAX + 1] = {
 434         [NFTA_TUNNEL_KEY_IP]    = { .type = NLA_NESTED, },
 435         [NFTA_TUNNEL_KEY_IP6]   = { .type = NLA_NESTED, },
 436         [NFTA_TUNNEL_KEY_ID]    = { .type = NLA_U32, },
 437         [NFTA_TUNNEL_KEY_FLAGS] = { .type = NLA_U32, },
 438         [NFTA_TUNNEL_KEY_TOS]   = { .type = NLA_U8, },
 439         [NFTA_TUNNEL_KEY_TTL]   = { .type = NLA_U8, },
 440         [NFTA_TUNNEL_KEY_SPORT] = { .type = NLA_U16, },
 441         [NFTA_TUNNEL_KEY_DPORT] = { .type = NLA_U16, },
 442         [NFTA_TUNNEL_KEY_OPTS]  = { .type = NLA_NESTED, },
 443 };
 444
 445 static int nft_tunnel_obj_init(const struct nft_ctx *ctx,
 446                                const struct nlattr * const tb[],
 447                                struct nft_object *obj)
 448 {
 449         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 450         struct ip_tunnel_info info;
 451         struct metadata_dst *md;
 452         int err;
 453
 454         if (!tb[NFTA_TUNNEL_KEY_ID])
 455                 return -EINVAL;
 456
 457         memset(&info, 0, sizeof(info));
 458         info.mode               = IP_TUNNEL_INFO_TX;
 459         info.key.tun_id         = key32_to_tunnel_id(nla_get_be32(tb[NFTA_TUNNEL_KEY_ID]));
 460         __set_bit(IP_TUNNEL_KEY_BIT, info.key.tun_flags);
 461         __set_bit(IP_TUNNEL_CSUM_BIT, info.key.tun_flags);
 462         __set_bit(IP_TUNNEL_NOCACHE_BIT, info.key.tun_flags);
 463
 464         if (tb[NFTA_TUNNEL_KEY_IP]) {
 465                 err = nft_tunnel_obj_ip_init(ctx, tb[NFTA_TUNNEL_KEY_IP], &info);
 466                 if (err < 0)
 467                         return err;
 468         } else if (tb[NFTA_TUNNEL_KEY_IP6]) {
 469                 err = nft_tunnel_obj_ip6_init(ctx, tb[NFTA_TUNNEL_KEY_IP6], &info);
 470                 if (err < 0)
 471                         return err;
 472         } else {
 473                 return -EINVAL;
 474         }
 475
 476         if (tb[NFTA_TUNNEL_KEY_SPORT]) {
 477                 info.key.tp_src = nla_get_be16(tb[NFTA_TUNNEL_KEY_SPORT]);
 478         }
 479         if (tb[NFTA_TUNNEL_KEY_DPORT]) {
 480                 info.key.tp_dst = nla_get_be16(tb[NFTA_TUNNEL_KEY_DPORT]);
 481         }
 482
 483         if (tb[NFTA_TUNNEL_KEY_FLAGS]) {
 484                 u32 tun_flags;
 485
 486                 tun_flags = ntohl(nla_get_be32(tb[NFTA_TUNNEL_KEY_FLAGS]));
 487                 if (tun_flags & ~NFT_TUNNEL_F_MASK)
 488                         return -EOPNOTSUPP;
 489
 490                 if (tun_flags & NFT_TUNNEL_F_ZERO_CSUM_TX)
 491                         __clear_bit(IP_TUNNEL_CSUM_BIT, info.key.tun_flags);
 492                 if (tun_flags & NFT_TUNNEL_F_DONT_FRAGMENT)
 493                         __set_bit(IP_TUNNEL_DONT_FRAGMENT_BIT,
 494                                   info.key.tun_flags);
 495                 if (tun_flags & NFT_TUNNEL_F_SEQ_NUMBER)
 496                         __set_bit(IP_TUNNEL_SEQ_BIT, info.key.tun_flags);
 497         }
 498         if (tb[NFTA_TUNNEL_KEY_TOS])
 499                 info.key.tos = nla_get_u8(tb[NFTA_TUNNEL_KEY_TOS]);
 500         if (tb[NFTA_TUNNEL_KEY_TTL])
 501                 info.key.ttl = nla_get_u8(tb[NFTA_TUNNEL_KEY_TTL]);
 502         else
 503                 info.key.ttl = U8_MAX;
 504
 505         if (tb[NFTA_TUNNEL_KEY_OPTS]) {
 506                 err = nft_tunnel_obj_opts_init(ctx, tb[NFTA_TUNNEL_KEY_OPTS],
 507                                                &info, &priv->opts);
 508                 if (err < 0)
 509                         return err;
 510         }
 511
 512         md = metadata_dst_alloc(priv->opts.len, METADATA_IP_TUNNEL, GFP_KERNEL);
 513         if (!md)
 514                 return -ENOMEM;
 515
 516         memcpy(&md->u.tun_info, &info, sizeof(info));
 517 #ifdef CONFIG_DST_CACHE
 518         err = dst_cache_init(&md->u.tun_info.dst_cache, GFP_KERNEL);
 519         if (err < 0) {
 520                 metadata_dst_free(md);
 521                 return err;
 522         }
 523 #endif
 524         ip_tunnel_info_opts_set(&md->u.tun_info, &priv->opts.u, priv->opts.len,
 525                                 priv->opts.flags);
 526         priv->md = md;
 527
 528         return 0;
 529 }
 530
 531 static inline void nft_tunnel_obj_eval(struct nft_object *obj,
 532                                        struct nft_regs *regs,
 533                                        const struct nft_pktinfo *pkt)
 534 {
 535         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 536         struct sk_buff *skb = pkt->skb;
 537
 538         skb_dst_drop(skb);
 539         dst_hold((struct dst_entry *) priv->md);
 540         skb_dst_set(skb, (struct dst_entry *) priv->md);
 541 }
 542
 543 static int nft_tunnel_ip_dump(struct sk_buff *skb, struct ip_tunnel_info *info)
 544 {
 545         struct nlattr *nest;
 546
 547         if (info->mode & IP_TUNNEL_INFO_IPV6) {
 548                 nest = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_IP6);
 549                 if (!nest)
 550                         return -1;
 551
 552                 if (nla_put_in6_addr(skb, NFTA_TUNNEL_KEY_IP6_SRC,
 553                                      &info->key.u.ipv6.src) < 0 ||
 554                     nla_put_in6_addr(skb, NFTA_TUNNEL_KEY_IP6_DST,
 555                                      &info->key.u.ipv6.dst) < 0 ||
 556                     nla_put_be32(skb, NFTA_TUNNEL_KEY_IP6_FLOWLABEL,
 557                                  info->key.label)) {
 558                         nla_nest_cancel(skb, nest);
 559                         return -1;
 560                 }
 561
 562                 nla_nest_end(skb, nest);
 563         } else {
 564                 nest = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_IP);
 565                 if (!nest)
 566                         return -1;
 567
 568                 if (nla_put_in_addr(skb, NFTA_TUNNEL_KEY_IP_SRC,
 569                                     info->key.u.ipv4.src) < 0 ||
 570                     nla_put_in_addr(skb, NFTA_TUNNEL_KEY_IP_DST,
 571                                     info->key.u.ipv4.dst) < 0) {
 572                         nla_nest_cancel(skb, nest);
 573                         return -1;
 574                 }
 575
 576                 nla_nest_end(skb, nest);
 577         }
 578
 579         return 0;
 580 }
 581
 582 static int nft_tunnel_opts_dump(struct sk_buff *skb,
 583                                 struct nft_tunnel_obj *priv)
 584 {
 585         struct nft_tunnel_opts *opts = &priv->opts;
 586         struct nlattr *nest, *inner;
 587
 588         nest = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS);
 589         if (!nest)
 590                 return -1;
 591
 592         if (test_bit(IP_TUNNEL_VXLAN_OPT_BIT, opts->flags)) {
 593                 inner = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS_VXLAN);
 594                 if (!inner)
 595                         goto failure;
 596                 if (nla_put_be32(skb, NFTA_TUNNEL_KEY_VXLAN_GBP,
 597                                  htonl(opts->u.vxlan.gbp)))
 598                         goto inner_failure;
 599                 nla_nest_end(skb, inner);
 600         } else if (test_bit(IP_TUNNEL_ERSPAN_OPT_BIT, opts->flags)) {
 601                 inner = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS_ERSPAN);
 602                 if (!inner)
 603                         goto failure;
 604                 if (nla_put_be32(skb, NFTA_TUNNEL_KEY_ERSPAN_VERSION,
 605                                  htonl(opts->u.erspan.version)))
 606                         goto inner_failure;
 607                 switch (opts->u.erspan.version) {
 608                 case ERSPAN_VERSION:
 609                         if (nla_put_be32(skb, NFTA_TUNNEL_KEY_ERSPAN_V1_INDEX,
 610                                          opts->u.erspan.u.index))
 611                                 goto inner_failure;
 612                         break;
 613                 case ERSPAN_VERSION2:
 614                         if (nla_put_u8(skb, NFTA_TUNNEL_KEY_ERSPAN_V2_HWID,
 615                                        get_hwid(&opts->u.erspan.u.md2)) ||
 616                             nla_put_u8(skb, NFTA_TUNNEL_KEY_ERSPAN_V2_DIR,
 617                                        opts->u.erspan.u.md2.dir))
 618                                 goto inner_failure;
 619                         break;
 620                 }
 621                 nla_nest_end(skb, inner);
 622         } else if (test_bit(IP_TUNNEL_GENEVE_OPT_BIT, opts->flags)) {
 623                 struct geneve_opt *opt;
 624                 int offset = 0;
 625
 626                 inner = nla_nest_start_noflag(skb, NFTA_TUNNEL_KEY_OPTS_GENEVE);
 627                 if (!inner)
 628                         goto failure;
 629                 while (opts->len > offset) {
 630                         opt = (struct geneve_opt *)opts->u.data + offset;
 631                         if (nla_put_be16(skb, NFTA_TUNNEL_KEY_GENEVE_CLASS,
 632                                          opt->opt_class) ||
 633                             nla_put_u8(skb, NFTA_TUNNEL_KEY_GENEVE_TYPE,
 634                                        opt->type) ||
 635                             nla_put(skb, NFTA_TUNNEL_KEY_GENEVE_DATA,
 636                                     opt->length * 4, opt->opt_data))
 637                                 goto inner_failure;
 638                         offset += sizeof(*opt) + opt->length * 4;
 639                 }
 640                 nla_nest_end(skb, inner);
 641         }
 642         nla_nest_end(skb, nest);
 643         return 0;
 644
 645 inner_failure:
 646         nla_nest_cancel(skb, inner);
 647 failure:
 648         nla_nest_cancel(skb, nest);
 649         return -1;
 650 }
 651
 652 static int nft_tunnel_ports_dump(struct sk_buff *skb,
 653                                  struct ip_tunnel_info *info)
 654 {
 655         if (nla_put_be16(skb, NFTA_TUNNEL_KEY_SPORT, info->key.tp_src) < 0 ||
 656             nla_put_be16(skb, NFTA_TUNNEL_KEY_DPORT, info->key.tp_dst) < 0)
 657                 return -1;
 658
 659         return 0;
 660 }
 661
 662 static int nft_tunnel_flags_dump(struct sk_buff *skb,
 663                                  struct ip_tunnel_info *info)
 664 {
 665         u32 flags = 0;
 666
 667         if (test_bit(IP_TUNNEL_DONT_FRAGMENT_BIT, info->key.tun_flags))
 668                 flags |= NFT_TUNNEL_F_DONT_FRAGMENT;
 669         if (!test_bit(IP_TUNNEL_CSUM_BIT, info->key.tun_flags))
 670                 flags |= NFT_TUNNEL_F_ZERO_CSUM_TX;
 671         if (test_bit(IP_TUNNEL_SEQ_BIT, info->key.tun_flags))
 672                 flags |= NFT_TUNNEL_F_SEQ_NUMBER;
 673
 674         if (nla_put_be32(skb, NFTA_TUNNEL_KEY_FLAGS, htonl(flags)) < 0)
 675                 return -1;
 676
 677         return 0;
 678 }
 679
 680 static int nft_tunnel_obj_dump(struct sk_buff *skb,
 681                                struct nft_object *obj, bool reset)
 682 {
 683         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 684         struct ip_tunnel_info *info = &priv->md->u.tun_info;
 685
 686         if (nla_put_be32(skb, NFTA_TUNNEL_KEY_ID,
 687                          tunnel_id_to_key32(info->key.tun_id)) ||
 688             nft_tunnel_ip_dump(skb, info) < 0 ||
 689             nft_tunnel_ports_dump(skb, info) < 0 ||
 690             nft_tunnel_flags_dump(skb, info) < 0 ||
 691             nla_put_u8(skb, NFTA_TUNNEL_KEY_TOS, info->key.tos) ||
 692             nla_put_u8(skb, NFTA_TUNNEL_KEY_TTL, info->key.ttl) ||
 693             nft_tunnel_opts_dump(skb, priv) < 0)
 694                 goto nla_put_failure;
 695
 696         return 0;
 697
 698 nla_put_failure:
 699         return -1;
 700 }
 701
 702 static void nft_tunnel_obj_destroy(const struct nft_ctx *ctx,
 703                                    struct nft_object *obj)
 704 {
 705         struct nft_tunnel_obj *priv = nft_obj_data(obj);
 706
 707         metadata_dst_free(priv->md);
 708 }
 709
 710 static struct nft_object_type nft_tunnel_obj_type;
 711 static const struct nft_object_ops nft_tunnel_obj_ops = {
 712         .type           = &nft_tunnel_obj_type,
 713         .size           = sizeof(struct nft_tunnel_obj),
 714         .eval           = nft_tunnel_obj_eval,
 715         .init           = nft_tunnel_obj_init,
 716         .destroy        = nft_tunnel_obj_destroy,
 717         .dump           = nft_tunnel_obj_dump,
 718 };
 719
 720 static struct nft_object_type nft_tunnel_obj_type __read_mostly = {
 721         .type           = NFT_OBJECT_TUNNEL,
 722         .family         = NFPROTO_NETDEV,
 723         .ops            = &nft_tunnel_obj_ops,
 724         .maxattr        = NFTA_TUNNEL_KEY_MAX,
 725         .policy         = nft_tunnel_key_policy,
 726         .owner          = THIS_MODULE,
 727 };
 728
 729 static int __init nft_tunnel_module_init(void)
 730 {
 731         int err;
 732
 733         err = nft_register_expr(&nft_tunnel_type);
 734         if (err < 0)
 735                 return err;
 736
 737         err = nft_register_obj(&nft_tunnel_obj_type);
 738         if (err < 0)
 739                 nft_unregister_expr(&nft_tunnel_type);
 740
 741         return err;
 742 }
 743
 744 static void __exit nft_tunnel_module_exit(void)
 745 {
 746         nft_unregister_obj(&nft_tunnel_obj_type);
 747         nft_unregister_expr(&nft_tunnel_type);
 748 }
 749
 750 module_init(nft_tunnel_module_init);
 751 module_exit(nft_tunnel_module_exit);
 752
 753 MODULE_LICENSE("GPL");
 754 MODULE_AUTHOR("Pablo Neira Ayuso <[email protected]>");
 755 MODULE_ALIAS_NFT_EXPR("tunnel");
 756 MODULE_ALIAS_NFT_OBJ(NFT_OBJECT_TUNNEL);
 757 MODULE_DESCRIPTION("nftables tunnel expression support");